feat:优化使用体验

v3.2-refact
panyy 2026-11-05 18:16:48 +08:00
parent 2baba0b19c
commit 5434504f9b
63 changed files with 2815 additions and 685 deletions

View File

@ -183,6 +183,10 @@ class BaseChatStep(IChatStep):
tool_ids=None,
mcp_output_enable=True,
**kwargs):
from system_manage.mcp_agent import context_for
enabled = (model_setting or {}).get('engine_tools', [])
self.context['engine_body'] = {**manage.context, 'workspace_id': workspace_id, 'chat_id': str(chat_id)}
self.context['engine_tools'] = enabled
chat_model = get_model_instance_by_model_workspace_id(model_id, workspace_id,
**model_params_setting) if model_id is not None else None
if stream:
@ -208,7 +212,9 @@ class BaseChatStep(IChatStep):
code_path = f'{executor.sandbox_path}/execute/{tool_id}.py'
if os.path.exists(code_path):
os.remove(code_path)
from system_manage.mcp_agent import call_records
return {
'engine_calls': call_records(self.context.get('engine_body', {})),
'step_type': 'chat_step',
'run_time': self.context['run_time'],
'model_id': str(manage.context['model_id']),
@ -231,25 +237,15 @@ class BaseChatStep(IChatStep):
def _handle_mcp_request(self, mcp_enable, tool_enable, mcp_source, mcp_servers, mcp_tool_ids, tool_ids,
mcp_output_enable, chat_model, message_list):
if not mcp_enable and not tool_enable:
if not mcp_enable and not tool_enable and not self.context.get('engine_tools'):
return None
mcp_servers_config = {}
# 迁移过来mcp_source是None
if mcp_source is None:
mcp_source = 'custom'
if mcp_enable:
# 兼容老数据
if not mcp_tool_ids:
mcp_tool_ids = []
if mcp_source == 'custom' and mcp_servers is not None and '"stdio"' not in mcp_servers:
mcp_servers_config = json.loads(mcp_servers)
elif mcp_tool_ids:
mcp_tools = QuerySet(Tool).filter(id__in=mcp_tool_ids).values()
for mcp_tool in mcp_tools:
if mcp_tool and mcp_tool['is_active']:
mcp_servers_config = {**mcp_servers_config, **json.loads(mcp_tool['code'])}
from system_manage.mcp_agent import context_for
engine_context = context_for(self.context['engine_body'], engine_tools=self.context.get('engine_tools'),
tool_ids=mcp_tool_ids if mcp_enable and mcp_source not in (None, 'custom') else [],
custom=mcp_servers if mcp_enable and mcp_source in (None, 'custom') else None)
if tool_enable:
if tool_ids and len(tool_ids) > 0: # 如果有工具ID,则将其转换为MCP
@ -269,8 +265,8 @@ class BaseChatStep(IChatStep):
self.context['execute_ids'].append(_id)
mcp_servers_config[str(tool.id)] = tool_config
if len(mcp_servers_config) > 0:
return mcp_response_generator(chat_model, message_list, json.dumps(mcp_servers_config), mcp_output_enable)
if mcp_servers_config or engine_context['tool_ids'] or engine_context['custom'] or engine_context['engine_tools']:
return mcp_response_generator(chat_model, message_list, json.dumps(mcp_servers_config), mcp_output_enable, engine_context)
return None
@ -332,6 +328,7 @@ class BaseChatStep(IChatStep):
no_references_setting, problem_text, mcp_enable, mcp_tool_ids, mcp_servers, mcp_source, tool_enable, tool_ids,
mcp_output_enable)
chat_record_id = uuid.uuid7()
self.context.get('engine_body', {})['chat_record_id'] = str(chat_record_id)
r = StreamingHttpResponse(
streaming_content=event_content(chat_result, chat_id, chat_record_id, paragraph_list,
post_response_handler, manage, self, chat_model, message_list, problem_text,
@ -375,7 +372,8 @@ class BaseChatStep(IChatStep):
chat_model, message_list,
)
if mcp_result:
return mcp_result, True
from application.flow.tools import collect_mcp_response
return collect_mcp_response(mcp_result), True
return chat_model.invoke(message_list), True
def execute_block(self, message_list: List[BaseMessage],
@ -401,6 +399,7 @@ class BaseChatStep(IChatStep):
reasoning = Reasoning(reasoning_content_start,
reasoning_content_end)
chat_record_id = uuid.uuid7()
self.context.get('engine_body', {})['chat_record_id'] = str(chat_record_id)
# 调用模型
try:
chat_result, is_ai_chat = self.get_block_result(message_list, chat_model, paragraph_list,

View File

@ -21,6 +21,7 @@ from .loop_continue_node import BaseLoopContinueNode
from .loop_node import *
from .loop_start_node import *
from .mcp_node import BaseMcpNode
from .component_node import BaseComponentNode
from .parameter_extraction_node import BaseParameterExtractionNode
from .question_node import *
from .reranker_node import *
@ -42,7 +43,7 @@ node_list = [BaseStartStepNode, BaseChatNode, BaseSearchKnowledgeNode, BaseSearc
BaseToolNodeNode, BaseToolLibNodeNode, BaseRerankerNode, BaseApplicationNode,
BaseDocumentExtractNode,
BaseImageUnderstandNode, BaseFormNode, BaseSpeechToTextNode, BaseTextToSpeechNode,
BaseImageGenerateNode, BaseVariableAssignNode, BaseMcpNode, BaseTextToVideoNode, BaseImageToVideoNode,
BaseImageGenerateNode, BaseVariableAssignNode, BaseMcpNode, BaseComponentNode, BaseTextToVideoNode, BaseImageToVideoNode,
BaseVideoUnderstandNode,
BaseIntentNode, BaseLoopNode, BaseLoopStartStepNode,
BaseLoopContinueNode,

View File

@ -206,29 +206,23 @@ class BaseChatNode(IChatNode):
def _handle_mcp_request(self, mcp_enable, tool_enable, mcp_source, mcp_servers, mcp_tool_id, mcp_tool_ids, tool_ids,
mcp_output_enable, chat_model, message_list, history_message, question):
if not mcp_enable and not tool_enable:
enabled = (self.node_params.get('model_setting') or {}).get('engine_tools', [])
if not mcp_enable and not tool_enable and not enabled:
return None
mcp_servers_config = {}
# 迁移过来mcp_source是None
if mcp_source is None:
mcp_source = 'custom'
if mcp_enable:
# 兼容老数据
if not mcp_tool_ids:
mcp_tool_ids = []
if mcp_tool_id:
mcp_tool_ids = list(set(mcp_tool_ids + [mcp_tool_id]))
if mcp_source == 'custom' and mcp_servers is not None and '"stdio"' not in mcp_servers:
mcp_servers_config = json.loads(mcp_servers)
mcp_servers_config = self.handle_variables(mcp_servers_config)
elif mcp_tool_ids:
mcp_tools = QuerySet(Tool).filter(id__in=mcp_tool_ids).values()
for mcp_tool in mcp_tools:
if mcp_tool and mcp_tool['is_active']:
mcp_servers_config = {**mcp_servers_config, **json.loads(mcp_tool['code'])}
mcp_servers_config = self.handle_variables(mcp_servers_config)
from system_manage.mcp_agent import context_for
ids = list(mcp_tool_ids or [])
if mcp_tool_id: ids.append(mcp_tool_id)
body = dict(self.workflow_manage.get_body())
# Explicit AI-node selection is limited to knowledge bases associated with this workflow.
from system_manage.mcp_workflows import knowledge_scope
body['knowledge_id_list'] = knowledge_scope(self.workflow_manage.flow)
engine_context = context_for(body, engine_tools=enabled,
tool_ids=ids if mcp_enable and mcp_source not in (None, 'custom') else [],
custom=self.handle_variables(json.loads(mcp_servers) if isinstance(mcp_servers, str) else mcp_servers) if mcp_enable and mcp_source in (None, 'custom') and mcp_servers else None,
transform=self.handle_variables, node_id=self.id)
if tool_enable:
if tool_ids and len(tool_ids) > 0: # 如果有工具ID,则将其转换为MCP
@ -248,8 +242,8 @@ class BaseChatNode(IChatNode):
self.context['execute_ids'].append(_id)
mcp_servers_config[str(tool.id)] = tool_config
if len(mcp_servers_config) > 0:
r = mcp_response_generator(chat_model, message_list, json.dumps(mcp_servers_config), mcp_output_enable)
if mcp_servers_config or engine_context['tool_ids'] or engine_context['custom'] or enabled:
r = mcp_response_generator(chat_model, message_list, json.dumps(mcp_servers_config), mcp_output_enable, engine_context)
return NodeResult(
{'result': r, 'chat_model': chat_model, 'message_list': message_list,
'history_message': [{'content': message.content, 'role': message.type} for message in
@ -315,7 +309,9 @@ class BaseChatNode(IChatNode):
code_path = f'{executor.sandbox_path}/execute/{tool_id}.py'
if os.path.exists(code_path):
os.remove(code_path)
from system_manage.mcp_agent import call_records
return {
'engine_calls': call_records(self.workflow_manage.get_body(), self.id),
'name': self.node.properties.get('stepName'),
"index": index,
'run_time': self.context.get('run_time'),

View File

@ -9,12 +9,16 @@ from application.flow.i_step_node import INode, NodeResult
class McpNodeSerializer(serializers.Serializer):
mcp_servers = serializers.JSONField(required=True, label=_("Mcp servers"))
mcp_servers = serializers.JSONField(required=False, default=dict, label=_("Mcp servers"))
mcp_server = serializers.CharField(required=True, label=_("Mcp server"))
mcp_tool = serializers.CharField(required=True, label=_("Mcp tool"))
mcp_tool_id = serializers.CharField(required=False, label=_("Mcp tool"), allow_null=True, allow_blank=True)
mcp_source = serializers.CharField(required=False, label=_("Mcp source"), allow_blank=True, allow_null=True)
tool_params = serializers.DictField(required=True, label=_("Tool parameters"))
tool_form_field = serializers.ListField(required=False, default=list)
timeout = serializers.IntegerField(default=60, min_value=1, max_value=600)
retries = serializers.IntegerField(default=0, min_value=0, max_value=2)
priority = serializers.IntegerField(default=5, min_value=0, max_value=9)
class IMcpNode(INode):

View File

@ -1,75 +1,59 @@
# coding=utf-8
import asyncio
import json
from copy import deepcopy
from typing import List
from django.db.models import QuerySet
from langchain_mcp_adapters.client import MultiServerMCPClient
from rest_framework.exceptions import ValidationError
from application.flow.i_step_node import NodeResult
from application.flow.step_node.mcp_node.i_mcp_node import IMcpNode
from tools.models import Tool
from application.flow.step_node.component_node.base_component_node import workflow_identity
from system_manage import mcp_runtime
class BaseMcpNode(IMcpNode):
def save_context(self, details, workflow_manage):
self.context['result'] = details.get('result')
self.context['tool_params'] = details.get('tool_params')
self.context['mcp_tool'] = details.get('mcp_tool')
self.context.update({key: details.get(key) for key in ('result', 'text', 'structured', 'tool_params', 'mcp_tool', 'mcp_server', 'attempts', 'task_id')})
def execute(self, mcp_servers, mcp_server, mcp_tool, mcp_tool_id, mcp_source, tool_params,**kwargs) -> NodeResult:
if mcp_source == 'referencing':
if not mcp_tool_id:
raise ValueError("MCP tool ID is required when mcp_source is 'referencing'.")
tool = QuerySet(Tool).filter(id=mcp_tool_id).first()
if not tool:
raise ValueError(f"Tool with ID {mcp_tool_id} not found.")
if not tool.is_active:
raise ValueError(f"Tool with ID {mcp_tool_id} is inactive.")
servers = json.loads(tool.code)
servers = self.handle_variables(servers) # 处理servers中的变量
params = json.loads(json.dumps(tool_params))
params = self.handle_variables(params)
else:
servers = json.loads(mcp_servers)
servers = self.handle_variables(servers) # 处理servers中的变量
params = json.loads(json.dumps(tool_params))
params = self.handle_variables(params)
def execute(self, mcp_servers, mcp_server, mcp_tool, tool_params, mcp_tool_id=None, mcp_source=None,
tool_form_field=None, timeout=60, retries=0, priority=5, **kwargs):
body, user = workflow_identity(self.workflow_manage)
params = self.handle_variables(deepcopy(tool_params), tool_form_field or [])
key = 'mcp:' + str(mcp_tool_id) if mcp_source == 'referencing' else 'mcp-custom'
result = mcp_runtime.invoke(body['workspace_id'], user, key,
{'servers': self.handle_variables(mcp_runtime.connections(mcp_servers)) if key == 'mcp-custom' else None,
'server': mcp_server, 'tool': mcp_tool, 'arguments': params},
parent=body.get('engine_task_id'), application_id=body['application_id'], chat_id=body.get('chat_id'),
scope={**body, 'node_id': self.id, 'source': 'workflow-mcp'}, timeout=timeout, retries=retries, priority=priority)
return NodeResult({**result, 'tool_params': params}, {})
async def call_tool(t, a):
client = MultiServerMCPClient(servers)
async with client.session(mcp_server) as s:
return await s.call_tool(t, a)
res = asyncio.run(call_tool(mcp_tool, params))
return NodeResult(
{'result': [content.text for content in res.content], 'tool_params': params, 'mcp_tool': mcp_tool}, {})
def handle_variables(self, tool_params):
# 处理参数中的变量
for k, v in tool_params.items():
if type(v) == str:
tool_params[k] = self.workflow_manage.generate_prompt(tool_params[k])
if type(v) == dict:
self.handle_variables(v)
if (type(v) == list) and (type(v[0]) == str):
tool_params[k] = self.get_reference_content(v)
return tool_params
def handle_variables(self, value, fields=None):
if isinstance(value, dict):
output = {}
for key, item in value.items():
field = next((f for f in fields or [] if f.get('field') == key), {})
if field.get('source') == 'custom' and field.get('input_type') == 'JsonInput' and isinstance(item, str):
try: output[key] = json.loads(self.workflow_manage.generate_prompt(item))
except ValueError: raise ValidationError('工具参数必须为有效 JSON') from None
elif field.get('source') == 'custom':
output[key] = self.workflow_manage.generate_prompt(item) if isinstance(item, str) else item
elif field.get('source') == 'referencing':
if not isinstance(item, list) or len(item) < 2: raise ValidationError('请选择工具参数的上游变量')
output[key] = self.get_reference_content(item)
else:
output[key] = self.handle_variables(item, fields)
return output
if isinstance(value, str): return self.workflow_manage.generate_prompt(value)
if isinstance(value, list):
known = set(self.workflow_manage.flow.node_map) | {'global', 'chat', 'loop'}
if len(value) >= 2 and all(isinstance(v, str) for v in value) and value[0] in known:
return self.get_reference_content(value)
return [self.handle_variables(item, fields) for item in value]
return value
def get_reference_content(self, fields: List[str]):
return self.workflow_manage.get_reference_field(
fields[0],
fields[1:])
return self.workflow_manage.get_reference_field(fields[0], fields[1:])
def get_details(self, index: int, **kwargs):
return {
'name': self.node.properties.get('stepName'),
"index": index,
'run_time': self.context.get('run_time'),
'status': self.status,
'err_message': self.err_message,
'type': self.node.type,
'mcp_tool': self.context.get('mcp_tool'),
'tool_params': self.context.get('tool_params'),
'result': self.context.get('result'),
}
return {'name': self.node.properties.get('stepName'), 'index': index,
'run_time': self.context.get('run_time'), 'status': self.status,
'err_message': self.err_message, 'type': self.node.type,
**{key: self.context.get(key) for key in ('mcp_tool', 'mcp_server', 'tool_params', 'result', 'text', 'structured', 'attempts', 'task_id')}}

View File

@ -13,7 +13,7 @@ import threading
from typing import Iterator
from django.http import StreamingHttpResponse
from langchain_core.messages import BaseMessageChunk, BaseMessage, ToolMessage, AIMessageChunk
from langchain_core.messages import BaseMessageChunk, BaseMessage, ToolMessage, AIMessageChunk, AIMessage, SystemMessage
from langchain_mcp_adapters.client import MultiServerMCPClient
from langgraph.prebuilt import create_react_agent
from application.flow.i_step_node import WorkFlowPostHandler
@ -267,17 +267,23 @@ def get_global_loop():
return _global_loop
async def _yield_mcp_response(chat_model, message_list, mcp_servers, mcp_output_enable=True):
client = MultiServerMCPClient(json.loads(mcp_servers))
tools = await client.get_tools()
async def _yield_mcp_response(chat_model, message_list, mcp_servers, mcp_output_enable=True, engine_context=None):
servers = json.loads(mcp_servers)
tools = await MultiServerMCPClient(servers).get_tools() if servers else []
if engine_context:
from system_manage.mcp_agent import tools as engine_tools
tools += await engine_tools(engine_context)
agent = create_react_agent(chat_model, tools)
response = agent.astream({"messages": message_list}, stream_mode='messages')
messages = message_list
if not any(isinstance(message, SystemMessage) for message in message_list):
messages = [SystemMessage(content='你是应用助手。根据用户请求选择已授权工具。需要外部数据或生成图谱时实际调用工具,不要编造工具结果。工具执行完成后依据结果回答,保存导图后返回打开链接,避免重复调用。'), *message_list]
response = agent.astream({"messages": messages}, stream_mode='messages')
# 用于存储工具调用信息
tool_calls_info = {}
async for chunk in response:
if isinstance(chunk[0], AIMessageChunk):
if isinstance(chunk[0], (AIMessageChunk, AIMessage)):
tool_calls = chunk[0].additional_kwargs.get('tool_calls', [])
for tool_call in tool_calls:
tool_id = tool_call.get('id', '')
@ -287,6 +293,10 @@ async def _yield_mcp_response(chat_model, message_list, mcp_servers, mcp_output_
'name': tool_call.get('function', {}).get('name', ''),
'input': tool_call.get('function', {}).get('arguments', '')
}
for tool_call in chunk[0].tool_calls:
if tool_call.get('id'):
tool_calls_info[tool_call['id']] = {'name': tool_call['name'],
'input': json.dumps(tool_call.get('args', {}), ensure_ascii=False)}
yield chunk[0]
if mcp_output_enable and isinstance(chunk[0], ToolMessage):
@ -303,14 +313,30 @@ async def _yield_mcp_response(chat_model, message_list, mcp_servers, mcp_output_
yield chunk[0]
def mcp_response_generator(chat_model, message_list, mcp_servers, mcp_output_enable=True):
def collect_mcp_response(response):
"""Use the same tool stream for the non-streaming chat API."""
from langchain_core.messages import AIMessage
parts, reasoning, metadata = [], [], {}
for chunk in response:
if isinstance(chunk.content, str):
parts.append(chunk.content)
else:
parts.extend(item.get('text', '') for item in chunk.content if isinstance(item, dict))
metadata.update(chunk.response_metadata)
value = chunk.additional_kwargs.get('reasoning_content')
if value: reasoning.append(value)
if reasoning: metadata['reasoning_content'] = ''.join(reasoning)
return AIMessage(content=''.join(parts), response_metadata=metadata)
def mcp_response_generator(chat_model, message_list, mcp_servers, mcp_output_enable=True, engine_context=None):
"""使用全局事件循环,不创建新实例"""
result_queue = queue.Queue()
loop = get_global_loop() # 使用共享循环
async def _run():
try:
async_gen = _yield_mcp_response(chat_model, message_list, mcp_servers, mcp_output_enable)
async_gen = _yield_mcp_response(chat_model, message_list, mcp_servers, mcp_output_enable, engine_context)
async for chunk in async_gen:
result_queue.put(('data', chunk))
except Exception as e:

View File

@ -151,6 +151,8 @@ class ModelKnowledgeAssociation(serializers.Serializer):
class ModelSettingSerializer(serializers.Serializer):
engine_tools = serializers.ListField(required=False, child=serializers.ChoiceField(
choices=['knowledge-search', 'mindmap-save']), max_length=2)
prompt = serializers.CharField(required=False, allow_null=True, allow_blank=True, max_length=102400,
label=_("Prompt word"))
system = serializers.CharField(required=False, allow_null=True, allow_blank=True, max_length=102400,
@ -677,21 +679,8 @@ class ApplicationOperateSerializer(serializers.Serializer):
if with_valid:
self.is_valid(raise_exception=True)
McpServersSerializer(data=instance).is_valid(raise_exception=True)
servers = json.loads(instance.get('mcp_servers'))
for server, config in servers.items():
if config.get('transport') not in ['sse', 'streamable_http']:
raise AppApiException(500, _('Only support transport=sse or transport=streamable_http'))
tools = []
for server in servers:
tools += [
{
'server': server,
'name': tool.name,
'description': tool.description,
'args_schema': tool.args_schema,
}
for tool in asyncio.run(get_mcp_tools({server: servers[server]}))]
return tools
from system_manage.mcp_runtime import discover
return discover(instance.get('mcp_servers'))
def delete(self, with_valid=True):
if with_valid:

View File

@ -136,7 +136,7 @@ class OpenView(APIView):
enforce(request, 'application', application_id)
return result.success(OpenChatSerializers(
data={'workspace_id': workspace_id, 'application_id': application_id,
'chat_user_id': str(uuid.uuid7()), 'chat_user_type': ChatUserType.ANONYMOUS_USER,
'chat_user_id': str(request.user.id), 'chat_user_type': ChatUserType.PLATFORM_USER.value,
'debug': True}).open())

View File

@ -8,6 +8,9 @@
"""
import os
import logging
import threading
from pathlib import Path
class MKTokenizer:
def __init__(self, tokenizer):
@ -17,14 +20,38 @@ class MKTokenizer:
return self.tokenizer.encode(text).ids
class ByteTokenEstimate:
"""Conservative token budget when an offline tokenizer is unavailable.
Provider-reported usage remains authoritative. This only supplies local
estimates for prompt budgeting; it never downloads model files.
"""
def encode(self, text):
return list(text.encode('utf-8'))
class TokenizerManage:
tokenizer = None
_lock = threading.Lock()
@staticmethod
def get_tokenizer():
from tokenizers import Tokenizer
# 创建Tokenizer
model_path = os.path.join("/opt/maxkb-app", "model", "tokenizer", "models--bert-base-cased")
with open(f"{model_path}/refs/main", encoding="utf-8") as f: snapshot = f.read()
TokenizerManage.tokenizer = Tokenizer.from_file(f"{model_path}/snapshots/{snapshot}/tokenizer.json")
return MKTokenizer(TokenizerManage.tokenizer)
if TokenizerManage.tokenizer is not None:
return TokenizerManage.tokenizer
with TokenizerManage._lock:
if TokenizerManage.tokenizer is not None:
return TokenizerManage.tokenizer
from tokenizers import Tokenizer
model_path = Path(os.environ.get('MAXKB_TOKENIZER_PATH') or
'/opt/maxkb-app/model/tokenizer/models--bert-base-cased')
try:
path = model_path if model_path.is_file() else model_path / 'tokenizer.json'
if not path.is_file():
snapshot = (model_path / 'refs/main').read_text(encoding='utf-8').strip()
path = model_path / 'snapshots' / snapshot / 'tokenizer.json'
TokenizerManage.tokenizer = MKTokenizer(Tokenizer.from_file(str(path)))
except Exception:
# A missing/corrupt tokenizer must not prevent remote-model chat.
logging.getLogger(__name__).warning('本地分词器不可用,使用保守字节估算;实际用量以模型返回为准')
TokenizerManage.tokenizer = ByteTokenEstimate()
return TokenizerManage.tokenizer

View File

@ -54,7 +54,7 @@ class XlsxSplitHandle(BaseSplitHandle):
try:
if type(limit) is str:
limit = int(limit)
workbook = openpyxl.load_workbook(io.BytesIO(buffer))
workbook = openpyxl.load_workbook(io.BytesIO(buffer), data_only=True)
try:
image_dict: dict = xlsx_embed_cells_images(io.BytesIO(buffer))
save_image([item for item in image_dict.values()])

View File

@ -37,6 +37,7 @@ def validate_params(params):
def knowledge_params(knowledge):
params = validate_params((knowledge.meta or {}).get('embedding_params') or {})
params.setdefault('normalize', True)
return {key: value for key, value in params.items()
if key in {'batch_size', 'normalize', 'query_prefix', 'document_prefix'}}
@ -48,7 +49,7 @@ class ConfiguredEmbedding(Embeddings):
self.batch_size = params.get('batch_size') or getattr(model, 'batch_size', 10) or 10
def _vector(self, vector):
if not self.params.get('normalize'):
if not self.params.get('normalize', True):
return vector
norm = math.sqrt(sum(float(v) ** 2 for v in vector))
return [float(v) / norm for v in vector] if norm else vector

View File

@ -1057,7 +1057,7 @@ class DocumentSerializers(serializers.Serializer):
knowledge = QuerySet(Knowledge).filter(id=self.data.get('knowledge_id')).first()
segmentation = (knowledge.meta or {}).get('segmentation') if knowledge is not None else None
segmentation = segmentation or {}
default_limit = segmentation.get('chunk_size') or 4096
default_limit = segmentation.get('chunk_size') or 500
default_overlap = segmentation.get('chunk_overlap') or 0
limit = instance.get("limit", None) or default_limit
overlap = instance.get("overlap", None)
@ -1740,7 +1740,7 @@ class DocumentSerializers(serializers.Serializer):
data={'workspace_id': self.data.get('workspace_id'), 'knowledge_id': self.data.get('knowledge_id')}
)
split_serializer.is_valid(instance={'file': [file]}, raise_exception=True)
parsed_documents = split_serializer.file_to_paragraph(file, None, None, 4096)
parsed_documents = split_serializer.file_to_paragraph(file, None, None, 500)
paragraphs = []
for parsed_document in parsed_documents:
paragraphs.extend(parsed_document.get('content', []))

View File

@ -346,9 +346,12 @@ class KnowledgeSerializer(serializers.Serializer):
)
)
), with_search_one=True)
from knowledge.embedding_guard import embedding_change_reason
return {
**knowledge_dict,
'meta': json.loads(knowledge_dict.get('meta', '{}')),
'embedding_change_block_reason': embedding_change_reason(
QuerySet(Knowledge).get(id=self.data.get('knowledge_id'))),
'application_id_list': list(filter(
lambda application_id: all_application_list.__contains__(application_id),
[
@ -371,6 +374,23 @@ class KnowledgeSerializer(serializers.Serializer):
).exclude(id=knowledge.id).exists():
raise AppApiException(500, _('Knowledge base name duplicate!'))
KnowledgeEditRequest(data=instance).is_valid(knowledge=knowledge)
from knowledge.embedding_params import knowledge_params, validate_params
from knowledge.embedding_guard import embedding_change_reason
from types import SimpleNamespace
current_params = knowledge_params(knowledge)
try:
params = validate_params(instance['embedding_params']) if 'embedding_params' in instance else None
except ValueError as exc:
raise AppApiException(500, str(exc))
new_params = knowledge_params(SimpleNamespace(meta={'embedding_params': params})) if params is not None else current_params
model_changed = ('embedding_model_id' in instance and
str(instance['embedding_model_id'] or '') != str(knowledge.embedding_model_id or ''))
vector_params_changed = ({k: v for k, v in current_params.items() if k != 'batch_size'} !=
{k: v for k, v in new_params.items() if k != 'batch_size'})
if model_changed or vector_params_changed:
reason = embedding_change_reason(knowledge)
if reason:
raise AppApiException(500, reason)
if 'embedding_model_id' in instance:
knowledge.embedding_model_id = instance.get('embedding_model_id')
if "name" in instance:
@ -386,11 +406,6 @@ class KnowledgeSerializer(serializers.Serializer):
incoming_meta[managed_key] = knowledge.meta[managed_key]
knowledge.meta = incoming_meta
if 'embedding_params' in instance:
from knowledge.embedding_params import validate_params
try:
params = validate_params(instance['embedding_params'])
except ValueError as exc:
raise AppApiException(500, str(exc))
knowledge.meta = {**(knowledge.meta or {}), 'embedding_params': params}
if 'folder_id' in instance:
knowledge.folder_id = instance.get('folder_id')

View File

@ -166,7 +166,7 @@ class KnowledgeSegmentationSerializer(serializers.Serializer):
保存后新建文档上传时默认使用该规则,上传时可覆盖。
"""
DEFAULTS = {"chunk_size": 4096, "chunk_overlap": 0}
DEFAULTS = {"chunk_size": 500, "chunk_overlap": 0}
@staticmethod
def one(workspace_id: str, knowledge_id: str):

View File

@ -54,19 +54,29 @@ class EmbeddingParamsTests(unittest.TestCase):
params = {'领域': '电力', 'dimensions': 512, '备注': '', 'batch_size': 2, 'query_prefix': '检索:'}
self.assertEqual(validate_params(params), params)
active = knowledge_params(SimpleNamespace(meta={'embedding_params': params}))
self.assertEqual(active, {'batch_size': 2, 'query_prefix': '检索:'})
self.assertEqual(active, {'batch_size': 2, 'query_prefix': '检索:', 'normalize': True})
base = RecordingModel()
model = ConfiguredEmbedding(base, active)
model.embed_query('问题')
self.assertEqual(base.calls, [('query', '检索:问题')])
def test_legacy_and_empty_params_keep_original_behavior(self):
self.assertEqual(knowledge_params(SimpleNamespace(meta={})), {})
def test_missing_normalization_defaults_to_unit_vectors(self):
self.assertEqual(knowledge_params(SimpleNamespace(meta={})), {'normalize': True})
self.assertEqual(knowledge_params(SimpleNamespace(meta={'embedding_params': {'normalize': None}})),
{'normalize': True})
self.assertEqual(validate_params({'batch_size': None, 'query_prefix': ''}), {})
base = RecordingModel()
self.assertEqual(ConfiguredEmbedding(base, {}).embed_query('x'), [3., 4.])
self.assertEqual(ConfiguredEmbedding(base, {}).embed_query('x'), [.6, .8])
self.assertEqual(ConfiguredEmbedding(base, {}).embed_documents(['x']), [[.6, .8]])
self.assertEqual(ConfiguredEmbedding(base, {}).embed_documents([]), [])
def test_explicit_false_keeps_original_vectors(self):
params = knowledge_params(SimpleNamespace(meta={'embedding_params': {'normalize': False}}))
self.assertEqual(params, {'normalize': False})
model = ConfiguredEmbedding(RecordingModel(), params)
self.assertEqual(model.embed_query('x'), [3., 4.])
self.assertEqual(model.embed_documents(['x']), [[3., 4.]])
def test_zero_vector_is_safe(self):
base = RecordingModel()
base.embed_query = lambda _: [0., 0.]

View File

@ -152,6 +152,11 @@ class MetadataTests(TestCase):
def test_upload_inherits_defaults_and_accepts_explicit_override(self):
from knowledge.serializers.document import DocumentSerializers
self.assertEqual(Rules.one('default', str(self.k.id))['chunk_size'], 500)
serializer = DocumentSerializers.Split(data={'knowledge_id':str(self.k.id),'workspace_id':'default'})
with patch.object(serializer,'file_to_paragraph',return_value=[]) as split:
serializer.parse({'file':[SimpleUploadedFile('default.txt',b'test')]})
self.assertEqual(split.call_args.args[-2:], (500,0))
Rules.update('default', str(self.k.id), self.user, {'chunk_size':100,'chunk_overlap':20})
serializer = DocumentSerializers.Split(data={'knowledge_id':str(self.k.id),'workspace_id':'default'})
with patch.object(serializer,'file_to_paragraph',return_value=[]) as split:

View File

@ -97,8 +97,11 @@ class MaxKBBaseModel(ABC):
@staticmethod
def filter_optional_params(model_kwargs):
from system_manage.summary_plan import LOCAL_PARAMETERS
optional_params = {}
for key, value in model_kwargs.items():
if key in LOCAL_PARAMETERS and key != 'timeout':
continue
if key not in ['model_id', 'use_local', 'streaming', 'show_ref_label', 'stream']:
if key == 'extra_body' and isinstance(value, dict):
optional_params = {**optional_params, **value}

View File

@ -14,6 +14,9 @@ class BaiLianChatModel(MaxKBBaseModel, BaseChatOpenAI):
@staticmethod
def new_instance(model_type, model_name, model_credential: Dict[str, object], **model_kwargs):
optional_params = MaxKBBaseModel.filter_optional_params(model_kwargs)
# Client controls belong to the SDK, not the provider's JSON body.
client_params = {key: optional_params.pop(key) for key in
('temperature', 'top_p', 'max_tokens', 'timeout', 'max_retries') if key in optional_params}
# if 'qwen-omni-turbo' in model_name or 'qwq' in model_name:
# optional_params['streaming'] = True
return BaiLianChatModel(
@ -21,5 +24,7 @@ class BaiLianChatModel(MaxKBBaseModel, BaseChatOpenAI):
openai_api_base=model_credential.get('api_base'),
openai_api_key=model_credential.get('api_key'),
streaming=True,
stream_usage=True,
**client_params,
extra_body=optional_params
)

View File

@ -36,6 +36,8 @@ class OllamaChatModel(MaxKBBaseModel, ChatOllama):
api_base = model_credential.get('api_base', '')
base_url = get_base_url(api_base)
optional_params = MaxKBBaseModel.filter_optional_params(model_kwargs)
from models_provider.context_policy import model_context
optional_params['num_ctx'] = model_context(model_kwargs)
return OllamaChatModel(model=model_name, base_url=base_url,
model_resource_id=str(model_kwargs['model_id']) if model_kwargs.get('model_id') else None,

View File

@ -1,11 +1,13 @@
"""Model file assets, independent of runtime/model connection records. Never executes model code."""
import json
import hashlib
import os
import shutil
import stat
import tempfile
import uuid
import zipfile
from contextlib import contextmanager
from datetime import datetime, timezone
from pathlib import Path, PurePosixPath
@ -18,6 +20,42 @@ class ImportError(ValueError):
pass
def model_fingerprint(directory):
"""Hash relative paths and complete file contents, independent of the archive name."""
root = Path(directory)
digest = hashlib.sha256()
for file in sorted((p for p in root.rglob('*') if p.is_file()), key=lambda p: p.relative_to(root).as_posix()):
name = file.relative_to(root).as_posix().encode('utf-8')
digest.update(len(name).to_bytes(8, 'big'))
digest.update(name)
digest.update(file.stat().st_size.to_bytes(8, 'big'))
with file.open('rb') as source:
for chunk in iter(lambda: source.read(1024 * 1024), b''):
digest.update(chunk)
return digest.hexdigest()
@contextmanager
def import_lock(root):
"""Serialize duplicate checks and record creation across Web processes."""
with (root / '.import.lock').open('a+b') as lock:
if os.name == 'nt':
import msvcrt
lock.seek(0)
msvcrt.locking(lock.fileno(), msvcrt.LK_LOCK, 1)
else:
import fcntl
fcntl.flock(lock.fileno(), fcntl.LOCK_EX)
try:
yield
finally:
if os.name == 'nt':
lock.seek(0)
msvcrt.locking(lock.fileno(), msvcrt.LK_UNLCK, 1)
else:
fcntl.flock(lock.fileno(), fcntl.LOCK_UN)
def inspect_model(directory):
root = Path(directory)
if not root.is_dir() or root.is_symlink():
@ -98,14 +136,33 @@ class ModelImportStore:
self.root = Path(root)
self.root.mkdir(parents=True, exist_ok=True)
def list(self):
def _records(self):
records = []
for path in self.root.glob('*/import.json'):
if path.parent.name.startswith('.import-'):
continue
try:
records.append(json.loads(path.read_text(encoding='utf-8')))
except (ValueError, OSError):
continue
return sorted(records, key=lambda r: r['created_at'], reverse=True)[:200]
return sorted(records, key=lambda r: r.get('created_at', ''), reverse=True)
def list(self):
records, paths, fingerprints = [], set(), set()
for record in self._records():
path = record.get('path')
canonical = os.path.normcase(os.path.realpath(path)) if path else None
fingerprint = record.get('fingerprint') if record.get('source') == 'UPLOAD' else None
if (canonical and canonical in paths) or (fingerprint and fingerprint in fingerprints):
continue
if canonical:
paths.add(canonical)
if fingerprint:
fingerprints.add(fingerprint)
records.append(record)
if len(records) == 200:
break
return records
def create(self, owner_id, path=None, upload=None):
if (path is None) == (upload is None):
@ -163,15 +220,37 @@ class ModelImportStore:
model_dir = Path(path).resolve(strict=True)
info = inspect_model(model_dir)
final_path, source_type, display_name = model_dir, 'PATH', model_dir.name
record = {
'id': identity, 'name': display_name, 'path': str(final_path),
'source': source_type, 'owner_id': str(owner_id),
'created_at': datetime.now(timezone.utc).isoformat(),
'status': 'IMPORTED', **info,
}
(staging / 'import.json').write_text(json.dumps(record, ensure_ascii=False), encoding='utf-8')
staging.rename(self.root / identity)
return record
with import_lock(self.root):
existing = self._records()
if source_type == 'PATH':
canonical = os.path.normcase(os.path.realpath(model_dir))
for record in existing:
if record.get('path') and os.path.normcase(os.path.realpath(record['path'])) == canonical:
return {**record, 'already_imported': True}
fingerprint = model_fingerprint(model_dir)
for record in existing:
if (record.get('size_bytes') != info['size_bytes'] or
record.get('file_count') != info['file_count']):
continue
previous_path = record.get('path')
if not previous_path or not Path(previous_path).is_dir():
continue
try:
inspect_model(previous_path)
previous = model_fingerprint(previous_path)
except (OSError, ImportError):
continue
if previous == fingerprint:
return {**record, 'already_imported': True}
record = {
'id': identity, 'name': display_name, 'path': str(final_path),
'source': source_type, 'owner_id': str(owner_id),
'created_at': datetime.now(timezone.utc).isoformat(),
'status': 'IMPORTED', 'fingerprint': fingerprint, **info,
}
(staging / 'import.json').write_text(json.dumps(record, ensure_ascii=False), encoding='utf-8')
staging.rename(self.root / identity)
return record
except (FileNotFoundError, PermissionError) as exc:
raise ImportError('目录不存在或无访问权限') from exc
finally:

View File

@ -105,5 +105,54 @@ class ImportTests(unittest.TestCase):
for kwargs in ({}, {'path':str(self.model),'upload':Upload([])}):
with self.assertRaises(ImportError):self.store.create('owner',**kwargs)
def test_repeated_path_and_historical_records_are_deduplicated(self):
first = self.store.create('owner', path=str(self.model))
repeated = self.store.create('owner', path=str(self.model / '..' / 'model'))
self.assertEqual(repeated['id'], first['id'])
self.assertTrue(repeated['already_imported'])
legacy = self.store.root / 'legacy'
legacy.mkdir()
(legacy / 'import.json').write_text(json.dumps({**first, 'id': 'legacy', 'created_at': '2000'}))
self.assertEqual([row['id'] for row in self.store.list()], [first['id']])
def test_zip_content_deduplicates_and_distinct_content_is_retained(self):
files = [('repo/config.json', '{"model_type":"bert"}'), ('repo/model.safetensors', b'weights')]
first = self.store.create('owner', upload=Upload(files))
second = self.store.create('owner', upload=Upload(list(reversed(files))))
self.assertEqual(first['id'], second['id'])
self.assertTrue(second['already_imported'])
different = self.store.create('owner', upload=Upload([
('repo/config.json', '{"model_type":"bert"}'), ('repo/model.safetensors', b'weights2')]))
self.assertNotEqual(first['id'], different['id'])
self.assertEqual(len(self.store.list()), 2)
def test_different_directory_and_zip_with_same_files_reuse_record(self):
first = self.store.create('owner', path=str(self.model))
copy = self.root / 'copy'
import shutil
shutil.copytree(self.model, copy)
second = self.store.create('owner', path=str(copy))
self.assertEqual(first['id'], second['id'])
zipped = Upload([('repo/config.json', (copy / 'config.json').read_bytes()),
('repo/model.safetensors', (copy / 'model.safetensors').read_bytes())])
self.assertEqual(self.store.create('owner', upload=zipped)['id'], first['id'])
def test_concurrent_imports_create_one_record(self):
from concurrent.futures import ThreadPoolExecutor
with ThreadPoolExecutor(max_workers=4) as pool:
results = list(pool.map(lambda _: self.store.create('owner', path=str(self.model)), range(4)))
self.assertEqual(len({row['id'] for row in results}), 1)
self.assertEqual(len(self.store._records()), 1)
def test_invalid_legacy_directory_does_not_block_import(self):
first = self.store.create('owner', path=str(self.model))
config = self.model / 'config.json'
config.write_text('x' * len(config.read_text()))
replacement = self.root / 'replacement'
replacement.mkdir()
(replacement / 'config.json').write_text(json.dumps({'model_type': 'bert', 'transformers_version': '4.45'}))
(replacement / 'model.safetensors').write_bytes(b'test fixture, not runtime weights')
self.assertNotEqual(self.store.create('owner', path=str(replacement))['id'], first['id'])
if __name__ == '__main__':unittest.main()

View File

@ -131,11 +131,15 @@ def get_model_default_params(model):
return value
return value
return {
params = {
p.get('field'): convert_to_int(p.get('default_value'))
for p in model.model_params_form
if p.get('default_value') is not None
}
if model.model_type == 'LLM':
from models_provider.context_policy import model_context
params['context_tokens'] = model_context(params)
return params
def get_model_instance_by_model_workspace_id(model_id, workspace_id, **kwargs):
@ -147,4 +151,8 @@ def get_model_instance_by_model_workspace_id(model_id, workspace_id, **kwargs):
"""
model = get_model_by_id(model_id, workspace_id)
s = get_model_default_params(model)
return ModelManage.get_model(model_id, lambda _id: get_model(model, **{**s, **kwargs}))
params = {**s, **kwargs}
if model.model_type == 'LLM' and any(kwargs.get(key) is not None for key in ('context_tokens', 'num_ctx')):
from models_provider.context_policy import model_context
params['context_tokens'] = model_context(kwargs)
return ModelManage.get_model(model_id, lambda _id: get_model(model, **params))

View File

@ -144,19 +144,23 @@ def serialize(obj, detail=False, user=None):
elif isinstance(obj, AIFlow):
data.update(revision=obj.revision, nodes=obj.nodes)
elif isinstance(obj, AIService):
from system_manage.ai_hub_service_api import get_config
data.update(flow_id=str(obj.flow_id), enabled=obj.enabled, allowed_users=obj.allowed_users,
requests_per_minute=obj.requests_per_minute,
key_configured=bool(obj.key_hash), url=f'ai_hub/published/{obj.id}/run')
key_configured=bool(obj.key_hash), url=f'ai_hub/published/{obj.id}/run', api_config=get_config(obj.id))
elif isinstance(obj, AITask):
from system_manage.ai_hub_progress import task_progress
data.update(status=obj.status, steps=obj.steps, error=obj.error, attempt=obj.attempt,
started_at=obj.started_at, finished_at=obj.finished_at)
if detail:
data.update(snapshot=obj.snapshot, result=obj.result, progress=task_progress(obj))
if any(n.get('type') == 'organize' for n in obj.snapshot.get('nodes', [])):
data['summary_input'] = {key: obj.input_data[key] for key in ('markdown', 'heading_structure') if key in obj.input_data}
elif isinstance(obj, AIArtifact):
data.update(revision=obj.revision, layout=obj.layout, task_id=str(obj.task_id) if obj.task_id else None)
if detail:
data.update(original=obj.original, markdown=obj.markdown, history=obj.history)
from system_manage.heading_structure import artifact_state, version_history, summary_state
data.update(original=obj.original, markdown=obj.markdown, history=version_history(obj), heading_structure=artifact_state(obj), summary_state=summary_state(obj))
return data
@ -251,22 +255,24 @@ def validate_nodes(workspace, user, nodes):
raise ValidationError('解析服务已停用')
clean['resource_id'] = str(parser.id)
if kind == 'organize':
if node.get('operation') not in (None, 'heading_repair'):
raise ValidationError('整理操作无效')
if node.get('operation'):
clean['operation'] = node['operation']
clean['model_id'] = str(model_for(workspace, user, node.get('model_id')).id)
clean['prompt'] = text(node.get('prompt', '保留事实、数值和原文语言,将以下文档整理成分级Markdown标题和要点。'), '提示词', limit=16000)
temperature = node.get('temperature')
if temperature is not None:
if isinstance(temperature, bool) or not isinstance(temperature, (int, float)) or not 0 <= temperature <= 2:
raise ValidationError('温度无效')
clean['temperature'] = temperature
# Absent parameters fall back to the per-model settings in model management.
for field, low, high in (
('context_tokens', 4096, 262144), ('output_tokens', 256, 32768),
('safety_tokens', 256, 8192), ('timeout', 30, 600)):
value = node.get(field)
if value is not None:
clean[field] = integer(value, low, high, '模型参数')
if clean.get('output_tokens', 4096) + clean.get('safety_tokens', 1024) >= clean.get('context_tokens', 32768):
raise ValidationError('输出与预留长度不能超过模型上下文')
from system_manage.summary_plan import validate_options
from system_manage.ai_hub_organize import resolve_settings
from models_provider.tools import get_model_default_params
from system_manage.mcp_business import summary_model_defaults, summary_supported_parameters
try:
clean.update(validate_options(node))
model = model_for(workspace, user, clean['model_id'])
resolve_settings(summary_model_defaults(model, get_model_default_params(model)), clean)
if clean.get('top_p') is not None and 'top_p' not in summary_supported_parameters(model):
raise ValueError('该模型未声明支持 top_p 参数')
except ValueError as exc:
raise ValidationError(str(exc)) from None
result.append(clean)
return result
@ -286,10 +292,14 @@ def service_create(workspace, user, data):
require_admin(user)
flow = one('flows', data.get('flow_id'), workspace, user, lock=True)
validate_nodes(workspace, user, flow.nodes)
return AIService.objects.create(workspace_id=workspace, owner_id=user.id,
obj = AIService.objects.create(workspace_id=workspace, owner_id=user.id,
name=text(data.get('name'), '名称', limit=128), flow=flow,
allowed_users=users(data.get('allowed_users', [])),
requests_per_minute=integer(data.get('requests_per_minute', 20), 1, 600, '调用频率'))
if 'api_config' in data:
from system_manage.ai_hub_service_api import save_config
save_config(obj.id, data['api_config'])
return obj
def make_snapshot(workspace, user, nodes):
@ -324,6 +334,10 @@ def submit(workspace, user, data, upload=None, service=None, request=None):
else:
nodes = one('flows', data['flow_id'], workspace, user, lock=True).nodes if data.get('flow_id') else data.get('nodes')
snapshot = make_snapshot(workspace, user, nodes)
from system_manage.mcp_registry import pipeline_resources, resolve
required = pipeline_resources(snapshot)
for key in required: resolve(workspace, user, key)
priority = integer(data.get('priority', 5), 0, 9, '优先级')
from system_manage.access_control import enforce_task
enforce_task(request, service.id if service else None, snapshot)
markdown = text(data.get('markdown', ''), '文档内容', required=False, limit=MAX_INLINE_TEXT)
@ -338,33 +352,98 @@ def submit(workspace, user, data, upload=None, service=None, request=None):
from system_manage.ai_hub_assets import asset_lock, normalize
with asset_lock():
markdown = text(normalize(markdown, workspace, user), '文档内容', required=False)
heading_input = {}
if data.get('heading_structure') is not None:
from system_manage.heading_structure import analyze, validate_manual
if not isinstance(data['heading_structure'], dict):
raise ValidationError('层级数据无效')
try:
_, structure = analyze(markdown, previous=data['heading_structure'], repair=False)
structure = validate_manual(structure, data['heading_structure'])
except ValueError as exc:
raise ValidationError(str(exc)) from None
heading_input = {'heading_structure': structure, 'original': text(data.get('original', markdown), '原稿', required=False)}
if data.get('_heading_source') is not None:
heading_input['heading_source'] = data['_heading_source']
# Plan without invoking a model. Reserve only useful internal parallel calls.
from system_manage.mcp_registry import settings_row
from system_manage.mcp_business import summary_model_defaults, summary_counter
from system_manage.summary_plan import token_counter
from system_manage.ai_hub_organize import plan_document, resolve_settings
from models_provider.tools import get_model_default_params
from system_manage.ai_hub_assets import protect_images
scheduler_settings = settings_row().meta
capacity = scheduler_settings.get('capacity', 4)
worker_capacity = max(1, capacity - min(scheduler_settings.get('interactive_reserve', 1), max(0, capacity - 1)))
resource_slots = {}
for node in snapshot:
if node['type'] != 'organize' or node.get('operation') == 'heading_repair':
continue
key = 'model:' + node['model_id']
slots = min(node.get('concurrency', 1) if node.get('parallel', True) else 1,
resolve(workspace, user, key)[0]['capacity'], worker_capacity)
if snapshot[0]['type'] != 'parse':
model = model_for(workspace, user, node['model_id'])
defaults = summary_model_defaults(model, get_model_default_params(model))
try:
options = {**node, 'max_input_tokens': min(int(defaults.get('max_input_tokens', 2000000)), node.get('max_input_tokens', 2000000))}
if 'input_tokens' not in options and defaults.get('input_tokens'):
options['input_tokens'] = int(defaults['input_tokens'])
settings = resolve_settings(defaults, node)
count, _ = summary_counter(model, defaults, markdown, workspace, user.id, settings, options)
if node.get('structure_mode', 'preserve') in ('preserve', 'reorganize'):
from system_manage.summary_sections import build_plan
planned = build_plan(protect_images(markdown)[0], node['prompt'], settings, options, count)
slots = min(slots, max(1, len(planned['batches'])))
else:
_, _, _, _, chunks = plan_document(protect_images(markdown)[0], node['prompt'], settings, options, count)
slots = min(slots, len(chunks))
except ValueError as exc:
raise ValidationError(str(exc)) from None
resource_slots[key] = slots
return AITask.objects.create(workspace_id=workspace, owner_id=user.id,
name=text(data.get('name') or (upload.name if upload else '导图任务'), '名称', limit=128), service=service,
snapshot={'nodes': snapshot, 'flow_id': str(service.flow_id) if service else data.get('flow_id'),
snapshot={'nodes': snapshot, 'resources': required, 'resource_slots': resource_slots, 'priority': priority, 'flow_id': str(service.flow_id) if service else data.get('flow_id'),
'flow_revision': service.flow.revision if service else None},
input_data={'markdown': markdown}, file_name=upload.name[:255] if upload else '',
input_data={'markdown': markdown, **heading_input}, file_name=upload.name[:255] if upload else '',
file_data=upload.read() if upload else None)
@transaction.atomic
def artifact_rename(obj, data):
if set(data) != {'action', 'name', 'revision'}:
raise ValidationError('重命名参数无效')
if isinstance(data['revision'], bool) or not isinstance(data['revision'], int) or data['revision'] != obj.revision:
raise ValidationError('内容已更新,请重新打开')
obj.name = text(data['name'], '名称', limit=128)
obj.save(update_fields=['name', 'updated_at'])
return obj
@transaction.atomic
def artifact_save(workspace, user, data, obj=None):
from system_manage.heading_structure import analyze, artifact_state, store_artifact_state, store_summary_state, summary_state, validate_manual
previous_structure = artifact_state(obj) if obj else {}
if obj:
obj = AIArtifact.objects.select_for_update().get(pk=obj.pk)
previous_structure = artifact_state(obj)
if data.get('revision') != obj.revision:
raise ValidationError('内容已更新,请重新打开')
history = list(obj.history)
if len(history) >= 30:
raise ValidationError('修订已达30次,请另存为新图谱')
history.append({'revision': obj.revision, 'markdown': obj.markdown, 'layout': obj.layout,
'operator': str(user.id), 'updated_at': obj.updated_at.isoformat()})
obj.history = history
obj.revision += 1
else:
from users.models import User
from system_manage.artifact_settings import limit
User.objects.select_for_update().get(pk=user.id)
maximum = limit()
if AIArtifact.objects.filter(workspace_id=workspace, owner_id=user.id).count() >= maximum:
raise ValidationError(f'已达到导图保存上限({maximum}条),请删除后再保存')
task = one('tasks', data['task_id'], workspace, user, lock=True) if data.get('task_id') else None
if task and task.status != 'SUCCESS':
raise ValidationError('任务尚未成功')
obj = AIArtifact(workspace_id=workspace, owner_id=user.id, task=task,
original=task.result.get('original', '') if task else text(data.get('original', ''), '原始内容', False, MAX_INLINE_TEXT))
if task:
previous_structure = (task.result.get('source') or {}).get('heading_structure', {})
obj.name = text(data.get('name'), '名称', limit=128)
obj.markdown = text(data.get('markdown'), '修订内容', limit=MAX_INLINE_TEXT)
obj.layout = data.get('layout', 'tree')
@ -374,8 +453,34 @@ def artifact_save(workspace, user, data, obj=None):
with asset_lock():
obj.markdown = text(normalize(obj.markdown, workspace, user), '修订内容')
obj.original = text(normalize(obj.original, workspace, user), '原始内容', required=False)
if 'summary_state' in data:
state = data['summary_state']
if not isinstance(state, dict) or set(state) - {'source_markdown', 'smart_markdown', 'mode'}:
raise ValidationError('智能总结内容无效')
if state.get('mode', 'source') not in ('source', 'smart'):
raise ValidationError('智能总结状态无效')
clean_state = {'mode': state.get('mode', 'source')}
for key in ('source_markdown', 'smart_markdown'):
clean_state[key] = text(normalize(text(state.get(key, ''), '总结内容', required=False, limit=MAX_INLINE_TEXT), workspace, user), '总结内容', required=False, limit=MAX_INLINE_TEXT)
store_summary_state(obj, clean_state)
elif summary_state(obj):
state = dict(summary_state(obj))
state.pop('_type', None)
state['smart_markdown' if state.get('mode') == 'smart' else 'source_markdown'] = obj.markdown
store_summary_state(obj, state)
for entry in obj.history:
entry['markdown'] = normalize(entry.get('markdown', ''), workspace, user)
if 'markdown' in entry:
entry['markdown'] = normalize(entry['markdown'], workspace, user)
incoming = data.get('heading_structure')
if incoming is not None or previous_structure:
try:
_, structure = analyze(obj.markdown, previous=previous_structure or incoming, repair=False)
if incoming is not None:
structure = validate_manual(structure, incoming)
_, structure = analyze(obj.markdown, previous=structure, repair=False)
except ValueError as exc:
raise ValidationError(str(exc)) from None
store_artifact_state(obj, structure)
obj.save()
return obj
@ -407,6 +512,9 @@ def delete_record(kind, pk, workspace, user):
from system_manage.models.platform_access import PlatformGrant
PlatformGrant.objects.filter(workspace_id=workspace, resource_type=resource_type,
target=str(obj.id)).delete()
if kind == 'services':
from system_manage.ai_hub_service_api import save_config
save_config(obj.id)
obj.delete()
@ -419,9 +527,10 @@ def parse_file(parser, name, content, options=None, progress_callback=None):
return parse(parser, name, content, opts, headers, progress_callback=progress_callback)
def markdown_tree(markdown):
def markdown_tree(markdown, excluded_lines=None):
"""Markdown转稳定ID结构树:标题层级、嵌套列表、表格;围栏内容不参与结构。"""
import re
excluded_lines = set(excluded_lines or [])
root = {'id': 'root', 'text': '文档', 'children': []}
heading_stack = [(0, root)]
list_stack = []
@ -459,6 +568,8 @@ def markdown_tree(markdown):
list_pattern = re.compile(r'^(\s*)(?:([-*+])|(\d{1,9})[.)])\s+(.+)$')
for index, line in enumerate(markdown.splitlines()):
if excluded_lines and index in excluded_lines:
continue
fence_match = re.match(r'^\s{0,3}(`{3,}|~{3,})', line)
if fence_match:
token = fence_match.group(1)

View File

@ -172,7 +172,7 @@ def protect_images(markdown):
return pattern.sub(replace, markdown), images
def restore_images(markdown, images):
def restore_images(markdown, images, append_missing=True):
# Summarizers may omit or repeat placeholders. The backend owns the image
# manifest, so preserve known positions and append omitted originals once.
known = {str(index + 1): image for index, image in enumerate(images)}
@ -185,7 +185,7 @@ def restore_images(markdown, images):
return known[index]
markdown = re.sub(r'AIHUBIMAGE[\s_*`-]*(\d+)[\s_*`-]*TOKEN', replace, markdown, flags=re.I)
missing = [image for index, image in known.items() if index not in restored]
if missing:
if missing and append_missing:
markdown = markdown.rstrip() + '\n\n## 原文图片\n\n' + '\n\n'.join(missing)
return markdown

View File

@ -1,6 +1,41 @@
"""Bounded hierarchical organization; source and completed chunks survive failures."""
import re
import time
import hashlib
import json
import threading
from models_provider.context_policy import DEFAULT_CONTEXT_TOKENS, summary_defaults
from concurrent.futures import ThreadPoolExecutor, as_completed
from system_manage.summary_plan import LOCAL_PARAMETERS, prompt_rules, output_estimate, source_language, language_rule, language_matches, boundaries_match
ORGANIZER_PARAMETERS = LOCAL_PARAMETERS
def resolve_settings(defaults, params):
"""Workflow overrides win; model-management fields configure the organizer locally."""
defaults = summary_defaults(defaults)
def number(name, value, low, high):
if isinstance(value, str) and value.isdigit():
value = int(value)
if type(value) is not int or not low <= value <= high:
raise ValueError(f'总结参数 {name} 无效,请填写 {low} 至 {high} 的整数')
return value
context = number('context_tokens', params.get('context_tokens', defaults['context_tokens']), 1024, 2000000)
if context > int(defaults['context_tokens']):
raise ValueError('上下文超过模型配置上限,请先在模型管理中配置能力参数')
output = number('output_tokens', params.get('max_tokens', params.get('output_tokens',
defaults['output_tokens'])), 1, 524288)
safety = number('safety_tokens', params.get('safety_tokens', defaults['safety_tokens']), 0, 200000)
if output + safety >= context:
raise ValueError('总结输出与预留长度不能超过上下文长度')
for key, value in [('context_tokens', context), ('max_input_tokens', params.get('max_input_tokens')),
('max_output_tokens', max(output, params.get('max_output_tokens') or 0))]:
limit = defaults.get(key)
if limit is not None and value is not None and int(value) > int(limit):
raise ValueError(f'总结参数超过模型配置的 {key} 上限')
if params.get('max_output_tokens') is not None and output > params['max_output_tokens']:
raise ValueError('本次输出长度超过模型最大输出长度')
return {'context_tokens': context, 'output_tokens': output, 'safety_tokens': safety}
def token_bound(text):
@ -8,25 +43,31 @@ def token_bound(text):
return len(text.encode('utf-8'))
def split_markdown(text, budget):
def split_markdown(text, budget, count=None):
count = count or token_bound
if budget < 256:
raise ValueError('模型上下文预算不足,请调大上下文或减少输出长度')
chunks, current = [], ''
blocks = re.split(r'(?m)(?=^#{1,6} )', text)
# Do not mistake headings inside fenced examples for document sections.
protected = list(re.finditer(r'(?ms)^```[^\n]*\n.*?^```[^\n]*(?:\n|$)|^~~~[^\n]*\n.*?^~~~[^\n]*(?:\n|$)', text))
boundaries = [m.start() for m in re.finditer(r'(?m)^#{1,6} ', text)
if not any(span.start() <= m.start() < span.end() for span in protected)]
boundaries = sorted(set([0, *boundaries, len(text)]))
blocks = [text[a:b] for a, b in zip(boundaries, boundaries[1:])]
for block in blocks:
while block:
if token_bound(current + block) <= budget:
if count(current + block) <= budget:
current += block
break
if current:
chunks.append(current); current = ''
if token_bound(block) <= budget:
if count(block) <= budget:
current = block
break
lo, hi = 1, min(len(block), budget)
lo, hi = 1, len(block)
while lo < hi:
mid = (lo + hi + 1) // 2
if token_bound(block[:mid]) <= budget:
if count(block[:mid]) <= budget:
lo = mid
else:
hi = mid - 1
@ -34,6 +75,13 @@ def split_markdown(text, budget):
newline = block.rfind('\n', 0, end)
if newline > end // 2:
end = newline + 1
# Keep a table, fenced example, or consecutive list together when
# the whole block fits; oversized blocks may still be continued.
spans = re.finditer(r'(?ms)^```[^\n]*\n.*?^```[^\n]*(?:\n|$)|^~~~[^\n]*\n.*?^~~~[^\n]*(?:\n|$)|(?:^\|[^\n]*\n?)+|(?:^[ \t]*(?:[-*+] |\d+[.)] )[^\n]*\n?)+', block)
for span in spans:
if span.start() < end < span.end() and count(span[0]) <= budget:
end = span.start() if span.start() else span.end()
break
# Never split an image placeholder between requests.
for match in re.finditer(r'AIHUBIMAGE\s*\d+\s*TOKEN', block[:end + 64]):
if match.start() < end < match.end():
@ -47,65 +95,301 @@ def split_markdown(text, budget):
return chunks
def organize(markdown, prompt, invoke, *, context_tokens=32768, output_tokens=4096,
safety_tokens=1024, progress=None, checkpoint=None):
budget = context_tokens - output_tokens - safety_tokens - token_bound(prompt) - 512
chunks = split_markdown(markdown, budget)
def plan_document(markdown, prompt, settings, options, count=token_bound):
context_tokens, output_tokens, safety_tokens = (settings[key] for key in ('context_tokens', 'output_tokens', 'safety_tokens'))
instruction = prompt_rules(prompt, options, source_language(markdown, prompt))
# Includes bounded batch context, retry guidance, and message framing.
input_limit = min(context_tokens - output_tokens - safety_tokens,
options.get('max_input_tokens') or context_tokens)
budget = input_limit - count(instruction) - 2048
if budget < 256:
raise ValueError('模型上下文预算不足,请减少提示词或输出预留')
ratio = max(.05, output_estimate(markdown, count, options) / max(1, count(markdown)))
# Output risk can require splitting even when input fits. Never constrain
# every map result to a fraction of the final merge's output allowance.
output_budget = max(256, int(output_tokens * .8 / ratio))
chunk_budget = min(budget, output_budget)
chunks = split_markdown(markdown, chunk_budget, count)
if len(chunks) > 128:
raise ValueError('文档分块过多,请调大模型上下文或拆分文档')
calls = 0
raise ValueError('文档分块过多,请调大模型预算或拆分文档')
if len(chunks) > 1 and not options.get('auto_split', True):
raise ValueError('文档超过输入或输出预算,请开启自动拆分')
return instruction, input_limit, budget, chunk_budget, chunks
def call(content, instruction):
nonlocal calls
calls += 1
if calls > 160:
raise ValueError('整理批次过多,请拆分文档')
if token_bound(content) + token_bound(instruction) > context_tokens - output_tokens - safety_tokens:
def organize(markdown, prompt, invoke, *, context_tokens=DEFAULT_CONTEXT_TOKENS, output_tokens=None,
safety_tokens=None, progress=None, checkpoint=None, metrics=None,
options=None, count=None, count_method='conservative_bytes', resume=None,
state_checkpoint=None, fingerprint_context=None):
options = options or {}
count = count or token_bound
inherited = summary_defaults({'context_tokens': context_tokens})
output_tokens = inherited['output_tokens'] if output_tokens is None else output_tokens
safety_tokens = inherited['safety_tokens'] if safety_tokens is None else safety_tokens
settings = {'context_tokens': context_tokens, 'output_tokens': output_tokens, 'safety_tokens': safety_tokens}
instruction, input_limit, budget, chunk_budget, chunks = plan_document(markdown, prompt, settings, options, count)
merge = len(chunks) > 1 and options.get('structure_mode', 'preserve') == 'reorganize'
concurrency = min(len(chunks), options.get('concurrency', 1)) if options.get('parallel', True) else 1
stats = metrics if metrics is not None else {}
stats.update(context_tokens=context_tokens, output_tokens=output_tokens, safety_tokens=safety_tokens,
input_bytes=token_bound(markdown), input_tokens_estimate=count(markdown), token_count_method=count_method,
input_budget=budget, chunk_budget=chunk_budget, map_chunks=len(chunks), concurrency=concurrency,
needs_merge=merge, expected_calls=len(chunks) + int(merge), model_calls=0,
compression_retries=0, language_retries=0, structure_retries=0, fact_retries=0, language_splits=0, split_retries=0, reused_chunks=0,
estimated_output_tokens=output_estimate(markdown, count, options))
fingerprint = hashlib.sha256(json.dumps([markdown, instruction, context_tokens, output_tokens,
safety_tokens, chunk_budget, options, fingerprint_context, 'summary-v4'],
ensure_ascii=False, sort_keys=True).encode()).hexdigest()
previous = (resume or {}).get('outputs', {}) if (resume or {}).get('fingerprint') == fingerprint else {}
saved = {key: value for key, value in previous.items() if isinstance(value, str) and value.strip()}
lock = threading.RLock()
last_saved = None
def persist():
nonlocal last_saved
with lock:
if checkpoint and saved != last_saved:
checkpoint(list(saved.values()))
last_saved = dict(saved)
if state_checkpoint:
state_checkpoint({'fingerprint': fingerprint, 'outputs': dict(saved), 'metrics': dict(stats)})
def report(stage, percent):
if progress:
progress({'stage': stage, 'percent': percent, 'plan': dict(stats)})
def call(content, guidance):
if count(content) + count(guidance) + 512 > input_limit:
raise ValueError('模型输入超过上下文预算')
for attempt in range(3):
for attempt in range(options.get('retries', 2) + 1):
with lock:
stats['model_calls'] += 1
if stats['model_calls'] > 160:
raise ValueError('总结调用超过限制,已保留完成分块')
try:
response = invoke(instruction, content)
response = invoke(guidance, content)
break
except Exception as exc:
if attempt == 2:
if attempt == options.get('retries', 2):
raise ValueError('模型调用失败,已保留原稿和完成的分块,请稍后重试') from exc
time.sleep(attempt + 1)
metadata = getattr(response, 'response_metadata', {})
if isinstance(metadata, dict) and metadata.get('finish_reason') in ('length', 'max_tokens'):
raise ValueError('模型输出被截断,请到模型管理调大该模型的输出上限,或减少整理内容')
result = getattr(response, 'content', response)
if isinstance(result, list):
result = ''.join(item.get('text', '') for item in result if isinstance(item, dict))
if not isinstance(result, str) or not result.strip():
metadata = getattr(response, 'response_metadata', {}) or {}
truncated = (metadata.get('finish_reason') or metadata.get('done_reason')) in ('length', 'max_tokens', 'max_output_tokens')
value = getattr(response, 'content', response)
if isinstance(value, list):
value = ''.join(item.get('text', '') for item in value if isinstance(item, dict))
if not truncated and (not isinstance(value, str) or not value.strip()):
raise ValueError('模型未返回有效整理内容')
return result.strip()
value = value.strip() if isinstance(value, str) else ''
if value.startswith('```'):
lines = value.splitlines()
closing = next((i for i, line in enumerate(lines[1:], 1) if re.fullmatch(r'```\s*', line)), len(lines))
value = '\n'.join(lines[1:closing]).strip()
elif '```' not in content:
closing = re.search(r'(?m)^```\s*$', value)
if closing:
value = value[:closing.start()].strip()
usage = getattr(response, 'usage_metadata', {}) or metadata.get('token_usage', {}) or {}
actual_output = usage.get('output_tokens', usage.get('completion_tokens', metadata.get('eval_count')))
return value, truncated, actual_output
outputs = []
chunk_prompt = prompt + '\n这是文档的一个分块,保留事实及标题关系,精简重复内容。'
for index, chunk in enumerate(chunks, 1):
if progress:
progress({'stage': f'智能整理分块 {index}/{len(chunks)}', 'percent': (index - 1) * 70 // len(chunks)})
outputs.append(call(chunk, prompt if len(chunks) == 1 else chunk_prompt))
if checkpoint:
checkpoint(list(outputs))
if len(outputs) == 1:
return outputs[0]
merge_prompt = prompt + '\n合并以下局部大纲,统一标题层级,压缩重复内容;保留事实、数字和图片占位符。'
previous_size = None
for round_index in range(4):
joined = '\n\n'.join(outputs)
groups = split_markdown(joined, budget)
if previous_size is not None and token_bound(joined) >= previous_size and len(groups) > 1:
raise ValueError('合并结果未收敛,已保留完成的分块,请缩小文档或调整提示词')
previous_size = token_bound(joined)
merged = []
for index, group in enumerate(groups, 1):
if progress:
progress({'stage': f'合并大纲 {round_index + 1}:{index}/{len(groups)}', 'percent': 70 + round_index * 6})
merged.append(call(group, merge_prompt))
if checkpoint:
checkpoint(merged)
if len(merged) == 1:
return merged[0]
outputs = merged
raise ValueError('合并批次超过限制,已保留完成的分块')
def summarize(content, key, guidance, depth=0):
with lock:
if key in saved:
stats['reused_chunks'] += 1
return saved[key]
original = content
language = source_language(original, prompt)
language_failed = False
outline_source = re.sub(r'(?ms)^```[^\n]*\n.*?^```[^\n]*(?:\n|$)|^~~~[^\n]*\n.*?^~~~[^\n]*(?:\n|$)', '', original)
required_headings = re.findall(r'(?m)^(#{1,6}) (.+)$', outline_source) if options.get('structure_mode') == 'preserve' else []
outline = '\n'.join(hashes + ' ' + title for hashes, title in required_headings)
# An exceptionally large TOC would defeat the body budget; it is already
# present in the source. Include only bounded, exact outline guidance.
outline_hint = ('\nREQUIRED OUTPUT OUTLINE. Copy these headings EXACTLY and add summary bullets beneath each:\n' + outline
if outline and count(outline) <= 400 else '')
structure_failed = False
facts_failed = False
for attempt in range(2):
hint = guidance + f'\nOutput at most {output_tokens} tokens. Summarize concisely, do not repeat the full source.'
if attempt:
hint += '\nPrevious result failed validation. Return a concise, complete summary in the required language.'
with lock:
stats['language_retries' if language_failed else 'structure_retries' if structure_failed else 'fact_retries' if facts_failed else 'compression_retries'] += 1
hint += '\n' + language_rule(language)
hint += outline_hint
if language_failed or structure_failed:
hint += '\nVALIDATION FAILED. Keep ALL original section headings in order. Fill EACH section using its source language.'
if facts_failed:
hint += '\nNUMERIC CONDITION VALIDATION FAILED. Preserve all explicit comparison operators and conditions EXACTLY. Above/exceeds means >, not >= or reaches. Do not weaken or reverse a threshold.'
value, truncated, actual_output = call(content, hint)
source_sections = list(re.finditer(r'(?m)^(#{1,6}) (.+)$', outline_source)) if required_headings else []
target_sections = list(re.finditer(r'(?m)^(#{1,6}) (.+)$', value))
bodies = [outline_source[src.end():source_sections[j+1].start() if j+1 < len(source_sections) else len(outline_source)].strip()
for j, src in enumerate(source_sections)]
# A heading-only ancestor chain has one unambiguous body owner.
# Restore that skeleton in code instead of paying for a relabel call.
if (source_sections and sum(bool(body) for body in bodies) == 1 and bodies[-1] and target_sections and
all(not value[dst.end():target_sections[j+1].start()].strip() for j, dst in enumerate(target_sections[:-1]))):
last_body = value[target_sections[-1].end():].strip()
if (len(target_sections) <= len(source_sections) and
language_matches(last_body, source_language(bodies[-1], prompt)) and
set(re.findall(r'\d+(?:\.\d+)?', bodies[-1])) <= set(re.findall(r'\d+(?:\.\d+)?', last_body))):
value = '\n'.join(src[1] + ' ' + src[2] for src in source_sections) + '\n' + last_body
target_sections = list(re.finditer(r'(?m)^(#{1,6}) (.+)$', value))
aligned = bool(source_sections) and len(source_sections) == len(target_sections)
section_language_failed = False
if aligned:
# Headings are owned by the document. Restore harmless translated
# or reformatted labels only when section order, body languages,
# and numeric anchors agree; do not ask the model to relabel them.
for j, (src, dst) in enumerate(zip(source_sections, target_sections)):
body = outline_source[src.end():source_sections[j+1].start() if j+1 < len(source_sections) else len(outline_source)].strip()
summary = value[dst.end():target_sections[j+1].start() if j+1 < len(target_sections) else len(value)].strip()
anchors = set(re.findall(r'\d+(?:\.\d+)?', body))
if body and not language_matches(summary, source_language(body, prompt)):
section_language_failed = True
aligned = False
if not anchors <= set(re.findall(r'\d+(?:\.\d+)?', summary)):
aligned = False
if aligned:
for src, dst in reversed(list(zip(source_sections, target_sections))):
value = value[:dst.start()] + src[1] + ' ' + src[2] + value[dst.end():]
# With no model tokenizer use the conservative estimator, explicitly
# recorded in the plan. Provider finish_reason remains authoritative.
language_failed = section_language_failed or not language_matches(value, language)
facts_failed = not boundaries_match(original, value)
actual_headings = re.findall(r'(?m)^#{1,6} (.+)$', value)
remaining = iter(actual_headings)
heading_key = lambda title: re.sub(r'[*_`]', '', title).strip()
structure_failed = any(not any(heading_key(candidate) == heading_key(title) for candidate in remaining) for _, title in required_headings)
# Some small models translate mixed-language documents despite valid
# instructions. Retry only language-local sections, with no global
# rewrite. Models that succeed initially still use a single call.
body_languages = {source_language(body, prompt) for body in bodies if body}
if (attempt == 0 and depth < 2 and language == 'mixed' and (language_failed or structure_failed)
and len(body_languages) > 1 and body_languages <= {'english', 'chinese'}):
with lock:
stats['language_splits'] += 1
stats['expected_calls'] += sum(bool(body) for body in bodies)
parts = []
for j, (section, body) in enumerate(zip(source_sections, bodies)):
heading = section[1] + ' ' + section[2]
if not body:
parts.append(heading)
continue
section_language = source_language(body, prompt)
parts.append(summarize(heading + '\n' + body, key + '/language' + str(j),
prompt_rules(prompt, options, section_language) +
f'\nLanguage-local recovery section {j+1}/{len(source_sections)}. Only summarize this section.', depth+1))
value = '\n\n'.join(parts)
with lock:
saved[key] = value
return value
if not truncated and not language_failed and not structure_failed and not facts_failed and (actual_output if isinstance(actual_output, int) else count(value)) <= output_tokens:
# Restore original levels by ordered occurrence, including duplicate titles.
pending_headings = list(required_headings)
def restore_level(match):
if pending_headings and heading_key(match[2]) == heading_key(pending_headings[0][1]):
hashes, title = pending_headings.pop(0)
return hashes + ' ' + title
return match[0]
value = re.sub(r'(?m)^(#{1,6}) (.+)$', restore_level, value)
with lock:
saved[key] = value
return value
if language_failed or structure_failed or facts_failed:
content = original
elif not truncated and count(value) <= budget:
content = value
if language_failed:
raise ValueError('总结未保持原文语言,已保留完成分块;请调整提示词或更换模型后重试')
if structure_failed:
raise ValueError('总结未保留原文标题,已保留完成分块;请调整提示词或更换模型后重试')
if facts_failed:
raise ValueError('总结改变了数值边界条件,已保留原稿和完成分块;请调整提示词或更换模型后重试')
if depth >= 3 or count(original) <= 512:
raise ValueError('模型总结截断或长度校验未通过,已保留完成的分块,请提高输出上限或减少细节')
with lock:
stats['split_retries'] += 1
parts = split_markdown(original, max(256, (count(original) + 1) // 2), count)
values = [summarize(part, key + '/' + str(i), guidance +
f'\n本批继续拆分为 {len(parts)} 段,当前第 {i+1} 段,只总结当前正文。', depth + 1)
for i, part in enumerate(parts)]
value = '\n\n'.join(values)
with lock:
saved[key] = value
return value
report(f'预计 {len(chunks)} 批,并发 {concurrency},' + ('完成后汇总' if merge else '完成后组装'), 0)
outputs = [None] * len(chunks)
offset = 0
jobs = []
for i, chunk in enumerate(chunks):
prefix = markdown[:offset]
path = []
for match in re.finditer(r'(?m)^(#{1,6}) (.+)$', prefix):
level, title = len(match[1]), match[2]
path = [(n, t) for n, t in path if n < level] + [(level, title)]
headings = re.findall(r'(?m)^#{1,6} (.+)$', chunk)
opening_heading = re.match(r'^(#{1,6}) ', chunk)
if opening_heading:
path = [(level, title) for level, title in path if level < len(opening_heading[1])]
metadata = ('\nStage: full document.' if len(chunks) == 1 else
f'\nCONTEXT ONLY, NEVER INCLUDE IN OUTPUT: batch {i+1}/{len(chunks)}; '
f'ancestor path: {" > ".join(t for _, t in path)[-80:] or "document start"}; '
f'headings: {" > ".join(headings[:3])[:80] or "continued body"}; '
f'source character range {offset+1}-{offset+len(chunk)} of {len(markdown)}; '
f'previous/next batch: {i or "none"}/{i+2 if i+1 < len(chunks) else "none"}. '
'Summarize ONLY the current source, never infer other batches.'
+ (f'\nPrevious source excerpt for context only, do not summarize it again: {prefix[-80:]}'
if prefix and not opening_heading else ''))
jobs.append((chunk, f'part{i}', instruction + metadata))
offset += len(chunk)
error = None
if concurrency == 1:
for i, job in enumerate(jobs):
try:
outputs[i] = summarize(*job)
finally:
persist()
report(f'总结分块 {i+1}/{len(chunks)}', (i+1)*85//len(chunks))
else:
with ThreadPoolExecutor(max_workers=concurrency) as pool:
futures = {pool.submit(summarize, *job): i for i, job in enumerate(jobs)}
done = 0
for future in as_completed(futures):
try:
outputs[futures[future]] = future.result()
except Exception as exc:
error = error or exc
for pending in futures:
pending.cancel()
done += 1
persist()
report(f'总结分块 {done}/{len(chunks)}', done*85//len(chunks))
if error:
raise error
if merge:
for round_index in range(4):
joined = '\n\n'.join(outputs)
groups = split_markdown(joined, chunk_budget, count)
report(f'合并大纲 {round_index+1},共 {len(groups)} 批', 88 + round_index*2)
merged = []
for i, group in enumerate(groups):
try:
merged.append(summarize(group, f'merge{round_index}/{i}', instruction +
f'\n全局汇总阶段,当前第 {i+1}/{len(groups)} 组章节摘要。统一主题、去重,保持原文语言及不同条件。'))
finally:
persist()
if len(merged) == 1:
outputs = merged
break
if count('\n\n'.join(merged)) >= count(joined):
raise ValueError('合并结果未收敛,已保留完成的分块')
outputs = merged
else:
raise ValueError('合并批次超过限制,已保留完成的分块')
report('总结完成', 100)
persist()
return '\n\n'.join(outputs)

View File

@ -34,4 +34,5 @@ def task_progress(task):
if percent is not None:
percent = min(99, int((complete_steps * 100 + percent) / total_steps))
return {'stage': detail.get('stage') or LABELS.get(current.get('type'), '处理中'),
'percent': percent, 'elapsed_seconds': elapsed}
'percent': percent, 'elapsed_seconds': elapsed,
**({'plan': detail['plan']} if isinstance(detail.get('plan'), dict) else {})}

View File

@ -6,18 +6,21 @@ from system_manage import ai_hub as hub
def claim():
with transaction.atomic():
task = AITask.objects.select_for_update(skip_locked=True).filter(status='PENDING').order_by('created_at').first()
if not task:
return None
task.status = 'RUNNING'
task.started_at = timezone.now()
task.save(update_fields=['status', 'started_at', 'updated_at'])
return task
from system_manage.mcp_scheduler import admit
return admit('worker')
def execute(task):
from users.models import User
if task.snapshot.get('engine'):
try:
from system_manage.mcp_workflows import execute as execute_workflow
return execute_workflow(task)
except Exception:
from system_manage.mcp_scheduler import finish
finish(task, 'FAILED', error='流程执行失败,请检查组件、授权及流程版本')
task.refresh_from_db()
return task
steps = []
try:
user = User.objects.get(pk=task.owner_id)
@ -31,17 +34,31 @@ def execute(task):
hub.require_use(user, service)
if not service.enabled:
raise ValueError('服务已停用')
from system_manage.mcp_registry import pipeline_resources, resolve
for key in pipeline_resources(task.snapshot['nodes']): resolve(task.workspace_id, user, key, check=False)
markdown = task.input_data.get('markdown', '')
original = markdown
source = {}
original = task.input_data.get('original', markdown)
source = task.input_data.get('heading_source', {})
if task.input_data.get('heading_structure'):
source = {**source, 'heading_structure': task.input_data['heading_structure']}
tree = {}
task.result = {'original': original, 'markdown': markdown, 'source': source, 'tree': tree}
previous_state = task.result.get('summary_state')
previous_parsed = task.result.get('parse_checkpoint')
task.result = {'original': original, 'markdown': markdown, 'source': source, 'tree': tree,
**({'summary_state': previous_state} if previous_state else {}),
**({'parse_checkpoint': previous_parsed} if previous_parsed else {})}
task.save(update_fields=['result', 'updated_at'])
for node in task.snapshot['nodes']:
step = {'type': node['type'], 'status': 'RUNNING'}
steps.append(step)
task.steps = steps
task.save(update_fields=['steps', 'updated_at'])
if node['type'] == 'parse' and previous_parsed and previous_parsed.get('node') == node:
markdown, original, source = (previous_parsed[key] for key in ('markdown', 'original', 'source'))
step.update(status='SUCCESS', progress={'stage': '复用已完成解析', 'percent': 100})
task.result.update(markdown=markdown, original=original, source=source)
task.save(update_fields=['steps', 'result', 'updated_at'])
continue
if node['type'] == 'parse':
parser = None
if node.get('processor') != 'builtin':
@ -54,63 +71,65 @@ def execute(task):
AITask.objects.filter(pk=task.pk, attempt=task.attempt, status='RUNNING').update(
steps=steps, updated_at=timezone.now())
report_progress({'stage': '上传文档到解析服务', 'percent': None})
if parser is None:
from system_manage.parsers.native import parse
report_progress({'stage': '使用内置处理器解析文档', 'percent': None})
parsed = parse(task.file_name, bytes(task.file_data))
else:
parsed = hub.parse_file(parser, task.file_name, bytes(task.file_data), node['options'], progress_callback=report_progress)
from system_manage.ai_hub_assets import asset_lock, normalize
with asset_lock():
markdown = original = hub.text(normalize(parsed['markdown'], task.workspace_id, user, parsed.get('images')), '解析结果')
source = parsed['source']
# Source JSON can contain image data too; preserve structure without inline payloads.
import json
source = json.loads(normalize(json.dumps(source), task.workspace_id, user))
task.result = {'original': original, 'markdown': markdown, 'source': source}
task.save(update_fields=['result', 'updated_at'])
from system_manage import mcp_runtime
key = 'file-reader' if parser is None else 'parser:' + str(parser.id)
parsed = mcp_runtime.invoke(task.workspace_id, user, key, {'options': node.get('options', {})},
parent=task.id, progress=report_progress, scope={'source': 'mindmap'})
markdown = parsed['text']
original = parsed.get('original', markdown)
source = parsed['structured']
task.refresh_from_db(fields=['result'])
task.result.update(original=original, markdown=markdown, source=source,
parse_checkpoint={'node': node, 'markdown': markdown, 'original': original, 'source': source})
task.save(update_fields=['result', 'updated_at'])
elif node['type'] == 'organize':
from models_provider.tools import get_model, get_model_default_params
from langchain_core.messages import HumanMessage, SystemMessage
model = hub.model_for(task.workspace_id, user, node['model_id'])
if model.update_time.isoformat() != node['resource_version']:
raise ValueError('模型配置已变更,请重新提交')
# Bound input; never silently truncate the original document.
from system_manage.ai_hub_assets import protect_images, restore_images
model_markdown, image_tokens = protect_images(markdown)
try:
# Per-model settings from model management win; task nodes override only when set.
defaults = get_model_default_params(model)
temperature = node.get('temperature')
if temperature is None:
temperature = defaults.get('temperature', 0.2)
max_tokens = node.get('output_tokens')
if max_tokens is None:
max_tokens = defaults.get('max_tokens', 4096)
instance = get_model(model, **{**defaults, 'temperature': temperature,
'max_tokens': max_tokens, 'timeout': node.get('timeout', 180)})
prompt = node['prompt'] + '\n保留 AIHUBIMAGE 数字 TOKEN 图片占位符,尽量放在相关内容旁;不要编造图片或重复占位符。'
except Exception as exc:
from rest_framework.exceptions import ValidationError
raise ValidationError('模型调用失败,请检查接入状态或输入长度') from exc
from system_manage.ai_hub_organize import organize
def organize_progress(progress):
step['progress'] = progress
AITask.objects.filter(pk=task.pk, attempt=task.attempt, status='RUNNING').update(steps=steps, updated_at=timezone.now())
def checkpoint(outputs):
task.result['organize_chunks'] = outputs
task.save(update_fields=['result', 'updated_at'])
organized = organize(model_markdown, prompt,
lambda instruction, content: instance.invoke([SystemMessage(content=instruction), HumanMessage(content=content)]),
context_tokens=node.get('context_tokens', 32768), output_tokens=max_tokens,
safety_tokens=node.get('safety_tokens', 1024), progress=organize_progress, checkpoint=checkpoint)
markdown = restore_images(hub.text(organized, '整理结果'), image_tokens)
if node.get('operation') == 'heading_repair':
from system_manage.heading_service import invoke_model
def heading_progress(progress):
step['progress'] = progress
AITask.objects.filter(pk=task.pk, attempt=task.attempt, status='RUNNING').update(steps=steps, updated_at=timezone.now())
organized = invoke_model(task.workspace_id, user, node, markdown, source, task.id, heading_progress)
markdown = organized['text']
source = {**source, **organized['structured']}
else:
from system_manage import mcp_runtime
model = hub.model_for(task.workspace_id, user, node['model_id'])
if model.update_time.isoformat() != node['resource_version']:
raise ValueError('模型配置已变更,请重新提交')
def organize_progress(progress):
step['progress'] = progress
AITask.objects.filter(pk=task.pk, attempt=task.attempt, status='RUNNING').update(steps=steps, updated_at=timezone.now())
def checkpoint(outputs):
task.refresh_from_db(fields=['result'])
task.result['summary_state' if isinstance(outputs, dict) else 'organize_chunks'] = outputs
if isinstance(outputs, dict) and outputs.get('metrics'):
step.setdefault('progress', {})['plan'] = outputs['metrics']
task.save(update_fields=['result', 'steps', 'updated_at'])
params = {'text': markdown, 'instruction': node['prompt'],
'operation': node.get('operation'),
'heading_structure': source.get('heading_structure', {}), 'heading_source': source,
'resume': task.result.get('summary_state'),
**{key: node[key] for key in (
'temperature', 'top_p', 'context_tokens', 'max_input_tokens', 'max_output_tokens',
'safety_tokens', 'retries', 'detail', 'structure_mode', 'parallel', 'auto_split',
'max_depth', 'max_children', 'node_max_chars', 'concurrency', 'input_tokens') if key in node}}
if node.get('output_tokens') is not None: params['max_tokens'] = node['output_tokens']
organized = mcp_runtime.invoke(task.workspace_id, user, 'model:' + str(model.id), params,
parent=task.id, timeout=node.get('timeout', 180), progress=organize_progress,
checkpoint=checkpoint, scope={'source': 'mindmap'})
markdown = organized['text']
if isinstance(organized.get('structured'), dict):
source = {**source, **organized['structured']}
else:
tree = hub.markdown_tree(markdown)
from system_manage import mcp_runtime
tree = mcp_runtime.invoke(task.workspace_id, user, 'mindmap-engine', {'text': markdown,
'heading_structure': source.get('heading_structure')},
parent=task.id, scope={'source': 'mindmap'})['structured']
step['status'] = 'SUCCESS'
task.steps = steps
# Persist intermediate original output so a later model failure does not lose OCR work.
task.result = {'original': original, 'markdown': markdown, 'source': source, 'tree': tree}
task.refresh_from_db(fields=['result'])
task.result.update(original=original, markdown=markdown, source=source, tree=tree)
task.save(update_fields=['steps', 'result', 'updated_at'])
task.status = 'SUCCESS'
except Exception as exc:

View File

@ -1,37 +1,65 @@
import signal
import threading
import time
from datetime import timedelta
from concurrent.futures import ThreadPoolExecutor
from django.core.management.base import BaseCommand
from django.db import close_old_connections
from django.utils import timezone
from system_manage.models.ai_hub import AITask
from system_manage.ai_hub_worker import claim, execute
from system_manage import mcp_scheduler
class Command(BaseCommand):
help = 'Run the isolated AI hub task consumer (does not run migrations).'
help = 'Run the isolated component/hub consumer; never consumes Celery queues or migrates.'
def add_arguments(self, parser):
parser.add_argument('--once', action='store_true')
parser.add_argument('--recover', action='store_true', help='Mark running tasks older than 30 minutes failed')
parser.add_argument('--recover', action='store_true', help='Mark tasks without heartbeat for 30 minutes failed')
parser.add_argument('--concurrency', type=int, default=1)
def handle(self, *args, **options):
if options['recover']:
AITask.objects.filter(status='RUNNING', updated_at__lt=timezone.now() - timedelta(minutes=30)).update(
status='FAILED', error='执行中断,请重试', finished_at=timezone.now())
concurrency = options.get('concurrency', 1)
if not 1 <= concurrency <= 8:
raise ValueError('concurrency must be between 1 and 8')
if options.get('recover'):
mcp_scheduler.recover()
stop = threading.Event()
if threading.current_thread() is threading.main_thread():
signal.signal(signal.SIGTERM, lambda *_: stop.set())
next_cleanup = time.monotonic() + 86400
while True:
def run(task):
close_old_connections()
if time.monotonic() >= next_cleanup:
from system_manage.ai_hub_assets import cleanup
try:
cleanup(apply=True)
except Exception:
self.stderr.write('图片缓存清理失败,将于下次重试')
next_cleanup = time.monotonic() + 86400
task = claim()
if task:
execute(task)
if options['once']:
return
if not task:
time.sleep(2)
try:
return execute(task)
except Exception:
mcp_scheduler.finish(task, 'FAILED', error='执行器异常,请检查后重试')
finally:
close_old_connections()
with ThreadPoolExecutor(max_workers=concurrency) as pool:
running = {}
while not stop.is_set() or running:
close_old_connections()
mcp_scheduler.heartbeat()
for future, task in list(running.items()):
if future.done():
future.result()
del running[future]
else:
AITask.objects.filter(pk=task.id, attempt=task.attempt, status='RUNNING').update(updated_at=timezone.now())
if not stop.is_set():
while len(running) < concurrency:
task = claim()
if not task: break
if stop.is_set():
AITask.objects.filter(pk=task.id, attempt=task.attempt, status='RUNNING').update(
status='PENDING', started_at=None, updated_at=timezone.now())
break
running[pool.submit(run, task)] = task
if options.get('once'): stop.set()
if time.monotonic() >= next_cleanup and not running:
from system_manage.ai_hub_assets import cleanup
try: cleanup(apply=True)
except Exception: self.stderr.write('图片缓存清理失败,将于下次重试')
next_cleanup = time.monotonic() + 86400
if running or not stop.is_set(): time.sleep(1)

View File

@ -15,6 +15,18 @@ def parse(name, content):
from system_manage.ai_hub import text
from . import extension
suffix = extension(name)
if suffix in ('md', 'txt'):
# Preserve headings and whitespace even when a document has no body paragraph.
try:
markdown = content.decode('utf-8-sig')
except UnicodeDecodeError:
from charset_normalizer import from_bytes
decoded = from_bytes(content).best()
if decoded is None:
raise ValidationError('无法识别文本编码,请使用 UTF-8')
markdown = str(decoded)
return {'markdown': text(markdown, '解析结果'), 'images': {},
'source': {'engine': 'maxkb-native', 'format': suffix}}
if suffix == 'doc':
command = shutil.which('libreoffice') or shutil.which('soffice')
if command is None:
@ -60,12 +72,31 @@ def parse(name, content):
mime = mimetypes.guess_type(file.file_name)[0] or 'image/jpeg'
images['./oss/file/' + str(file.id)] = f'data:{mime};base64,' + base64.b64encode(raw).decode()
upload = SimpleUploadedFile(Path(name.replace('\\', '/')).name, content)
result = handler.handle(upload, [], False, 2_000_000, lambda _: content, capture, 0)
sections = result if isinstance(result, list) else [result]
markdown = '\n\n'.join('\n\n'.join(
((('# ' + item.get('title', '') + '\n') if item.get('title') else '') + item.get('content', ''))
for item in section.get('content', [])) for section in sections if isinstance(section, dict))
evidence = {}
if suffix == 'docx':
from docx import Document
from .native_outline import word_markdown
# Use the converter's Markdown before knowledge-base chunking flattens headings.
markdown, evidence = word_markdown(Document(io.BytesIO(content)), handler, capture)
elif suffix in ('html', 'htm'):
from html2text import html2text
from common.handle.impl.text.html_split_handle import get_encoding
markdown = html2text(content.decode(get_encoding(content)))
from .native_outline import html_evidence
evidence = html_evidence(content)
elif suffix == 'pdf':
from .native_outline import pdf_markdown
markdown, evidence = pdf_markdown(content)
if markdown is None:
result = handler.handle(upload, [], False, 2_000_000, lambda _: content, capture, 0)
markdown = '\n\n'.join(item.get('content', '') for item in result.get('content', []))
else:
result = handler.handle(upload, [], False, 2_000_000, lambda _: content, capture, 0)
sections = result if isinstance(result, list) else [result]
markdown = '\n\n'.join('\n\n'.join(
((('# ' + item.get('title', '') + '\n') if item.get('title') else '') + item.get('content', ''))
for item in section.get('content', [])) for section in sections if isinstance(section, dict))
for path, uri in images.items():
markdown = markdown.replace(path, uri)
return {'markdown': text(markdown, '解析结果'), 'images': {},
'source': {'engine': 'maxkb-native', 'format': suffix}}
'source': {'engine': 'maxkb-native', 'format': suffix, **evidence}}

View File

@ -11,6 +11,83 @@ from system_manage.ai_hub_worker import claim, execute
from system_manage.models.ai_hub import AIFlow, AIParserResource, AITask, AIArtifact, AIService
class NativeParserTests(TestCase):
def test_word_preserves_heading_only_sections_tables_and_images(self):
import io
from docx import Document
from PIL import Image
from system_manage.parsers.native import parse
document = Document()
document.add_heading('Chapter42', 1)
document.add_heading('Section42', 2)
document.add_heading('Detail42', 3)
document.add_paragraph('Body42')
table = document.add_table(rows=2, cols=2)
table.cell(0, 0).text = 'Name'
table.cell(0, 1).text = 'Value'
table.cell(1, 0).text = 'Count'
table.cell(1, 1).text = '42'
image = io.BytesIO()
Image.new('RGB', (2, 2), 'white').save(image, format='PNG')
image.seek(0)
document.add_picture(image)
buffer = io.BytesIO()
document.save(buffer)
with patch('common.handle.impl.text.doc_split_handle.DocSplitHandle.handle',
side_effect=AssertionError('Knowledge chunking must not run')):
markdown = parse('headings.docx', buffer.getvalue())['markdown']
self.assertEqual([line for line in markdown.splitlines() if line.startswith('#')],
['# Chapter42', '## Section42', '### Detail42'])
self.assertIn('Body42', markdown)
self.assertIn('| Count | 42 |', markdown)
self.assertIn('data:image/png;base64,', markdown)
self.assertNotIn('./oss/file/', markdown)
tree = hub.markdown_tree(markdown)
self.assertEqual(tree['children'][0]['children'][0]['children'][0]['text'], 'Detail42')
def test_html_and_htm_preserve_headings_lists_and_links_without_chunking(self):
from system_manage.parsers.native import parse
content = (b'<meta charset="utf-8"><h1>Chapter42</h1><h2>Section42</h2>'
b'<h3>Detail42</h3><p>Body42</p><ul><li>Parent42'
b'<ul><li>Child42</li></ul></li></ul>'
b'<p><a href="https://example.org">Link42</a></p>')
for suffix in ('html', 'htm'):
with self.subTest(suffix=suffix), patch(
'common.handle.impl.text.html_split_handle.HTMLSplitHandle.handle',
side_effect=AssertionError('Knowledge chunking must not run')):
markdown = parse('headings.' + suffix, content)['markdown']
self.assertEqual([line for line in markdown.splitlines() if line.startswith('#')],
['# Chapter42', '## Section42', '### Detail42'])
self.assertIn('Child42', markdown)
self.assertIn('[Link42](https://example.org)', markdown)
def test_xlsx_uses_cached_formula_values_and_does_not_evaluate_formulas(self):
import io
import zipfile
from openpyxl import Workbook
from system_manage.parsers.native import parse
workbook = Workbook()
workbook.active.append(['Cached', 'Uncached', 'Literal'])
workbook.active.append(['=6*7', '=2+3', 17])
buffer = io.BytesIO()
workbook.save(buffer)
cached = io.BytesIO()
with zipfile.ZipFile(io.BytesIO(buffer.getvalue())) as source, zipfile.ZipFile(cached, 'w') as target:
for item in source.infolist():
data = source.read(item.filename)
if item.filename == 'xl/worksheets/sheet1.xml':
self.assertIn(b'<f>6*7</f><v></v>', data)
data = data.replace(b'<f>6*7</f><v></v>', b'<f>6*7</f><v>42</v>')
target.writestr(item, data)
markdown = parse('cached.xlsx', cached.getvalue())['markdown']
self.assertIn('| 42 | | 17 |', markdown)
self.assertNotIn('=6*7', markdown)
self.assertNotIn('=2+3', markdown)
class HubTests(TestCase):
def setUp(self):
import tempfile

View File

@ -1,10 +1,59 @@
import unittest
from types import SimpleNamespace
from unittest.mock import patch
from system_manage.ai_hub_organize import organize, split_markdown, token_bound
from system_manage.ai_hub_organize import organize, split_markdown, token_bound, resolve_settings
class OrganizeTests(unittest.TestCase):
def test_model_management_budgets_and_workflow_precedence(self):
defaults = {'context_tokens': '131072', 'output_tokens': '8192', 'safety_tokens': '2048', 'max_tokens': 16384}
self.assertEqual(resolve_settings(defaults, {}), {'context_tokens': 131072, 'output_tokens': 8192, 'safety_tokens': 2048})
self.assertEqual(resolve_settings(defaults, {'max_tokens': 4096, 'context_tokens': 65536})['output_tokens'], 4096)
self.assertEqual(resolve_settings(defaults, {'output_tokens': 1024})['output_tokens'], 1024)
for params in ({'context_tokens': True}, {'safety_tokens': -1}, {'context_tokens': 4096, 'max_tokens': 8192}):
with self.assertRaises(ValueError): resolve_settings(defaults, params)
def test_truncated_response_retries_with_original_not_partial_draft(self):
calls, saved, metrics = [], [], {}
def invoke(instruction, content):
calls.append((instruction, content))
return SimpleNamespace(content='incomplete' if len(calls) == 1 else '# Result\n- fact42',
response_metadata={'finish_reason': 'length' if len(calls) == 1 else 'stop'})
self.assertIn('fact42', organize('Original fact42', '总结', invoke, checkpoint=saved.append, metrics=metrics))
self.assertEqual(calls[0][1], calls[1][1])
self.assertNotIn('incomplete', str(saved))
self.assertEqual(metrics['compression_retries'], 1)
def test_complete_overlong_summary_is_semantically_compressed(self):
contents = []
def invoke(prompt, content):
contents.append(content)
return 'complete verbose result ' * 100 if len(contents) == 1 else '# Summary\n- fact42'
result = organize('Original fact42', '总结', invoke, context_tokens=8192, output_tokens=256)
self.assertIn('fact42', result)
self.assertIn('complete verbose result', contents[1])
def test_repeated_truncation_splits_source_and_preserves_completed_sibling(self):
saved, metrics = [], {}
def invoke(prompt, content):
if len(content) > 700:
return SimpleNamespace(content='partial', response_metadata={'finish_reason': 'length'})
if 'FAIL' in content:
return SimpleNamespace(content='partial', response_metadata={'finish_reason': 'length'})
return '# Completed42'
with self.assertRaisesRegex(ValueError, '截断'):
organize('A' * 700 + '\nFAIL' + 'B' * 700, '总结', invoke, checkpoint=saved.append, metrics=metrics)
self.assertTrue(any('Completed42' in value for snapshot in saved for value in snapshot))
self.assertNotIn('partial', str(saved))
self.assertGreater(metrics['split_retries'], 0)
def test_echo_model_cannot_grow_merge_forever_or_be_accepted(self):
calls = []
def invoke(prompt, content): calls.append(1); return content + '\nextra'
with self.assertRaisesRegex(ValueError, '长度'):
organize('facts42 ' * 500, '总结', invoke, context_tokens=8192, output_tokens=256)
self.assertLessEqual(len(calls), 160)
def test_split_preserves_long_unbroken_unicode_and_image_tokens(self):
text = '# 标题\n' + '超长段落42' * 8000 + '\nAIHUBIMAGE0001TOKEN\n结束'
chunks = split_markdown(text, 2048)
@ -19,7 +68,7 @@ class OrganizeTests(unittest.TestCase):
calls.append(content)
return SimpleNamespace(content='# Result\n- fact42', response_metadata={})
result = organize(source, '整理事实', invoke, context_tokens=8192, output_tokens=1024,
progress=stages.append, checkpoint=saved.append)
progress=stages.append, checkpoint=saved.append, options={'structure_mode': 'reorganize'})
self.assertIn('fact42', result)
self.assertGreater(len(calls), 2)
self.assertGreater(len(saved), 2)
@ -39,6 +88,261 @@ class OrganizeTests(unittest.TestCase):
return '# saved42'
with patch('system_manage.ai_hub_organize.time.sleep'):
with self.assertRaisesRegex(ValueError, '已保留') as failure:
organize('word42 ' * 5000, '整理', invoke, context_tokens=4096, output_tokens=512, checkpoint=saved.append)
organize('word42 ' * 5000, '整理', invoke, context_tokens=8192, output_tokens=512, checkpoint=saved.append)
self.assertNotIn('private', str(failure.exception))
self.assertEqual(saved, [['# saved42']])
class SummaryPlanningTests(unittest.TestCase):
def test_short_document_one_call_and_language_rules(self):
calls, metrics = [], {}
def invoke(prompt, content):
calls.append((prompt, content))
return '# Operations\n- Inspect every 42 days.'
result = organize('# Operations\nInspect every 42 days.', '保留事实', invoke, metrics=metrics)
self.assertEqual(len(calls), 1)
self.assertIn('42 days', result)
self.assertIn('ENGLISH ONLY', calls[0][0])
self.assertIn('Stage: full document', calls[0][0])
self.assertEqual(metrics['expected_calls'], 1)
self.assertFalse(metrics['needs_merge'])
def test_budget_counts_prompt_metadata_and_framing(self):
calls, metrics = [], {}
def invoke(prompt, content):
calls.append(1)
self.assertLessEqual(token_bound(prompt) + token_bound(content) + 256, 8192-1024-1024)
return '- fact42'
organize('English data 42. ' * 1500, 'Only facts', invoke, context_tokens=8192,
output_tokens=1024, metrics=metrics)
self.assertEqual(len(calls), metrics['map_chunks'])
self.assertGreater(len(calls), 1)
def test_output_budget_and_detail_can_require_split(self):
source = '# Operations\n' + 'Inspect equipment 42 days. ' * 180
brief, detailed = {}, {}
for detail, metric in [('brief', brief), ('detailed', detailed)]:
organize(source, 'Summarize', lambda *a: '- 42 days', output_tokens=1024,
options={'detail': detail}, metrics=metric)
self.assertEqual(brief['map_chunks'], 1)
self.assertGreater(detailed['map_chunks'], 1)
self.assertGreater(detailed['estimated_output_tokens'], brief['estimated_output_tokens'])
def test_parallel_preserves_order_and_bounds_concurrency(self):
import threading, time
active, peak = 0, 0
lock = threading.Lock()
source = ''.join(f'## Chapter {i}\n' + 'facts ' * 700 + '\n' for i in range(4))
def invoke(prompt, content):
nonlocal active, peak
with lock:
active += 1; peak = max(peak, active)
time.sleep(.02 if 'Chapter 0' in content else .005)
with lock: active -= 1
return '- ' + content[:35].replace('\n', ' ')
metrics = {}
result = organize(source, 'Summarize', invoke, context_tokens=8192, output_tokens=2048,
options={'concurrency': 2}, metrics=metrics)
self.assertEqual(peak, 2)
self.assertEqual(metrics['model_calls'], metrics['map_chunks'])
self.assertLess(result.index('Chapter 0'), result.index('Chapter 1'))
def test_failure_resume_reuses_only_matching_configuration(self):
source = 'facts42 ' * 1200
states, calls = [], []
def fail(prompt, content):
calls.append(1)
if len(calls) > 1: raise RuntimeError('secret')
return '- retained42'
with self.assertRaises(ValueError):
organize(source, 'Summarize', fail, context_tokens=8192, output_tokens=1024,
options={'retries': 0}, state_checkpoint=states.append)
metrics = {}
organize(source, 'Summarize', lambda *a: '- resumed42', context_tokens=8192, output_tokens=1024,
options={'retries': 0}, resume=states[-1], metrics=metrics)
self.assertEqual(metrics['reused_chunks'], 1)
self.assertEqual(metrics['model_calls'], metrics['map_chunks']-1)
other = {}
organize(source, 'New prompt', lambda *a: '- new42', context_tokens=8192, output_tokens=1024,
options={'retries': 0}, resume=states[-1], metrics=other)
self.assertEqual(other['reused_chunks'], 0)
def test_chunk_metadata_and_mixed_language_survive_merge(self):
seen = []
def invoke(prompt, content):
seen.append(prompt)
return '# Maintenance\n- 42 days\n## 检修\n- 每月一次' if '检修' in content else '# Maintenance\n- 42 days'
organize('## Maintenance 检修\n' + '42 days ' * 1400, 'Summarize', invoke,
context_tokens=8192, output_tokens=2048, options={'structure_mode': 'reorganize'})
self.assertTrue(any('batch' in p and 'character range' in p and 'ancestor path' in p for p in seen))
self.assertTrue(any('全局汇总阶段' in p for p in seen))
self.assertTrue(all('original' in p.lower() for p in seen))
def test_no_split_option_rejects_before_call(self):
invoke = unittest.mock.Mock()
with self.assertRaisesRegex(ValueError, '开启自动拆分'):
organize('facts ' * 5000, 'Summarize', invoke, options={'auto_split': False})
invoke.assert_not_called()
def test_invalid_settings_and_declared_model_limits(self):
from system_manage.summary_plan import validate_options
for options in ({'concurrency': True}, {'top_p': float('nan')}, {'detail': 'unknown'},
{'parallel': 'true'}, {'context_tokens': 1024}, {'max_depth': 7}):
with self.assertRaises(ValueError): validate_options(options)
with self.assertRaisesRegex(ValueError, '上限'):
resolve_settings({'context_tokens': 8192}, {'context_tokens': 16384})
with self.assertRaises(ValueError):
resolve_settings({'max_output_tokens': 1024}, {'max_tokens': 2048})
def test_max_input_independent_limit(self):
stats = {}
organize('data42 ' * 2000, 'Summarize', lambda *a: '- data42', output_tokens=512,
options={'max_input_tokens': 8192}, metrics=stats)
self.assertLess(stats['input_budget'], 8192)
self.assertGreater(stats['map_chunks'], 1)
def test_table_list_and_fenced_example_stay_whole_when_they_fit(self):
table = '| Key | Value |\n| --- | --- |\n| Count | 42 |\n'
code = '```text\n# not a heading\nvalue42\n```\n'
items = '- step 1\n- step 2\n'
for block in (table, code, items):
text = 'Intro ' * 45 + '\n' + block + 'After ' * 60
parts = split_markdown(text, 300)
self.assertEqual(''.join(parts), text)
self.assertTrue(any(block in part for part in parts))
self.assertTrue(all(token_bound(part) <= 300 for part in parts))
def test_custom_node_length_changes_output_planning(self):
metrics = []
for length in (20, 120):
stats = {}
organize('Equipment 42 days. ' * 240, 'Summarize', lambda *a: '- 42 days', output_tokens=1024,
options={'node_max_chars': length}, metrics=stats)
metrics.append(stats)
self.assertGreater(metrics[1]['map_chunks'], metrics[0]['map_chunks'])
def test_provider_usage_prevents_unnecessary_compression_from_byte_estimate(self):
calls = []
def invoke(*a):
calls.append(1)
return SimpleNamespace(content='中文结论42。' * 30, response_metadata={'finish_reason': 'stop'},
usage_metadata={'output_tokens': 200})
organize('原文42', 'Summarize', invoke, output_tokens=256)
self.assertEqual(len(calls), 1)
def test_wrong_language_retries_only_current_source_and_unwraps_any_fence(self):
contents, metrics = [], {}
def invoke(prompt, content):
contents.append(content)
return '中文错误结果' if len(contents) == 1 else '```ruby\n# Maintenance\n- Inspect every 42 days.'
output = organize('Inspect every 42 days.', 'Summarize', invoke, metrics=metrics)
self.assertEqual(len(contents), 2)
self.assertEqual(contents[0], contents[1])
self.assertEqual(output, '# Maintenance\n- Inspect every 42 days.')
self.assertEqual(metrics['language_retries'], 1)
def test_repeated_wrong_language_is_not_success(self):
with self.assertRaisesRegex(ValueError, '未保持原文语言'):
organize('Inspect every 42 days.', 'Summarize', lambda *a: '中文错误结果')
def test_language_detection_preserves_source_despite_translation_prompt(self):
from system_manage.summary_plan import source_language
self.assertEqual(source_language('English text', '请翻译为中文'), 'english')
self.assertEqual(source_language('これは日本語の説明です'), 'japanese')
self.assertEqual(source_language('한국어 설명입니다'), 'korean')
def test_original_headings_are_validated_and_levels_restored(self):
calls=[]
def invoke(*a):
calls.append(1)
return '- fact42' if len(calls)==1 else '### Maintenance\n#### Inspection\n- fact42'
result=organize('# Maintenance\n## Inspection\nInspect every 42 days.', 'Summarize', invoke,
options={'structure_mode':'preserve'})
self.assertTrue(result.startswith('# Maintenance\n## Inspection'))
self.assertEqual(len(calls),2)
def test_ollama_length_stop_is_never_accepted(self):
with self.assertRaisesRegex(ValueError,'截断'):
organize('English source', 'Summarize', lambda *a: SimpleNamespace(content='Partial',
response_metadata={'done_reason':'length','eval_count':256}),output_tokens=256)
def test_mixed_language_failure_recovers_only_language_local_sections(self):
calls, stats = [], {}
source='# Maintenance / 维护\n## English operations\nInspect every 42 days.\n## 中文操作\n每31天记录一次运行情况。'
def invoke(prompt, content):
calls.append(content)
if len(calls)==1:return '## 英文操作\n每42天检查一次。\n## 中文操作\n每31天记录一次运行情况。'
if 'English operations' in content:return '## English operations\n- Inspect every 42 days.'
return '## 中文操作\n- 每31天记录一次运行情况。'
output=organize(source,'Summarize',invoke,options={'structure_mode':'preserve'},metrics=stats)
self.assertEqual(len(calls),3)
self.assertNotIn('中文操作',calls[1])
self.assertNotIn('English operations',calls[2])
self.assertIn('42 days',output)
self.assertIn('每31天',output)
self.assertEqual(stats['language_splits'],1)
self.assertEqual(stats['expected_calls'],3)
def test_single_body_heading_chain_is_restored_without_another_call(self):
calls=[]
def invoke(*a):calls.append(1);return '### Maintenance\n- Inspect every 42 days.'
output=organize('# Maintenance\n## Inspection\nInspect every 42 days.','Summarize',invoke,
options={'structure_mode':'preserve'})
self.assertEqual(len(calls),1)
self.assertIn('# Maintenance\n## Inspection',output)
def test_customized_default_prompt_is_not_discarded(self):
from system_manage.summary_plan import DEFAULT_SUMMARY_PROMPT,prompt_rules
custom=DEFAULT_SUMMARY_PROMPT+'\nKeep the STOP condition explicit.'
self.assertIn(custom,prompt_rules(custom,{},'english'))
def test_independent_chapter_does_not_inherit_previous_sibling_context(self):
seen=[]
source='## Alpha\n'+'Alpha fact42. '*180+'\n## Beta\n'+'Beta fact43. '*180
def invoke(prompt,content):
seen.append((prompt,content));return '- fact42 and fact43'
organize(source,'Summarize',invoke,context_tokens=8192,output_tokens=1024)
beta=[prompt for prompt,content in seen if content.startswith('## Beta')]
self.assertTrue(beta)
self.assertTrue(all('ancestor path: document start' in prompt for prompt in beta))
self.assertTrue(all('Previous source excerpt' not in prompt for prompt in beta))
def test_model_wrapper_and_trailing_self_comment_are_removed_without_call(self):
for wrapped in ('```markdown\n# Result\n- fact42\n```\nThis summary is complete.',
'# Result\n- fact42\n```\nThis summary is complete.'):
calls=[]
def invoke(*a):calls.append(1);return wrapped
result=organize('Original fact42','Summarize',invoke)
self.assertEqual(result,'# Result\n- fact42')
self.assertEqual(len(calls),1)
def test_explicit_numeric_conditions_keep_strictness_negation_and_sign(self):
from system_manage.summary_plan import boundaries_match
self.assertTrue(boundaries_match('Stop above 91 degrees.', 'Stop when temperature exceeds 91.0 degrees.'))
self.assertFalse(boundaries_match('Stop above 91 degrees.', 'Stop when it reaches 91 degrees.'))
self.assertFalse(boundaries_match('温度超过90℃停止。', '温度达到90℃停止。'))
self.assertTrue(boundaries_match('Do not exceed 90.', 'At most 90.0.'))
self.assertFalse(boundaries_match('At least -5.', 'Below -5.'))
self.assertTrue(boundaries_match('Thunder 42.', 'Thunder 42.'))
def test_changed_numeric_boundary_repairs_from_original_only(self):
contents,stats=[],{}
def invoke(prompt,content):
contents.append(content)
return '- Stop when it reaches 91 degrees.' if len(contents)==1 else '- Stop above 91 degrees.'
output=organize('Stop above 91 degrees.','Summarize',invoke,metrics=stats)
self.assertEqual(len(contents),2)
self.assertEqual(contents[0],contents[1])
self.assertIn('above 91',output)
self.assertEqual(stats['fact_retries'],1)
def test_repeated_numeric_boundary_change_cannot_be_success(self):
with self.assertRaisesRegex(ValueError,'数值边界条件'):
organize('Stop above 91 degrees.','Summarize',lambda *a:'- Stop when it reaches 91 degrees.')

View File

@ -1,13 +1,15 @@
from system_manage.views.mcp_engine import EngineView
from django.urls import path
from . import views
from .views.ai_hub import HubView, PublishedRun, ParserTest, AssetContent, MarkdownExport, DocumentExport
from .views.ai_hub import HubView, PublishedRun, ParserTest, AssetContent, MarkdownExport, DocumentExport, ServiceCalls
from .views.platform_access import AccessView, RoleView, GrantsView, UserAccessView
from .views.access_control_policy import AccessControlRuntimeView, AccessControlDecisionView
app_name = "system_manage"
# @formatter:off
urlpatterns = [
path('workspace/<str:workspace_id>/mcp_engine/<str:kind>', EngineView.as_view()),
path('platform_access', AccessView.as_view()),
path('platform_access/roles', RoleView.as_view()),
path('platform_access/roles/<uuid:pk>', RoleView.as_view()),
@ -18,6 +20,7 @@ urlpatterns = [
path('workspace/<str:workspace_id>/ai_hub/exports', MarkdownExport.as_view()),
path('workspace/<str:workspace_id>/ai_hub/assets/<uuid:pk>/content', AssetContent.as_view()),
path('workspace/<str:workspace_id>/ai_hub/parsers/<uuid:pk>/test', ParserTest.as_view()),
path('workspace/<str:workspace_id>/ai_hub/services/<uuid:pk>/calls', ServiceCalls.as_view()),
path('workspace/<str:workspace_id>/ai_hub/published/<uuid:pk>/run', PublishedRun.as_view()),
path('workspace/<str:workspace_id>/ai_hub/<str:kind>', HubView.as_view()),
path('workspace/<str:workspace_id>/ai_hub/<str:kind>/<uuid:pk>', HubView.as_view()),

View File

@ -35,10 +35,18 @@ class HubView(APIView):
raise PermissionDenied('未授权使用思维导图')
def get(self, request, workspace_id, kind, pk=None):
if kind == 'artifact_limit' and not pk:
from system_manage import artifact_settings
hub.require_admin(request.user)
return result.success({'limit': artifact_settings.limit()})
if kind == 'workbench_defaults' and not pk:
from system_manage import workbench_defaults
return result.success(workbench_defaults.read(workspace_id, request.user))
if kind == 'parser_catalog':
from system_manage.parsers.catalog import catalogue
return result.success(catalogue())
if kind == 'resources':
from system_manage.mcp_business import summary_model_metadata
parsers = hub.visible(hub.AIParserResource.objects.filter(workspace_id=workspace_id, enabled=True), request.user)
from users.models import User
from users.models.user_manage_ext import UserActivationApplication
@ -48,7 +56,9 @@ class HubView(APIView):
resource_type__in=['PARSER', 'SERVICE'], actions__contains=['AUTHORIZE']).exists()
return result.success({'parsers': [hub.serialize(p, user=request.user) for p in parsers
if 'USE' in hub.serialize(p, user=request.user)['actions']],
'models': list(hub.authorized_models(workspace_id, request.user).values('id', 'name', 'model_type')),
'models': [summary_model_metadata(model)
if model.model_type == 'LLM' else {'id': str(model.id), 'name': model.name, 'model_type': model.model_type}
for model in hub.authorized_models(workspace_id, request.user)],
'users': [{'id': str(u.id), 'name': u.nick_name or u.username} for u in User.objects.filter(is_active=True).exclude(id__in=pending)] if may_authorize else [],
'is_admin': hub.admin(request.user)})
if pk:
@ -61,9 +71,33 @@ class HubView(APIView):
raise ValidationError('资源类型无效')
query = hub.visible(hub.KINDS[kind].objects.filter(workspace_id=workspace_id), request.user)
if kind == 'tasks':
from django.db.models import Q
query = query.filter(Q(snapshot__engine__isnull=True) | Q(snapshot__engine=False))
query = query.defer('file_data', 'input_data', 'snapshot', 'result')
elif kind == 'artifacts':
query = query.defer('original', 'markdown', 'history')
if any(key in request.query_params for key in ('mine', 'current_page', 'page_size', 'name')):
mine = request.query_params.get('mine', 'false')
if mine not in ('true', 'false'):
raise ValidationError('查询参数无效')
try:
page = int(request.query_params.get('current_page', '1'))
size = int(request.query_params.get('page_size', '20'))
if page < 1 or size < 1 or size > 100:
raise ValueError()
except (ValueError, TypeError):
raise ValidationError('分页参数无效')
name = request.query_params.get('name', '')
if len(name) > 128:
raise ValidationError('名称无效')
if mine == 'true':
query = query.filter(owner_id=request.user.id)
if name:
query = query.filter(name__icontains=name)
total = query.count()
records = query.order_by('-updated_at', '-id')[(page - 1) * size:page * size]
return result.success({'records': [hub.serialize(obj, user=request.user) for obj in records],
'total': total, 'current': page})
return result.success([hub.serialize(obj, user=request.user) for obj in query.order_by('-created_at')[:200]])
def post(self, request, workspace_id, kind, pk=None):
@ -76,6 +110,20 @@ class HubView(APIView):
raise ValidationError('任务参数无效')
if not isinstance(data, dict):
raise ValidationError('请求参数无效')
if kind == 'artifact_limit' and not pk:
from system_manage import artifact_settings
return result.success(artifact_settings.save(request.user, data))
if kind == 'workbench_defaults' and not pk:
from system_manage import workbench_defaults
return result.success(workbench_defaults.save(workspace_id, request.user, data))
if kind == 'summary_preview':
from system_manage.mcp_business import summary_preview
from system_manage.access_control import enforce_task
enforce_task(request, nodes=data.get('nodes'))
return result.success(summary_preview(workspace_id, request.user, data))
if kind == 'heading_structure':
from system_manage.heading_service import process
return result.success(process(workspace_id, request.user, data, request))
if kind == 'tasks':
from system_manage.access_control import enforce_task
if pk:
@ -97,7 +145,12 @@ class HubView(APIView):
if kind == 'parsers':
obj = hub.parser_save(workspace_id, request.user, data, obj)
elif kind == 'artifacts':
obj = hub.artifact_save(workspace_id, request.user, data, obj)
if 'action' in data:
if data['action'] != 'rename':
raise ValidationError('操作无效')
obj = hub.artifact_rename(obj, data)
else:
obj = hub.artifact_save(workspace_id, request.user, data, obj)
elif kind == 'services':
if set(data) - {'allowed_users'}:
hub.require_resource_action(request.user, obj, 'EDIT')
@ -120,12 +173,17 @@ class HubView(APIView):
if data.get('revoke_key'):
obj.key_hash = ''
obj.save()
if 'api_config' in data:
from system_manage.ai_hub_service_api import save_config
save_config(obj.id, data['api_config'])
output = hub.serialize(obj)
if key:
output['api_key'] = key
return result.success(output)
elif kind == 'tasks':
obj = AITask.objects.select_for_update().get(pk=obj.pk)
if obj.snapshot.get('engine'):
raise ValidationError('请在 MCP 引擎任务页重试该任务')
if obj.status != 'FAILED':
raise ValidationError('仅失败任务可重试')
if hub.make_snapshot(workspace_id, request.user, obj.snapshot['nodes']) != obj.snapshot['nodes']:
@ -151,6 +209,8 @@ class HubView(APIView):
raise ValidationError('待处理任务已达上限')
obj = hub.submit(workspace_id, request.user, data, request.FILES.get('file'), request=request)
elif kind == 'artifacts':
if 'action' in data:
raise ValidationError('操作无效')
obj = hub.artifact_save(workspace_id, request.user, data)
else:
raise ValidationError('资源类型无效')
@ -225,7 +285,79 @@ class MarkdownExport(HubView):
class PublishedRun(HubView):
"""Internal service: login AND explicit resource grants; Key is an optional extra credential."""
authentication_classes = [TokenAuth]
http_method_names = ['post', 'options']
http_method_names = ['get', 'post', 'options']
def handle_exception(self, exc):
from rest_framework.exceptions import APIException
from rest_framework.response import Response
from common.exception.app_exception import AppApiException
if isinstance(exc, (APIException, AppApiException)):
status = exc.status_code
defaults = {400: '请求参数无效', 401: '请先登录', 403: '无权调用该服务',
404: '服务不存在或无权访问', 405: '不支持的请求方法', 429: '调用过于频繁'}
message = defaults.get(status, '调用失败')
detail = getattr(exc, 'detail', None)
while isinstance(detail, (list, tuple, dict)) and detail:
detail = next(iter(detail.values())) if isinstance(detail, dict) else detail[0]
safe_messages = {'服务Key无效', '调用过于频繁', '请求方法与服务配置不一致', '请求参数过长',
'任务参数无效', '包含未知参数', '参数不能重复', '文件参数无效', '参数类型无效',
'缺少必填参数', '优先级无效', '服务已停用', '请选择文件', '文档内容不能为空',
'文件任务需要解析节点', '待处理任务已达上限', '文件不能超过20MB'}
if str(detail) in safe_messages:
message = str(detail)
self._call_error = message
return Response({'code': getattr(exc, 'code', status), 'message': message}, status=status)
raise exc
def dispatch(self, request, *args, **kwargs):
import time
self._call_started = time.monotonic()
try:
with transaction.atomic():
return super().dispatch(request, *args, **kwargs)
except Exception:
import logging
from system_manage.models import Log
logging.getLogger(__name__).error('AI service call failed')
try:
Log.objects.create(menu='AI_SERVICE', operate='CALL', workspace_id=str(kwargs.get('workspace_id', '')),
operation_object={'service_id': str(kwargs.get('pk', ''))}, user={}, status=500, ip_address='',
details={'method': request.method, 'http_status': 500,
'elapsed_ms': max(0, int((time.monotonic() - self._call_started) * 1000)),
'status': 'FAILED', 'error': '服务执行失败', 'task_id': None})
except Exception:
return result.error('调用日志保存失败', response_status=503)
return result.error('服务执行失败', response_status=500)
def finalize_response(self, request, response, *args, **kwargs):
import time
from system_manage.models import Log
from rest_framework.response import Response
response = super().finalize_response(request, response, *args, **kwargs)
body = getattr(response, 'data', None)
if body is None and response.get('Content-Type', '').startswith('application/json'):
import json
try:
body = json.loads(response.content)
except (ValueError, TypeError):
body = None
code = body.get('code') if isinstance(body, dict) else None
success = response.status_code < 400 and code in (None, 0, 200)
try:
Log.objects.create(menu='AI_SERVICE', operate='CALL', workspace_id=str(kwargs.get('workspace_id', '')),
operation_object={'service_id': str(kwargs.get('pk', ''))},
user={'id': str(getattr(getattr(request, '_user', None), 'id', '') or '')}, status=response.status_code if success or response.status_code >= 400 else 400, ip_address='',
details={'method': request.method, 'http_status': response.status_code,
'elapsed_ms': max(0, int((time.monotonic() - self._call_started) * 1000)),
'status': 'SUCCESS' if success else 'FAILED', 'error': '' if success else getattr(self, '_call_error', '调用失败'),
'task_id': getattr(self, '_task_id', None) if success else None})
except Exception:
transaction.set_rollback(True)
response = super().finalize_response(request, Response({'code': 503, 'message': '调用日志保存失败'}, status=503), *args, **kwargs)
return response
def get(self, request, workspace_id, pk):
return self.post(request, workspace_id, pk)
def post(self, request, workspace_id, pk):
service = hub.one('services', pk, workspace_id, request.user)
@ -241,17 +373,46 @@ class PublishedRun(HubView):
cache.add(key, 0, 70)
if cache.incr(key) > service.requests_per_minute:
raise ValidationError('调用过于频繁')
data = request.data
if request.FILES:
import json
try:
data = json.loads(request.data.get('data', '{}'))
except (ValueError, TypeError):
raise ValidationError('任务参数无效')
task = hub.submit(workspace_id, request.user, data, request.FILES.get('file'), service=service, request=request)
from system_manage.ai_hub_service_api import get_config, parse_request
data, upload = parse_request(request, get_config(service.id))
task = hub.submit(workspace_id, request.user, data, upload, service=service, request=request)
self._task_id = str(task.id)
return result.success(hub.serialize(task))
class ServiceCalls(HubView):
http_method_names = ['get', 'options']
def get(self, request, workspace_id, pk):
from system_manage.models import Log
service = hub.one('services', pk, workspace_id, request.user)
hub.require_resource_action(request.user, service, 'EDIT')
try:
page = int(request.query_params.get('current_page', 1))
size = int(request.query_params.get('page_size', 20))
if page < 1 or size < 1 or size > 100:
raise ValueError()
except (ValueError, TypeError):
raise ValidationError('分页参数无效')
logs = Log.objects.filter(menu='AI_SERVICE', operate='CALL', workspace_id=workspace_id,
operation_object__service_id=str(pk)).order_by('-create_time')
total = logs.count()
rows = list(logs[(page - 1) * size:page * size])
ids = [row.details.get('task_id') for row in rows if row.details.get('task_id')]
tasks = {str(t.id): t for t in AITask.objects.filter(id__in=ids, workspace_id=workspace_id)}
records = []
for row in rows:
details = row.details
task = tasks.get(details.get('task_id'))
duration = int((task.finished_at - task.started_at).total_seconds() * 1000) if task and task.started_at and task.finished_at else None
records.append({'id': str(row.id), 'created_at': row.create_time.isoformat(),
'method': details.get('method'), 'status': details.get('status'),
'http_status': details.get('http_status'), 'elapsed_ms': details.get('elapsed_ms'),
'error': (task.error if task and task.status == 'FAILED' else details.get('error')), 'task_id': details.get('task_id'),
'task_status': task.status if task else None, 'execution_ms': duration})
return result.success({'records': records, 'total': total})
class DocumentExport(HubView):
http_method_names = ['post', 'options']
@ -259,6 +420,8 @@ class DocumentExport(HubView):
from django.http import HttpResponse
from system_manage.document_export import export_document, ExportError, MAX_HTML
from system_manage.ai_hub_assets import ASSET, image_bytes
import time
started = time.perf_counter()
def image(src):
match = ASSET.fullmatch(src)
if not match or match[1] != workspace_id:
@ -275,6 +438,7 @@ class DocumentExport(HubView):
except (ExportError, ValueError, TypeError) as exc:
raise ValidationError(str(exc) if isinstance(exc, ExportError) else '文档内容格式无效') from exc
response = HttpResponse(raw, content_type=mime)
response['Server-Timing'] = f'document_export;dur={(time.perf_counter() - started) * 1000:.1f}'
response['Content-Disposition'] = 'attachment; filename="document.' + format_name + '"'
response['Cache-Control'] = 'private, no-store'
response['X-Content-Type-Options'] = 'nosniff'

View File

@ -9,6 +9,7 @@ def main():
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument('action', choices=['check', 'plan', 'migrate', 'worker', 'cleanup-assets'], nargs='?', default='check')
parser.add_argument('--apply', action='store_true', help='Apply expired image cleanup (default: report only)')
parser.add_argument('--concurrency', type=int, default=1)
parser.add_argument('--once', action='store_true')
parser.add_argument('--recover', action='store_true')
args = parser.parse_args()
@ -35,7 +36,7 @@ def main():
elif args.action == 'cleanup-assets':
call_command('ai_hub_cleanup_assets', apply=args.apply)
elif args.action == 'worker':
call_command('ai_hub_worker', once=args.once, recover=args.recover)
call_command('ai_hub_worker', once=args.once, recover=args.recover, concurrency=args.concurrency)
else:
from system_manage.models.ai_hub import AIParserResource, AIFlow, AIService, AITask, AIArtifact
tables = set(connection.introspection.table_names())

View File

@ -29,5 +29,6 @@ export async function hubBinary(kind: 'assets' | 'exports' | 'document_export',
try { const detail = JSON.parse(await response.data.text()); if (typeof detail.message === 'string') message = detail.message.slice(0, 200) } catch { /* retain a safe fallback */ }
throw new Error(message)
}
return { content: await response.data.arrayBuffer(), mime: response.data.type }
return { content: await response.data.arrayBuffer(), mime: response.data.type,
serverTimings: response.headers?.['server-timing'] || '' }
}

View File

@ -38,6 +38,7 @@
</div>
<el-collapse-transition>
<div class="mt-12" v-if="data['show']">
<EngineCallRecords :calls="data.engine_calls" />
<template v-if="data.status === 200 || data.type == WorkflowType.LoopNode">
<!-- 开始 -->
<template
@ -915,6 +916,12 @@
</div>
</div>
</template>
<template v-if="data.type === WorkflowType.ComponentNode">
<div class="card-never border-r-6 p-8-12 pre-wrap">
<p>组件:{{ data.component_id }}</p><p v-if="data.task_id">任务:{{ data.task_id }}</p>
<p>{{ data.text }}</p><pre v-if="data.structured">{{ JSON.stringify(data.structured, null, 2) }}</pre>
</div>
</template>
<!-- MCP 节点 -->
<template v-if="data.type === WorkflowType.McpNode">
<div class="card-never border-r-6">
@ -1087,6 +1094,7 @@
</template>
<script setup lang="ts">
import { ref, computed } from 'vue'
import EngineCallRecords from '@/components/engine-call-records/index.vue'
import ParagraphCard from '@/components/ai-chat/component/knowledge-source-component/ParagraphCard.vue'
import DynamicsForm from '@/components/dynamics-form/index.vue'
import { iconComponent } from '@/workflow/icons/utils'

View File

@ -25,6 +25,7 @@ export enum WorkflowType {
SpeechToTextNode = 'speech-to-text-node',
ImageGenerateNode = 'image-generate-node',
McpNode = 'mcp-node',
ComponentNode = 'component-node',
IntentNode = 'intent-node',
TextToVideoGenerateNode = 'text-to-video-node',
ImageToVideoGenerateNode = 'image-to-video-node',

View File

@ -142,25 +142,23 @@
{{ menuText.roleManagement }}
</router-link>
<router-link
v-if="user.is_admin()"
to="/system/mindmap"
style="
display: block;
padding: 10px 16px;
border-radius: 8px;
text-decoration: none;
color: #333;
font-size: 14px;
"
:style="
activeSystemMenu === '/system/mindmap'
? { backgroundColor: '#e6f0ff', color: '#1890ff' }
: {}
"
>
{{ menuText.mindmapManagement }}
</router-link>
<div v-if="user.is_admin()">
<div style="display:flex;justify-content:space-between;align-items:center;padding:10px 16px;border-radius:8px;cursor:pointer"
:style="isServiceMenu ? { backgroundColor:'#e6f0ff', color:'#1890ff' } : {}" @click="serviceExpanded = !serviceExpanded">
<span>{{ menuText.mindmapManagement }}</span><span>{{ serviceExpanded ? '▾' : '▸' }}</span>
</div>
<div v-if="serviceExpanded" style="padding-left:32px;margin-top:4px;display:flex;flex-direction:column;gap:4px">
<router-link to="/system/services" style="display:block;padding:8px 16px;border-radius:8px;text-decoration:none;color:#666;font-size:13px"
:style="activeSystemMenu === '/system/services' ? { backgroundColor:'#e6f0ff', color:'#1890ff' } : {}">{{ menuText.mindmapSettings }}</router-link>
<router-link to="/system/advanced-flows" style="display:block;padding:8px 16px;border-radius:8px;text-decoration:none;color:#666;font-size:13px"
:style="activeSystemMenu === '/system/advanced-flows' ? { backgroundColor:'#e6f0ff', color:'#1890ff' } : {}">{{ menuText.advancedFlows }}</router-link>
</div>
</div>
<router-link v-if="user.is_admin()" to="/system/mcp-engine"
style="display:block;padding:10px 16px;border-radius:8px;text-decoration:none;"
:style="activeSystemMenu === '/system/mcp-engine' ? {backgroundColor:'#e6f0ff',color:'#1890ff'} : {}">
{{ menuText.mcpEngine }}
</router-link>
<div>
<div
@ -241,7 +239,7 @@
</template>
<script setup lang="ts">
import { computed, onMounted, ref } from 'vue'
import { computed, onMounted, ref, watch } from 'vue'
import { useRoute } from 'vue-router'
import { useI18n } from 'vue-i18n'
import UserAvatar from '@/layout/layout-header/avatar/index.vue'
@ -256,7 +254,10 @@ const menuText = computed(() => {
return {
aiApplication: 'AI Applications',
mindmap: 'Mind Map',
mindmapManagement: 'Mind Map Management',
mindmapManagement: 'Service Management',
mindmapSettings: 'Mind Map Settings',
advancedFlows: 'Advanced Orchestration',
mcpEngine: 'MCP Engine',
knowledge: 'Knowledge Base',
toolManagement: 'Tool Management',
modelManagement: 'Model Management',
@ -273,7 +274,10 @@ const menuText = computed(() => {
return {
aiApplication: 'AI應用',
mindmap: '思維導圖',
mindmapManagement: '思維導圖管理',
mindmapManagement: '服務管理',
mindmapSettings: '導圖設定',
advancedFlows: '高級編排',
mcpEngine: 'MCP 引擎',
knowledge: '知識庫',
toolManagement: '工具管理',
modelManagement: '模型管理',
@ -289,7 +293,10 @@ const menuText = computed(() => {
return {
aiApplication: 'AI应用',
mindmap: '思维导图',
mindmapManagement: '思维导图管理',
mindmapManagement: '服务管理',
mindmapSettings: '导图设置',
advancedFlows: '高级编排',
mcpEngine: 'MCP 引擎',
knowledge: '知识库',
toolManagement: '工具管理',
modelManagement: '模型管理',
@ -320,8 +327,10 @@ const activeMainMenu = computed(() => {
const activeSystemMenu = computed(() => {
const path = route.path
if (path.endsWith('/system/mcp-engine')) return '/system/mcp-engine'
if (path === '/system/role') return '/system/role'
if (path === '/system/mindmap' || path === '/admin/system/mindmap') return '/system/mindmap'
if (path === '/system/services' || path === '/admin/system/services') return '/system/services'
if (path === '/system/advanced-flows' || path === '/admin/system/advanced-flows') return '/system/advanced-flows'
if (path.startsWith('/system/user') || path.startsWith('/admin/system/user'))
return '/system/user'
if (path.startsWith('/system/authorization') || path.startsWith('/admin/system/authorization')) {
@ -344,6 +353,9 @@ const activeSystemMenu = computed(() => {
const resourceAuthExpanded = ref(false)
const systemSettingsExpanded = ref(false)
const isServiceMenu = computed(() => ['/system/services', '/system/advanced-flows'].includes(activeSystemMenu.value))
const serviceExpanded = ref(isServiceMenu.value)
watch(isServiceMenu, value => { if (value) serviceExpanded.value = true })
const toggleResourceAuth = () => {
resourceAuthExpanded.value = !resourceAuthExpanded.value

View File

@ -176,7 +176,7 @@ onMounted(() => {
}
:deep(.el-dropdown-menu__item) {
padding: 12px 11px;
padding: 12px 20px;
&:hover {
background: var(--app-text-color-light-1);

View File

@ -8,13 +8,17 @@ const systemRouter = {
hidden: true,
component: () => import('@/layout/layout-template/SystemMainLayout.vue'),
children: [
{ path: '/system/mcp-engine', name: 'mcpEngine',
meta: { icon: 'Connection', iconActive: 'Connection', title: 'MCP 引擎', activeMenu: '/system',
parentPath: '/system', parentName: 'system', sameRoute: 'mcpEngine', permission: [RoleConst.ADMIN] },
component: () => import('@/views/system/mcp-engine/index.vue') },
{
path: '/system/mindmap',
path: '/system/services',
name: 'mindmapManagement',
meta: {
icon: 'app-document',
iconActive: 'app-document',
title: '思维导图管理',
title: '导图设置',
activeMenu: '/system',
parentPath: '/system',
parentName: 'system',
@ -23,6 +27,13 @@ const systemRouter = {
},
component: () => import('@/views/system/mindmap/index.vue'),
},
{
path: '/system/advanced-flows',
name: 'advancedFlows',
meta: { title: '高级编排', activeMenu: '/system', parentPath: '/system', parentName: 'system', permission: [RoleConst.ADMIN] },
component: () => import('@/views/system/mindmap/AdvancedFlowsPage.vue'),
},
{ path: '/system/mindmap', redirect: (to: any) => ({ path: '/system/services', query: { ...to.query, tab: 'mindmap' } }) },
{
path: '/system/user',
name: 'user',

View File

@ -35,6 +35,7 @@
<AppIcon iconName="app-debug-outlined" class="mr-4"></AppIcon>
{{ $t('views.applicationWorkflow.setting.debug') }}
</el-button>
<el-button @click="engineRunVisible = true" v-if="permissionPrecise.edit(id)">后台任务</el-button>
<el-button @click="saveApplication(true)" v-if="permissionPrecise.edit(id)">
<AppIcon iconName="app-save-outlined" class="mr-4"></AppIcon>
{{ $t('common.save') }}
@ -139,8 +140,19 @@
@refreshVersion="refreshVersion"
/>
</div>
<el-dialog v-model="engineRunVisible" title="提交流程任务" width="520px">
<el-form label-position="top"><el-form-item label="输入内容"><el-input v-model="engineQuestion" type="textarea" :rows="4" /></el-form-item>
<el-form-item label="输入文件"><input type="file" @change="engineFile = ($event.target as HTMLInputElement).files?.[0]" /></el-form-item>
<el-form-item label="优先级(0~9)"><el-input-number v-model="enginePriority" :min="0" :max="9" /></el-form-item>
<el-text type="info">运行已发布版本;草稿修改需先发布。</el-text>
<p v-if="engineTask">任务 {{ engineTask.id }}:{{ engineTask.status }} {{ engineTask.error }}</p>
<pre v-if="engineTask?.result?.text" style="white-space:pre-wrap">{{ engineTask.result.text }}</pre>
</el-form>
<template #footer><el-button @click="engineRunVisible = false">关闭</el-button><el-button type="primary" :loading="engineSubmitting" @click="submitEngineTask">提交任务</el-button></template>
</el-dialog>
</template>
<script setup lang="ts">
import engineApi from '@/api/system/mcp-engine'
import { ref, onMounted, onBeforeUnmount, computed, nextTick, provide } from 'vue'
import { useRouter, useRoute } from 'vue-router'
import type { Action } from 'element-plus'
@ -163,6 +175,21 @@ import { WorkflowMode } from '@/enums/application'
provide('getApplicationDetail', () => detail)
provide('workflowMode', WorkflowMode.Application)
const { theme } = useStore()
const engineRunVisible = ref(false), engineQuestion = ref(''), enginePriority = ref(5), engineSubmitting = ref(false), engineTask = ref<any>()
let engineFile: File | undefined, engineTimer: ReturnType<typeof setInterval> | undefined
async function submitEngineTask() {
engineSubmitting.value = true
try {
engineTask.value = (await engineApi.submit({application_id: route.params.id,question:engineQuestion.value,priority:enginePriority.value},engineFile)).data
if(engineTimer)clearInterval(engineTimer)
engineTimer=setInterval(async()=>{
const task=(await engineApi.list('tasks',{id:engineTask.value.id})).data
engineTask.value=task
if(!['PENDING','RUNNING'].includes(task.status)&&engineTimer)clearInterval(engineTimer)
},2000)
} finally { engineSubmitting.value=false }
}
onBeforeUnmount(()=>{if(engineTimer)clearInterval(engineTimer)})
const router = useRouter()
const route = useRoute()
const {

View File

@ -54,6 +54,15 @@ const dynamicsFormRef = ref<InstanceType<typeof DynamicsForm>>()
const form_data = ref<any>({})
const dialogVisible = ref(false)
const loading = ref(false)
// These fields configure the mind-map summarization pipeline, not chat generation.
const summaryFields = new Set([
'context_tokens', 'max_input_tokens', 'input_tokens', 'max_output_tokens',
'output_tokens', 'safety_tokens', 'retries', 'concurrency', 'max_depth',
'max_children', 'node_max_chars', 'detail', 'structure_mode', 'parallel',
'auto_split', 'tokenizer_path', 'supported_parameters', 'summary_temperature',
'summary_enable_thinking',
])
const chatFields = (fields: FormField[]) => fields.filter((field) => !summaryFields.has(field.field))
const getApi = (model_id: string, application_id?: string) => {
return loadSharedApi({ type: 'model', systemType: apiType.value }).getModelParamsForm(
model_id,
@ -72,7 +81,7 @@ const open = (model_id: string, application_id?: string, model_setting_data?: an
form_data.value = {}
const api = getApi(model_id, application_id)
api.then((ok: any) => {
model_form_field.value = ok.data
model_form_field.value = chatFields(ok.data)
// 渲染动态表单
dynamicsFormRef.value?.render(model_form_field.value, model_setting_data)
})
@ -82,8 +91,8 @@ const open = (model_id: string, application_id?: string, model_setting_data?: an
const reset_default = (model_id: string, application_id?: string) => {
const api = getApi(model_id, application_id)
api.then((ok: any) => {
model_form_field.value = ok.data
const model_setting_data = ok.data
model_form_field.value = chatFields(ok.data)
const model_setting_data = model_form_field.value
.map((item: any) => {
if (item.show_default_value === false) {
return { [item.field]: undefined }

View File

@ -175,7 +175,7 @@ const form = reactive<{
[propName: string]: any
}>({
patterns: [],
limit: 4096,
limit: 500,
overlap: 0,
with_filter: true,
})
@ -274,7 +274,7 @@ onMounted(async () => {
const res = await loadSharedApi({ type: 'knowledge', systemType: apiType.value })
.getKnowledgeDetail(id, loading)
const defaults = res.data?.meta?.segmentation || {}
form.limit = defaults.chunk_size ?? 4096
form.limit = defaults.chunk_size ?? 500
form.overlap = defaults.chunk_overlap ?? 0
rulesReady.value = true
splitDocument()

View File

@ -2,18 +2,21 @@
<div class="p-16-24">
<div class="flex-between mb-16">
<h2>{{ $t('common.setting') }}</h2>
<KnowledgeVersions
:knowledge-id="String(id)"
:can-edit="!route.path.includes('share/') && permissionPrecise.edit(id)"
/>
<el-button
v-if="!route.path.includes('share/') && permissionPrecise.edit(id)"
type="primary"
:loading="saving"
:disabled="!settingsReady"
@click="submit"
>保存</el-button
>
<div class="flex align-center" style="gap: 12px">
<KnowledgeVersions
:knowledge-id="String(id)"
:can-edit="!route.path.includes('share/') && permissionPrecise.edit(id)"
@changed="getDetail"
/>
<el-button
v-if="!route.path.includes('share/') && permissionPrecise.edit(id)"
type="primary"
:loading="saving"
:disabled="!settingsReady"
@click="submit"
>保存</el-button
>
</div>
</div>
<el-card style="--el-card-padding: 0">
<div class="knowledge-setting main-calc-height">
@ -247,7 +250,11 @@
:apiType="apiType"
section="embedding"
:form-value="baseSettings"
:embedding-disabled="Boolean(detail?.embedding_change_block_reason)"
/>
<el-text v-if="detail?.embedding_change_block_reason" type="info">
{{ detail.embedding_change_block_reason }}
</el-text>
<el-form label-position="top" class="mt-16">
<el-row :gutter="24">
<el-col :span="12">
@ -263,7 +270,7 @@
</el-col>
<el-col :span="12">
<el-form-item label="输出归一化">
<el-checkbox v-model="embeddingParams.normalize">启用</el-checkbox>
<el-checkbox v-model="embeddingParams.normalize" :disabled="Boolean(detail?.embedding_change_block_reason)">启用</el-checkbox>
</el-form-item>
</el-col>
</el-row>
@ -272,7 +279,7 @@
<el-button
link
type="primary"
:disabled="optionalEmbeddingParams.length >= 50"
:disabled="Boolean(detail?.embedding_change_block_reason) || optionalEmbeddingParams.length >= 50"
@click="addEmbeddingParam"
>添加参数</el-button
>
@ -287,10 +294,11 @@
v-model="param.key"
placeholder="参数名称"
:maxlength="64"
:disabled="Boolean(detail?.embedding_change_block_reason)"
style="width: 180px; flex-shrink: 0"
/>
<el-input v-model="param.value" placeholder="参数值" :maxlength="2000" />
<el-button link type="danger" @click="optionalEmbeddingParams.splice(index, 1)"
<el-input v-model="param.value" placeholder="参数值" :maxlength="2000" :disabled="Boolean(detail?.embedding_change_block_reason)" />
<el-button link type="danger" :disabled="Boolean(detail?.embedding_change_block_reason)" @click="optionalEmbeddingParams.splice(index, 1)"
>删除</el-button
>
</div>
@ -499,7 +507,7 @@ const settingsReady = computed(
)
const detail = ref<any>({})
const cloneModelId = ref('')
const embeddingParams = reactive({ batch_size: 10, normalize: false })
const embeddingParams = reactive({ batch_size: 10, normalize: true })
const embeddingParamLabels: Record<string, string> = {
query_prefix: '查询指令',
document_prefix: '文档指令',
@ -541,7 +549,7 @@ const metadataForm = ref<Array<{ name: string; value: string }>>([])
const metadataCreator = ref('')
const metadataReady = ref(false)
const segmentation = ref<any>({
chunk_size: 4096,
chunk_size: 500,
chunk_overlap: 0,
updated_by: '',
updated_at: '',
@ -568,7 +576,7 @@ function editMetadata() {
function loadSegmentation() {
if (!id) return
return knowledgeApi.getKnowledgeSegmentation(id, segmentationLoading).then((res: any) => {
segmentation.value = res.data || { chunk_size: 4096, chunk_overlap: 0 }
segmentation.value = res.data || { chunk_size: 500, chunk_overlap: 0 }
savedSegmentation.value = JSON.stringify({
chunk_size: segmentation.value.chunk_size,
chunk_overlap: segmentation.value.chunk_overlap,
@ -824,7 +832,7 @@ function getDetail() {
const params = res.data?.meta?.embedding_params || {}
Object.assign(embeddingParams, {
batch_size: params.batch_size ?? 10,
normalize: params.normalize ?? false,
normalize: params.normalize ?? true,
})
optionalEmbeddingParams.value = Object.entries(params)
.filter(([key]) => !['batch_size', 'normalize'].includes(key))

View File

@ -45,6 +45,7 @@
:placeholder="$t('views.knowledge.form.EmbeddingModel.placeholder')"
:options="modelOptions"
:model-type="'EMBEDDING'"
:disabled="embeddingDisabled"
showFooter
></ModelSelect>
</el-form-item>
@ -63,6 +64,7 @@ const props = defineProps<{
}
apiType: 'systemShare' | 'workspace' | 'systemManage'
section?: 'basic' | 'embedding'
embeddingDisabled?: boolean
formValue?: knowledgeData
}>()
const form = ref<knowledgeData>(props.formValue || {

View File

@ -1,10 +1,10 @@
<template>
<div class="hub-manager" v-loading="loading">
<el-tabs v-model="tab" @tab-change="load">
<el-tab-pane label="服务" name="services" />
<el-tab-pane label="流程" name="flows" />
<el-tab-pane label="任务" name="tasks" />
<el-tab-pane label="图谱" name="artifacts" />
<div class="hub-manager" :class="{ embedded: hideTabs }" v-loading="loading">
<el-tabs v-if="!hideTabs" v-model="tab" @tab-change="load">
<el-tab-pane label="导图服务" name="services" />
<el-tab-pane label="导图处理流程" name="flows" />
<el-tab-pane label="运行记录" name="tasks" />
<el-tab-pane label="导图管理" name="artifacts" />
</el-tabs>
<div class="mb-16">
<el-button
@ -14,11 +14,16 @@
>{{ tab === 'services' ? '发布服务' : '新建流程' }}</el-button
>
<el-button @click="load">刷新</el-button>
<template v-if="tab === 'artifacts' && resources.is_admin">
<span class="ml-16 mr-8">每人保存上限</span>
<el-input-number v-model="artifactLimit" :min="1" :max="10000" :precision="0" controls-position="right" />
<el-button class="ml-8" :loading="limitSaving" @click="saveArtifactLimit">保存</el-button>
</template>
</div>
<el-table :data="rows" row-key="id">
<el-table :data="pagedRows" row-key="id">
<el-table-column prop="name" label="名称" min-width="180" />
<el-table-column
v-if="tab === 'flows' || tab === 'artifacts'"
v-if="tab === 'flows'"
prop="revision"
label="版本"
width="90"
@ -37,14 +42,18 @@
<el-table-column label="操作" min-width="300"
><template #default="{ row }">
<el-button link type="primary" @click="open(row)">{{
tab === 'flows' ? '查看 / 新版本' : tab === 'artifacts' ? '打开 / 复核' : '详情'
tab === 'flows' ? '查看流程' : tab === 'artifacts' ? '查看导图' : tab === 'services' ? '接口信息' : '运行详情'
}}</el-button>
<el-button v-if="tab === 'flows'" link type="primary" @click="openFlowSubmit(row)"
>试运行</el-button
>
<template
v-if="tab === 'services' && (resources.is_admin || row.actions?.includes('EDIT'))"
>
<el-button link type="primary" @click="toggle(row)">{{
row.enabled ? '停用' : '启用'
}}</el-button>
<el-button link type="primary" @click="showCalls(row)">调用日志</el-button>
<el-button link type="primary" @click="rotate(row)">更新Key</el-button>
<el-button v-if="row.key_configured" link @click="revoke(row)">撤销Key</el-button>
</template>
@ -66,15 +75,16 @@
</template></el-table-column
>
</el-table>
<el-dialog v-model="dialog" :title="title" width="740px" destroy-on-close>
<SystemPagination v-model:current-page="listPage" v-model:page-size="listPageSize" :total="rows.length" />
<el-dialog v-model="dialog" :title="title" width="740px" destroy-on-close @closed="runKey = ''">
<el-form label-position="top">
<el-form-item v-if="tab !== 'tasks'" label="名称"
><el-input v-model="form.name" maxlength="128"
><el-input v-model="form.name" maxlength="128" :disabled="tab === 'services' && !canEdit"
/></el-form-item>
<template v-if="tab === 'flows'">
<el-form-item label="文档解析"
><el-select v-model="form.parser_id" clearable placeholder="不解析文件"
><el-option
><el-option label="内置解析" value="builtin" /><el-option
v-for="p in resources.parsers"
:key="p.id"
:label="p.name"
@ -84,47 +94,103 @@
><el-select v-model="form.model_id" clearable placeholder="按原文生成"
><el-option v-for="m in llms" :key="m.id" :label="m.name" :value="m.id" /></el-select
></el-form-item>
<el-form-item v-if="form.model_id" label="提示词"
><el-input v-model="form.prompt" type="textarea" :rows="4"
/></el-form-item>
<el-form-item v-if="form.model_id" label="温度"
><el-input-number v-model="form.temperature" :min="0" :max="2" :step="0.1"
/></el-form-item>
<el-form-item label="输出">思维导图</el-form-item>
<el-collapse v-if="form.model_id"
><el-collapse-item title="高级参数"
><el-form-item label="提示词"
><el-input v-model="form.prompt" type="textarea" :rows="4"
/></el-form-item>
<el-form-item label="温度"
><el-input-number v-model="form.temperature" :min="0" :max="2" :step="0.1"
/></el-form-item> </el-collapse-item></el-collapse
><el-form-item label="输出">思维导图</el-form-item>
</template>
<template v-if="tab === 'services'">
<el-form-item label="流程版本"
><el-select v-model="form.flow_id" :disabled="!!form.id"
><el-select v-model="form.flow_id" :disabled="!!form.id || !canEdit"
><el-option
v-for="f in flows"
:key="f.id"
:label="`${f.name} · v${f.revision}`"
:value="f.id" /></el-select
></el-form-item>
<el-form-item v-if="!form.id" label="每分钟调用次数"
><el-input-number v-model="form.requests_per_minute" :min="1" :max="600"
<el-form-item label="每分钟调用次数"
><el-input-number
v-model="form.requests_per_minute"
:min="1"
:max="600"
:disabled="!canEdit"
/></el-form-item>
<el-form-item label="请求方法"
><el-select v-model="form.api_config.method" :disabled="!canEdit" @change="changeMethod"
><el-option label="POST" value="POST" /><el-option
label="GET"
value="GET" /></el-select
></el-form-item>
<el-form-item label="请求参数">
<el-table :data="form.api_config.parameters">
<el-table-column label="业务字段" width="110"
><template #default="{ row }">{{ parameterLabels[row.target] }}</template></el-table-column
>
<el-table-column label="参数名称" min-width="150"
><template #default="{ row }"
><el-input v-model="row.name" :disabled="!canEdit" /></template
></el-table-column>
<el-table-column label="类型" width="90"
><template #default="{ row }">{{ row.type === 'file' ? '文件' : '文本' }}</template></el-table-column
>
<el-table-column label="必填" width="80"
><template #default="{ row }"
><el-switch v-model="row.required" :disabled="!canEdit" /></template
></el-table-column>
<el-table-column label="操作" width="70"
><template #default="{ $index }"
><el-button
link
type="danger"
:disabled="!canEdit"
@click="form.api_config.parameters.splice($index, 1)"
>删除</el-button
></template
></el-table-column
>
</el-table>
<el-select
v-if="canEdit && availableTargets.length"
placeholder="添加参数"
@change="addParameter"
><el-option
v-for="target in availableTargets"
:key="target"
:label="parameterLabels[target]"
:value="target"
/></el-select>
</el-form-item>
<el-form-item v-if="form.id" label="服务接口"
><el-input :model-value="serviceUrl(form)" readonly
/></el-form-item>
<el-form-item
v-if="form.id && (resources.is_admin || form.actions?.includes('USE'))"
label="调用"
><el-input
v-model="runMarkdown"
type="textarea"
:rows="3"
placeholder="Markdown"
/><el-input
v-model="runKey"
type="password"
show-password
placeholder="服务Key(如已启用)"
/><input
type="file"
@change="runFile = ($event.target as HTMLInputElement).files?.[0] || null"
/><el-button @click="runService">提交任务</el-button></el-form-item
>
<template v-if="form.id && (resources.is_admin || form.actions?.includes('USE'))">
<el-form-item
v-for="param in form.api_config.parameters"
:key="param.target"
:label="param.name"
:required="param.required"
>
<input
v-if="param.type === 'file'"
type="file"
@change="runFile = ($event.target as HTMLInputElement).files?.[0] || null"
/>
<el-input
v-else
v-model="runValues[param.target]"
:type="param.target === 'markdown' ? 'textarea' : 'text'"
/>
</el-form-item>
<el-form-item label="服务Key"
><el-input v-model="runKey" type="password" show-password
/></el-form-item>
<el-button @click="runService">调用服务</el-button>
</template>
</template>
<el-form-item
v-if="tab === 'services' && (resources.is_admin || form.actions?.includes('AUTHORIZE'))"
@ -161,7 +227,7 @@
><el-input v-model="reviewMarkdown" type="textarea" :rows="8"
/></el-form-item>
<el-button v-if="form.status === 'SUCCESS'" type="primary" @click="saveTaskResult"
>确认保存图谱</el-button
>保存导图</el-button
>
</template>
<template v-if="tab === 'artifacts'">
@ -179,14 +245,6 @@
:label="name"
:value="value" /></el-select
></el-form-item>
<el-form-item label="历史修订"
><el-select placeholder="查看历史" @change="showHistory"
><el-option
v-for="h in form.history || []"
:key="h.revision"
:label="`v${h.revision}`"
:value="h.revision" /></el-select
></el-form-item>
<el-button @click="openWorkbench">在工作台打开</el-button>
</template>
</el-form>
@ -197,31 +255,222 @@
}}</el-button></template
>
</el-dialog>
<el-dialog v-model="flowSubmitDialog" title="试运行流程" width="600px" destroy-on-close>
<el-form label-position="top">
<el-form-item label="Markdown"
><el-input v-model="flowMarkdown" type="textarea" :rows="6"
/></el-form-item>
<el-form-item label="文件"
><input
type="file"
@change="flowFile = ($event.target as HTMLInputElement).files?.[0] || null"
/></el-form-item>
<el-collapse
><el-collapse-item title="高级参数"
><el-form-item label="优先级"
><el-input-number
v-model="flowPriority"
:min="0"
:max="9" /></el-form-item></el-collapse-item
></el-collapse>
</el-form>
<template #footer
><el-button @click="flowSubmitDialog = false">取消</el-button
><el-button type="primary" :loading="flowSubmitting" @click="submitFlow"
>提交任务</el-button
></template
>
</el-dialog>
<el-dialog v-model="callsDialog" title="调用日志" width="900px">
<el-button @click="loadCalls">刷新</el-button>
<el-table :data="calls" v-loading="callsLoading">
<el-table-column prop="created_at" label="时间" min-width="160" /><el-table-column
prop="method"
label="方法"
width="75"
/>
<el-table-column label="状态" width="100"
><template #default="{ row }">{{
statuses[row.status] || row.status
}}</template></el-table-column
>
<el-table-column prop="http_status" label="HTTP" width="75" />
<el-table-column label="HTTP耗时" width="100"
><template #default="{ row }">{{ duration(row.elapsed_ms) }}</template></el-table-column
>
<el-table-column prop="task_id" label="任务ID" min-width="180" show-overflow-tooltip />
<el-table-column label="任务状态" width="100"
><template #default="{ row }">{{
statuses[row.task_status] || row.task_status || '-'
}}</template></el-table-column
>
<el-table-column label="任务耗时" width="100"
><template #default="{ row }">{{ duration(row.execution_ms) }}</template></el-table-column
>
<el-table-column label="错误" min-width="150" show-overflow-tooltip
><template #default="{ row }">{{
row.error ? String(row.error).slice(0, 120) : '-'
}}</template></el-table-column
>
</el-table>
<SystemPagination
v-model:current-page="callsPage"
v-model:page-size="callsPageSize"
:total="callsTotal"
@change="changeCallsPage"
/>
</el-dialog>
</div>
</template>
<script setup lang="ts">
import { computed, onMounted, ref } from 'vue'
import { computed, onMounted, ref, watch } from 'vue'
import { ElMessage, ElMessageBox } from 'element-plus'
import { hubGet, hubPost, hubDelete, hubBase } from '@/api/ai-hub'
import { request } from '@/request/index'
const emit = defineEmits(['open-artifact'])
import SystemPagination from '@/components/system-pagination/index.vue'
const emit = defineEmits(['open-artifact', 'tab-change'])
const props = withDefaults(defineProps<{initialTab?: 'services' | 'flows' | 'tasks' | 'artifacts'; hideTabs?: boolean}>(), {initialTab: 'services', hideTabs: false})
function openWorkbench() {
emit('open-artifact', form.value)
dialog.value = false
}
const tab = ref('services'),
const tab = ref(props.initialTab),
loading = ref(false),
saving = ref(false),
dialog = ref(false)
watch(tab, value => { listPage.value = 1; if (props.hideTabs) emit('tab-change', value) })
const rows = ref<any[]>([]),
flows = ref<any[]>([]),
form = ref<any>({}),
resources = ref<any>({ parsers: [], models: [] })
const artifactLimit = ref(10), limitSaving = ref(false)
async function saveArtifactLimit() {
limitSaving.value = true
try {
artifactLimit.value = (await hubPost('artifact_limit', { limit: artifactLimit.value })).data.limit
ElMessage.success('已保存')
} finally { limitSaving.value = false }
}
const listPage = ref(1), listPageSize = ref(10)
const pagedRows = computed(() => rows.value.slice((listPage.value - 1) * listPageSize.value, listPage.value * listPageSize.value))
watch([() => rows.value.length, listPageSize], ([total, size]) => { listPage.value = Math.min(listPage.value, Math.max(1, Math.ceil(total / size))) })
const reviewMarkdown = ref(''),
runMarkdown = ref(''),
runValues = ref<Record<string, string>>({}),
runKey = ref(''),
runFile = ref<File | null>(null)
const flowSubmitDialog = ref(false),
flowSubmitting = ref(false),
flowId = ref(''),
flowMarkdown = ref(''),
flowFile = ref<File | null>(null),
flowPriority = ref(5)
function openFlowSubmit(row: any) {
flowId.value = row.id
flowMarkdown.value = ''
flowFile.value = null
flowPriority.value = 5
flowSubmitDialog.value = true
}
async function submitFlow() {
if (!flowMarkdown.value.trim() && !flowFile.value) {
ElMessage.warning('请输入 Markdown 或选择文件')
return
}
flowSubmitting.value = true
try {
const payload = {
flow_id: flowId.value,
markdown: flowMarkdown.value,
priority: flowPriority.value,
}
if (flowFile.value) {
const data = new FormData()
data.append('data', JSON.stringify(payload))
data.append('file', flowFile.value)
await request({ url: `${hubBase()}/tasks`, method: 'post', data })
} else await hubPost('tasks', payload)
ElMessage.success('任务已提交')
flowSubmitDialog.value = false
tab.value = 'tasks'
await load()
} finally {
flowSubmitting.value = false
}
}
const callsDialog = ref(false),
callsLoading = ref(false),
callsPage = ref(1),
callsPageSize = ref(20),
callsTotal = ref(0)
let previousCallsPageSize = 20
function changeCallsPage(_page: number, size: number) {
if (size !== previousCallsPageSize) callsPage.value = 1
previousCallsPageSize = size
loadCalls()
}
const calls = ref<any[]>([]),
callsServiceId = ref('')
const parameterLabels: Record<string, string> = { markdown: 'Markdown 内容', name: '任务名称', file: '输入文件' }
const defaultApiConfig = () => ({
method: 'POST',
parameters: ['markdown', 'name', 'file'].map((target) => ({
name: target,
target,
type: target === 'file' ? 'file' : 'string',
required: false,
})),
})
const canEdit = computed(
() => !form.value.id || resources.value.is_admin || form.value.actions?.includes('EDIT'),
)
const availableTargets = computed(() =>
['markdown', 'name', 'file'].filter(
(target) =>
!(target === 'file' && form.value.api_config?.method === 'GET') &&
!form.value.api_config?.parameters?.some((p: any) => p.target === target),
),
)
function changeMethod(method: string) {
if (method === 'GET') {
form.value.api_config.parameters = form.value.api_config.parameters.filter(
(p: any) => p.target !== 'file',
)
runFile.value = null
}
}
function addParameter(target: string) {
form.value.api_config.parameters.push({
name: target,
target,
type: target === 'file' ? 'file' : 'string',
required: false,
})
}
function duration(value: number | null) {
return value == null ? '-' : `${value} ms`
}
async function showCalls(row: any) {
callsServiceId.value = row.id
callsPage.value = 1
calls.value = []
callsDialog.value = true
await loadCalls()
}
async function loadCalls() {
callsLoading.value = true
try {
const result = await request({
url: `${hubBase()}/services/${callsServiceId.value}/calls`,
method: 'get',
params: { current_page: callsPage.value, page_size: callsPageSize.value },
})
calls.value = result.data.data.records || []
callsTotal.value = result.data.data.total || 0
} finally {
callsLoading.value = false
}
}
const llms = computed(() => resources.value.models.filter((m: any) => m.model_type === 'LLM'))
const statuses: Record<string, string> = {
PENDING: '待处理',
@ -237,7 +486,7 @@ const nodeLabels: Record<string, string> = {
const layouts = { tree: '树状' }
const title = computed(
() =>
({ services: '导图服务', flows: '处理流程', tasks: '任务详情', artifacts: '图谱复核' })[
({ services: '导图服务', flows: '导图处理流程', tasks: '运行详情', artifacts: '导图详情' })[
tab.value as 'services'
],
)
@ -257,6 +506,8 @@ async function load() {
hubGet('flows'),
])
resources.value = r.data
if (tab.value === 'artifacts' && resources.value.is_admin)
artifactLimit.value = (await hubGet('artifact_limit')).data.limit
rows.value = list.data
flows.value = f.data
} finally {
@ -273,28 +524,45 @@ function create() {
prompt: '保留原文事实、数值和语言,整理为分级Markdown标题和要点。',
temperature: 0.2,
requests_per_minute: 20,
api_config: defaultApiConfig(),
}
dialog.value = true
}
async function open(row: any) {
form.value = (await hubGet(tab.value, row.id)).data
if (tab.value === 'services') form.value.api_config = form.value.api_config || defaultApiConfig()
if (tab.value === 'flows') {
const p = form.value.nodes.find((n: any) => n.type === 'parse'),
m = form.value.nodes.find((n: any) => n.type === 'organize')
Object.assign(form.value, {
parser_id: p?.resource_id || '',
parser_id: p?.processor === 'builtin' ? 'builtin' : p?.resource_id || '',
model_id: m?.model_id || '',
prompt: m?.prompt || '',
temperature: m?.temperature ?? 0.2,
})
}
reviewMarkdown.value = form.value.result?.markdown || ''
runMarkdown.value = ''
runValues.value = {}
runKey.value = ''
runFile.value = null
dialog.value = true
}
async function save() {
if (tab.value === 'services' && canEdit.value) {
const config = form.value.api_config
const names = config.parameters.map((p: any) => p.name.trim())
if (names.some((name: string) => !name) || new Set(names).size !== names.length) {
ElMessage.warning('参数名称不能为空或重复')
return
}
if (config.method === 'GET' && config.parameters.some((p: any) => p.target === 'file')) {
ElMessage.warning('GET 不支持文件参数')
return
}
config.parameters.forEach((p: any) => {
p.name = p.name.trim()
})
}
saving.value = true
try {
let body: any = { ...form.value },
@ -315,7 +583,11 @@ async function save() {
body = { name: body.name, previous_id: id, nodes: [] }
id = undefined
if (form.value.parser_id)
body.nodes.push({ type: 'parse', resource_id: form.value.parser_id })
body.nodes.push(
form.value.parser_id === 'builtin'
? { type: 'parse', processor: 'builtin' }
: { type: 'parse', resource_id: form.value.parser_id },
)
if (form.value.model_id)
body.nodes.push({
type: 'organize',
@ -342,10 +614,15 @@ async function remove(row: any) {
const effects: Record<string, string> = {
services: '对应授权和服务Key将失效,已有任务与图谱保留。',
flows: '已有任务与图谱保留。',
tasks: '任务记录及输入文件将删除,已保存图谱保留。',
artifacts: '图谱及其历史修订将删除,原任务保留。',
tasks: '任务记录及输入文件将删除,已保存导图保留。',
artifacts: '导图将删除,原任务保留。',
}
const confirmed = await ElMessageBox.confirm(`删除“${row.name}”?${effects[kind]}`, '确认删除', { type: 'warning' }).then(() => true, () => false)
const confirmed = await ElMessageBox.confirm(`删除“${row.name}”?${effects[kind]}`, '确认删除', {
type: 'warning',
}).then(
() => true,
() => false,
)
if (!confirmed) return
await hubDelete(kind, row.id)
ElMessage.success('已删除')
@ -370,16 +647,30 @@ function serviceUrl(row: any) {
return `${location.origin}${import.meta.env.BASE_URL}api${hubBase()}/published/${row.id}/run`
}
async function runService() {
const config = form.value.api_config
if (config.method === 'GET' && config.parameters.some((p: any) => p.type === 'file')) {
ElMessage.warning('GET 不支持文件参数')
return
}
const values: Record<string, string | File> = {}
for (const param of config.parameters) {
const value = param.type === 'file' ? runFile.value : runValues.value[param.target]
if (param.required && !value) {
ElMessage.warning(`请填写${param.name}`)
return
}
if (value) values[param.name] = value
}
const data = new FormData()
data.append('data', JSON.stringify({ markdown: runMarkdown.value }))
if (runFile.value) data.append('file', runFile.value)
Object.entries(values).forEach(([name, value]) => data.append(name, value))
await request({
url: `${hubBase()}/published/${form.value.id}/run`,
method: 'post',
data,
headers: { 'X-Service-Key': runKey.value },
method: config.method.toLowerCase(),
...(config.method === 'GET' ? { params: values } : { data }),
headers: runKey.value ? { 'X-Service-Key': runKey.value } : {},
timeout: 30000,
})
runKey.value = ''
ElMessage.success('任务已提交')
dialog.value = false
tab.value = 'tasks'
@ -397,18 +688,6 @@ async function saveTaskResult() {
tab.value = 'artifacts'
await load()
}
function showHistory(revision: number) {
const h = form.value.history.find((item: any) => item.revision === revision)
if (h)
ElMessageBox.alert(h.markdown, `v${revision}`, {
customStyle: {
maxWidth: '80vw',
whiteSpace: 'pre-wrap',
maxHeight: '80vh',
overflow: 'auto',
},
})
}
onMounted(load)
</script>
<style scoped>
@ -417,6 +696,11 @@ onMounted(load)
height: 100%;
overflow: auto;
}
.hub-manager.embedded {
padding: 0;
height: auto;
overflow: visible;
}
.el-select {
width: 100%;
}

View File

@ -18,6 +18,7 @@ export interface ParseParams {
}
export interface ParseResult {
task_id?: string
backend: string
version: string
results: Record<
@ -28,6 +29,8 @@ export interface ParseResult {
model_output?: string
content_list?: string
images?: Record<string, string>
original_md?: string
heading_structure?: any
}
>
}
@ -41,7 +44,15 @@ export interface ParseProgress {
}
export interface MindmapOrganizeProgress extends ParseProgress {
plan?: Record<string, any>
task_id?: string
result_md?: string | null
heading_structure?: any
original_md?: string
sources?: any[]
warnings?: string[]
summary_input?: { markdown?: string; heading_structure?: any }
summary_node?: Record<string, any>
}
const ids = new Map<string, string>()
@ -52,7 +63,7 @@ const states: Record<string, ParseProgress['status']> = {
FAILED: 'failed',
}
async function progress(taskId: string): Promise<MindmapOrganizeProgress> {
const id = ids.get(taskId)
const id = ids.get(taskId) || (/^[a-f0-9-]{36}$/i.test(taskId) ? taskId : '')
if (!id)
return { progress: null, status: 'pending', stage: '准备中', error: null, file_names: '' }
const task = await hubCall('tasks', 'get', undefined, id)
@ -70,10 +81,40 @@ async function progress(taskId: string): Promise<MindmapOrganizeProgress> {
: '失败'),
error: task.error,
file_names: task.name,
task_id: task.id,
result_md: task.result?.markdown,
heading_structure: task.result?.source?.heading_structure,
original_md: task.result?.original,
plan: task.progress?.plan || task.result?.source?.organization || task.result?.summary_state?.metrics,
sources: task.result?.source?.summary_sources || [],
warnings: task.result?.source?.summary_warnings || [],
summary_input: task.summary_input,
summary_node: task.snapshot?.nodes?.find((node: any) => node.type === 'organize'),
}
}
export interface SummaryOptions {
context_tokens: number | null
max_input_tokens: number | null
input_tokens: number | null
max_output_tokens: number | null
output_tokens: number | null
safety_tokens: number | null
top_p: number | null
timeout: number
retries: number
concurrency: number
max_depth: number
max_children: number
node_max_chars: number
detail: 'brief' | 'standard' | 'detailed'
structure_mode: 'preserve' | 'reorganize'
parallel: boolean
auto_split: boolean
}
export const documentApi = {
async previewSummary(markdown: string, node: Record<string, unknown>) {
return hubCall('summary_preview', 'post', { markdown, nodes: [node, { type: 'render' }] })
},
async parseDocument(params: ParseParams, taskId: string = uuid()): Promise<ParseResult> {
const results: ParseResult['results'] = {}
for (const file of params.files) {
@ -108,6 +149,8 @@ export const documentApi = {
results[file.name] = {
md_content: current.result.markdown,
middle_json: JSON.stringify(current.result.source),
original_md: current.result.original,
heading_structure: current.result.source?.heading_structure,
}
completed = true
break
@ -116,7 +159,7 @@ export const documentApi = {
}
if (!completed) throw new Error('任务仍在处理中,请到管理页查看')
}
return { backend: 'shared', version: '1', results }
return { backend: 'shared', version: '1', results, task_id: ids.get(taskId) }
},
async createParseTask(_taskId: string): Promise<ParseProgress> {
return { progress: null, stage: '准备中', status: 'pending', error: null, file_names: '' }
@ -127,19 +170,30 @@ export const documentApi = {
markdown: string,
_mode = 'smart',
prompt?: string,
config?: { model_id: string; temperature?: number | null },
config?: { model_id: string; temperature?: number | null; options?: SummaryOptions },
context?: { heading_structure?: any; original?: string },
): Promise<MindmapOrganizeProgress> {
if (!config?.model_id) throw new Error('请选择整理模型')
if (!config?.model_id) throw new Error('请选择总结模型')
const node: Record<string, unknown> = { type: 'organize', model_id: config.model_id, prompt }
if (config.temperature !== null && config.temperature !== undefined)
node.temperature = config.temperature
Object.entries(config.options || {}).forEach(([key, value]) => {
if (value !== null && value !== undefined) node[key] = value
})
const task = await hubCall('tasks', 'post', {
name: '智能整理',
name: '智能总结',
markdown,
nodes: [node, { type: 'render' }],
...context,
})
ids.set(taskId, task.id)
return progress(taskId)
return { progress: null, status: 'pending', stage: '等待任务处理', error: null,
file_names: task.name || '智能总结', task_id: task.id }
},
getMindmapProgress: progress,
async retryMindmapTask(taskId: string) {
const id = ids.get(taskId) || (/^[a-f0-9-]{36}$/i.test(taskId) ? taskId : '')
if (!id) throw new Error('任务已失效,请重新总结')
return hubCall('tasks', 'post', {}, id)
},
}

View File

@ -1,5 +1,6 @@
<template>
<div class="config-panel">
<div class="config-panel" :class="{ 'platform-config': platform }">
<div v-if="!platform" class="export-hint">个人设置仅在当前页面会话有效,刷新后恢复平台默认。</div>
<el-form label-position="top">
<el-form-item label="PDF / 图片解析服务"
><el-select v-model="draft.parserResourceId" placeholder="请选择"
@ -9,21 +10,6 @@
:label="p.name"
:value="p.id" /></el-select
></el-form-item>
<el-form-item label="整理模型"
><el-select v-model="draft.mindmapModelId" clearable placeholder="请选择"
><el-option v-for="m in models" :key="m.id" :label="m.name" :value="m.id" /></el-select
></el-form-item>
<el-form-item label="温度(留空跟随模型配置)"
><el-input-number
v-model="draft.mindmapLlmTemperature"
:min="0"
:max="2"
:step="0.1"
placeholder="跟随模型配置"
/></el-form-item>
<el-form-item label="整理提示词"
><el-input v-model="draft.mindmapPrompt" type="textarea" :rows="6"
/></el-form-item>
<el-form-item label="导出长边">
<el-select v-model="draft.exportResolution" aria-label="导出长边">
<el-option label="1024 px" :value="1024" />
@ -33,10 +19,35 @@
</el-select>
<span class="export-hint">仅 PNG/JPEG;SVG 保留矢量尺寸。</span>
</el-form-item>
<div v-if="selectedModel" class="export-hint effective-summary-budget">
本次输入 ≤ {{ effectiveInput.toLocaleString() }} Token · 输出 ≤ {{ effectiveOutput.toLocaleString() }} Token · 保留原文语言
</div>
<el-form-item label="总结模型"
><el-select v-model="draft.mindmapModelId" clearable placeholder="请选择" @change="draft.summaryOptions.top_p = null"
><el-option v-for="m in models" :key="m.id" :label="m.name" :value="m.id" /></el-select
></el-form-item>
<el-collapse v-model="advancedPanels">
<el-collapse-item title="高级参数" name="advanced">
<span class="export-hint">长度单位为 Token,留空使用默认值。</span>
<el-form-item v-for="field in visibleBudgetFields" :key="field.key" :label="field.label">
<el-input-number v-model="draft.summaryOptions[field.key]" :min="field.min" :max="field.max"
:step="field.step || 1" :placeholder="modelDefault(field.key)" :aria-label="field.label" />
</el-form-item>
<el-form-item label="总结粒度">
<el-select v-model="draft.summaryOptions.detail"><el-option label="精简" value="brief" /><el-option label="标准" value="standard" /><el-option label="详细" value="detailed" /></el-select>
</el-form-item>
<el-form-item label="结构策略">
<el-select v-model="draft.summaryOptions.structure_mode"><el-option label="保留原目录" value="preserve" /><el-option label="按主题重组" value="reorganize" /></el-select>
</el-form-item>
<el-form-item label="总结提示词"
><el-input v-model="draft.mindmapPrompt" type="textarea" :rows="6"
/></el-form-item>
</el-collapse-item>
</el-collapse>
<div v-if="error" class="error">{{ error }}</div>
<div class="actions">
<el-button @click="load">刷新资源</el-button
><el-button type="primary" @click="confirm">确认</el-button>
><el-button type="primary" :loading="saving" @click="confirm">{{ platform ? '保存' : '确认' }}</el-button>
</div>
</el-form>
</div>
@ -44,20 +55,54 @@
<script setup lang="ts">
import { computed, onMounted, reactive, ref, watch } from 'vue'
import type { DocumentConfig } from '@/views/mindmap/workbench/composables/useDocumentProcessor'
import type { SummaryOptions } from '@/views/mindmap/workbench/api/document'
import { hubCall } from '@/views/mindmap/workbench/api/hub'
import { DEFAULT_SUMMARY_BUDGET } from '@/views/mindmap/workbench/utils/contextDefaults'
const props = defineProps<{
modelValue: DocumentConfig
backendOptions: any[]
languageOptions: any[]
platform?: boolean
saving?: boolean
}>()
const emit = defineEmits(['update:modelValue', 'backendChange', 'close'])
const draft = reactive({ ...props.modelValue })
const emit = defineEmits(['update:modelValue', 'backendChange', 'close', 'confirm'])
const copyConfig = (value: DocumentConfig) => ({ ...value, summaryOptions: { ...value.summaryOptions } })
const draft = reactive(copyConfig(props.modelValue))
const advancedPanels = ref<string[]>([])
const resources = ref<any>({ parsers: [], models: [] }),
error = ref('')
const models = computed(() => resources.value.models.filter((m: any) => m.model_type === 'LLM'))
const selectedModel = computed(() => models.value.find((m: any) => m.id === draft.mindmapModelId))
type NumericKey = { [K in keyof SummaryOptions]: SummaryOptions[K] extends number | null ? K : never }[keyof SummaryOptions]
const budgetFields: { key: NumericKey; label: string; min: number; max: number; step?: number }[] = [
{ key: 'context_tokens', label: '上下文窗口', min: 4096, max: 2000000 },
{ key: 'max_input_tokens', label: '最大输入长度', min: 1024, max: 2000000 },
{ key: 'input_tokens', label: '本次输入上限', min: 1024, max: 2000000 },
{ key: 'max_output_tokens', label: '最大输出长度', min: 256, max: 524288 },
{ key: 'output_tokens', label: '本次输出上限', min: 256, max: 524288 },
{ key: 'safety_tokens', label: '安全预留长度', min: 256, max: 200000 },
{ key: 'timeout', label: '调用超时(秒)', min: 30, max: 600 },
{ key: 'retries', label: '失败重试次数', min: 0, max: 2 },
{ key: 'concurrency', label: '并发上限', min: 1, max: 8 },
{ key: 'max_depth', label: '最大层级', min: 2, max: 6 },
{ key: 'max_children', label: '子节点数量目标', min: 2, max: 30 },
{ key: 'node_max_chars', label: '节点长度目标', min: 10, max: 300 },
]
const visibleBudgetFields = budgetFields.filter((field) => ['input_tokens', 'output_tokens'].includes(field.key))
function modelDefault(key: string) {
return String(selectedModel.value?.summary_defaults?.[key]
?? DEFAULT_SUMMARY_BUDGET[key as keyof typeof DEFAULT_SUMMARY_BUDGET] ?? '跟随模型配置')
}
const effectiveOutput = computed(() => Number(draft.summaryOptions.output_tokens ?? selectedModel.value?.summary_defaults?.output_tokens ?? DEFAULT_SUMMARY_BUDGET.output_tokens))
const effectiveInput = computed(() => Math.max(0, Math.min(
Number(draft.summaryOptions.context_tokens ?? selectedModel.value?.summary_defaults?.context_tokens ?? DEFAULT_SUMMARY_BUDGET.context_tokens)
- effectiveOutput.value - Number(draft.summaryOptions.safety_tokens ?? selectedModel.value?.summary_defaults?.safety_tokens ?? DEFAULT_SUMMARY_BUDGET.safety_tokens),
Number(draft.summaryOptions.max_input_tokens ?? selectedModel.value?.summary_limits?.max_input_tokens ?? DEFAULT_SUMMARY_BUDGET.max_input_tokens),
Number(draft.summaryOptions.input_tokens ?? selectedModel.value?.summary_defaults?.input_tokens ?? DEFAULT_SUMMARY_BUDGET.input_tokens),
)))
watch(
() => props.modelValue,
(value) => Object.assign(draft, value),
(value) => Object.assign(draft, copyConfig(value)),
{ deep: true },
)
async function load() {
@ -67,14 +112,31 @@ async function load() {
if (!draft.parserResourceId && resources.value.parsers?.length)
draft.parserResourceId = resources.value.parsers[0].id
if (!draft.mindmapModelId && models.value.length) draft.mindmapModelId = models.value[0].id
if (draft.parserResourceId || draft.mindmapModelId) emit('update:modelValue', { ...draft })
} catch (e: any) {
error.value = e.message
}
}
function confirm() {
emit('update:modelValue', { ...draft })
emit('close')
const options = draft.summaryOptions
for (const field of budgetFields) {
const value = options[field.key]
if (value != null && (!Number.isInteger(value) || value < field.min || value > field.max)) {
error.value = `${field.label}超出范围`; return
}
}
for (const key of ['context_tokens', 'max_input_tokens', 'max_output_tokens'] as const) {
const limit = selectedModel.value?.summary_limits?.[key]
if (limit && options[key] && options[key]! > Number(limit)) { error.value = '参数超过模型配置上限'; return }
}
const context = options.context_tokens ?? selectedModel.value?.summary_defaults?.context_tokens ?? DEFAULT_SUMMARY_BUDGET.context_tokens
const output = options.output_tokens ?? selectedModel.value?.summary_defaults?.output_tokens ?? DEFAULT_SUMMARY_BUDGET.output_tokens
const safety = options.safety_tokens ?? selectedModel.value?.summary_defaults?.safety_tokens ?? DEFAULT_SUMMARY_BUDGET.safety_tokens
if (output + safety >= context || output > Number(options.max_output_tokens ?? selectedModel.value?.summary_limits?.max_output_tokens ?? DEFAULT_SUMMARY_BUDGET.max_output_tokens)) {
error.value = '输出与预留长度超过模型预算'; return
}
emit('update:modelValue', copyConfig(draft))
emit('confirm', copyConfig(draft))
if (!props.platform) emit('close')
}
onMounted(load)
</script>
@ -82,6 +144,9 @@ onMounted(load)
.config-panel {
padding: 12px;
}
.config-panel.platform-config {
padding: 0;
}
.el-select {
width: 100%;
}

View File

@ -38,7 +38,7 @@
<div ref="htmlRef" class="markdown-content rendered-html" v-html="renderedContent"></div>
</div>
<div v-else-if="mode === 'pdf'" class="render-shell">
<div v-else-if="mode === 'pdf'" class="render-shell pdf-render-shell">
<iframe v-if="pdfUrl" :src="pdfUrl" class="actual-pdf-preview" title="PDF预览"></iframe>
<el-empty v-else description="点击生成预览,查看实际分页效果" />
</div>
@ -67,6 +67,7 @@ import { markdownEngine } from '@/views/mindmap/workbench/utils/graphModel'
import { hubCall } from '@/views/mindmap/workbench/api/hub'
import 'katex/dist/katex.min.css'
import { useImageAssets, hasImageAssets, exportMarkdownBundle, embedImages } from '@/views/mindmap/workbench/utils/imageAssets'
import { createExportImageCache, runLimited } from '@/views/mindmap/workbench/utils/exportPreparation'
type RenderMode = 'markdown' | 'html' | 'pdf' | 'richtext'
type FlavorMode = 'commonmark' | 'gfm'
@ -126,13 +127,18 @@ function createMarkdownRenderer(flavor: FlavorMode) {
const exporting = ref(false), pdfUrl = ref(''), exportStage = ref('')
const exportCache = new Map<string, Blob>()
const formulaImages = createExportImageCache()
let pdfBlob: Blob | undefined, preparedHtml: Promise<string> | undefined, exportRevision = 0
function revokePdf() { if (pdfUrl.value) URL.revokeObjectURL(pdfUrl.value); pdfUrl.value = ''; pdfBlob = undefined }
function invalidateExports() { exportRevision++; exportCache.clear(); preparedHtml = undefined; revokePdf() }
watch(() => props.content, invalidateExports)
onBeforeUnmount(invalidateExports)
onBeforeUnmount(() => { invalidateExports(); formulaImages.clear() })
const elapsed = (start: number) => Math.round(performance.now() - start)
let preparationTimings: Record<string, number> = {}
async function prepareDocumentHtml() {
const source = props.content
const revision = exportRevision, timings: Record<string, number> = {}, started = performance.now()
const checkCurrent = () => { if (revision !== exportRevision) throw new Error('内容已更新,请重新导出') }
if (assetsLoading.value) {
exportStage.value = '正在加载图片'
await new Promise<void>(resolve => { const stop = watch(assetsLoading, loading => { if (!loading) { stop(); resolve() } }, { flush:'sync' }) })
@ -147,31 +153,63 @@ async function prepareDocumentHtml() {
document.body.appendChild(holder)
try {
await document.fonts.ready
checkCurrent()
timings.assetsAndFontsMs = elapsed(started)
const formulas = Array.from(holder.querySelectorAll<HTMLElement>('.export-math'))
if (formulas.length) {
const formulaStarted = performance.now()
exportStage.value = `正在处理公式 0/${formulas.length}`
const { toPng, getFontEmbedCSS } = await import('html-to-image')
const fontEmbedCSS = await getFontEmbedCSS(holder)
const formulaImages = new Map<string, string>()
for (const formula of formulas) {
const fontEmbedCSS = await getFontEmbedCSS(holder, { preferredFontFormat: 'woff2' })
// Include font content in the signature without retaining it in every cache key.
let fontSignature = 2166136261
for (let i = 0; i < fontEmbedCSS.length; i++) fontSignature = Math.imul(fontSignature ^ fontEmbedCSS.charCodeAt(i), 16777619)
const jobs = formulas.map(formula => {
const target = formula.querySelector<HTMLElement>('.katex') || formula
target.style.display = 'inline-block'
return { formula, target }
}).map(({ formula, target }) => {
const rect = target.getBoundingClientRect()
const style = getComputedStyle(target)
const width = Math.max(1, Math.ceil(rect.width)), height = Math.max(1, Math.ceil(rect.height))
const key = JSON.stringify([formula.dataset.formula, width, height, !!formula.querySelector('.katex-display'), style.fontFamily, style.fontSize, style.fontWeight, style.fontStyle, style.letterSpacing, style.color, style.lineHeight, fontSignature])
return { formula, target, width, height, key, src: '' }
})
let completed = 0
await runLimited(jobs, 3, async job => {
checkCurrent()
job.src = await formulaImages.get(job.key, () => toPng(job.target, { width:job.width, height:job.height, pixelRatio:2, backgroundColor:'#ffffff', cacheBust:false, fontEmbedCSS }))
checkCurrent()
exportStage.value = `正在处理公式 ${++completed}/${jobs.length}`
})
// Replace only after all measurements and renders settle, avoiding layout churn.
for (const { formula, src, width, height } of jobs) {
const img = document.createElement('img')
const key = `${formula.dataset.formula}:${rect.width}:${rect.height}`
if (!formulaImages.has(key)) formulaImages.set(key, await toPng(target, { width:Math.max(1, Math.ceil(rect.width)), height:Math.max(1, Math.ceil(rect.height)), pixelRatio:2, backgroundColor:'#ffffff', cacheBust:false, fontEmbedCSS }))
img.src = formulaImages.get(key)!
img.width = Math.ceil(rect.width); img.height = Math.ceil(rect.height)
img.src = src
img.width = width; img.height = height
img.alt = formula.dataset.formula || formula.textContent || '公式'
formula.replaceWith(img)
}
timings.formulasMs = elapsed(formulaStarted)
timings.formulaCount = formulas.length
}
return await embedImages(holder.innerHTML)
checkCurrent()
const imagesStarted = performance.now()
const html = await embedImages(holder.innerHTML, false, (done, total) => {
checkCurrent(); exportStage.value = `正在整理图片 ${done}/${total}`
})
checkCurrent()
timings.imagesMs = elapsed(imagesStarted)
timings.totalPreparationMs = elapsed(started)
preparationTimings = timings
return html
} finally { holder.remove() }
}
async function documentBlob(format: 'docx' | 'pdf' | 'rtf') {
const cached = exportCache.get(format)
if (cached) return cached
const revision = exportRevision
const started = performance.now(), reusedPreparation = !!preparedHtml
exportStage.value = '正在整理文档'
if (!preparedHtml) {
const pending = prepareDocumentHtml()
@ -179,8 +217,14 @@ async function documentBlob(format: 'docx' | 'pdf' | 'rtf') {
void pending.catch(() => { if (preparedHtml === pending) preparedHtml = undefined })
}
const html = await preparedHtml
if (revision !== exportRevision) throw new Error('内容已更新,请重新导出')
const preparationWaitMs = elapsed(started), requestStarted = performance.now()
exportStage.value = `正在生成 ${format.toUpperCase()}`
const result = await hubCall('document_export', 'post', { html, format, paper:'A4' })
if (revision !== exportRevision) throw new Error('内容已更新,请重新导出')
// Numeric metadata only: never log document text, formulas, images or credentials.
console.info('[document-export]', { format, reusedPreparation, preparationWaitMs, ...preparationTimings,
requestMs: elapsed(requestStarted), totalMs: elapsed(started), serverTimings: result.serverTimings || '' })
const blob = new Blob([result.content], { type:result.mime })
if (revision === exportRevision) exportCache.set(format, blob)
return blob
@ -240,7 +284,9 @@ function createFullHtmlDocument(innerHtml: string, title: string) {
}
async function exportHtml(innerHtml: string, fileName: string) {
const html = createFullHtmlDocument(await embedImages(innerHtml), fileName)
const html = createFullHtmlDocument(await embedImages(innerHtml, false, (done, total) => {
exportStage.value = `正在整理图片 ${done}/${total}`
}), fileName)
downloadBlob(new Blob([html], { type: 'text/html;charset=utf-8' }), fileName)
}
@ -285,7 +331,8 @@ async function handleDownload() {
<style scoped>
.export-stage { font-size:12px; color:var(--el-text-color-secondary); }
.actual-pdf-preview { display:block; width:100%; height:100%; min-height:520px; border:0; }
.actual-pdf-preview { display:block; width:100%; flex:1; min-height:0; border:0; }
.pdf-render-shell { height:100%; min-height:0; }
.markdown-renderer {
width: 100%;
height: 100%;

View File

@ -37,10 +37,12 @@
<el-option label="不自动换行" :value="0" />
</el-select>
<el-switch v-model="showImages" active-text="显示图片" aria-label="显示图片" size="small" />
<el-button size="small" @click="resetView"><el-icon><Refresh /></el-icon><span>重置视图</span></el-button>
</div>
</el-popover>
<span class="content-toolbar-meta" :title="nodeSummary">{{ totalNodeCount }} 节点<span v-if="hiddenNodeCount"> · 隐藏 {{ hiddenNodeCount }}</span></span>
<el-button size="small" :disabled="!editable || !content" @click="emit('heading-edit')">标题层级</el-button>
<el-button size="small" @click="resetView"><el-icon><Refresh /></el-icon><span>重置视图</span></el-button>
<span v-if="isExporting" class="content-toolbar-meta">{{ exportStage }}</span>
<el-popover v-model:visible="downloadVisible" trigger="click" placement="bottom-end" :width="220">
<template #reference>
<el-button type="primary" size="small" :disabled="!content || isRendering || isExporting" aria-label="下载">
@ -94,6 +96,7 @@ const props = defineProps({
title: { type: String, default: '思维导图' },
sourceName: { type: String, default: '未命名文档' },
editable: { type: Boolean, default: true },
excludedLines: { type: Array as () => number[], default: () => [] },
content: {
type: String,
default: ''
@ -111,6 +114,7 @@ const props = defineProps({
const emit = defineEmits<{
(e: 'node-edit', payload: { start: number; end: number; original: string; newText: string }): void
(e: 'update:layout', value: GraphLayout): void
(e: 'heading-edit'): void
}>()
const layout = computed({ get: () => props.layout as GraphLayout, set: value => emit('update:layout', value) })
@ -119,7 +123,7 @@ const downloadVisible = ref(false)
const exportScope = ref<'filtered' | 'viewport' | 'full'>('filtered')
let exportingFull = false, skipRemember = false
const fullTree = computed(() => {
const root = graphFromMarkdown(props.content, props.title)
const root = graphFromMarkdown(props.content, props.title, props.excludedLines)
const original = [...props.content.matchAll(/\/admin\/api\/workspace\/[^/\s"<>]+\/ai_hub\/assets\/[a-f0-9-]{36}\/content/g)]
const resolved = [...graphContent.value.matchAll(/blob:[^\s)"<>]+/g)]
if (showImages.value && original.length === resolved.length) walkGraph(root).forEach(node => { original.forEach((match, i) => { node.content = node.content.split(match[0]).join(resolved[i][0]) }) })
@ -200,13 +204,13 @@ function readNodeWidth() {
}
}
const nodeWidth = ref(readNodeWidth())
const viewportSize = ref('')
watch(nodeWidth, width => {
try { localStorage.setItem(NODE_WIDTH_KEY, String(width)) } catch { /* Storage may be disabled. */ }
})
const scale = ref(1)
const isRendering = ref(false)
const isExporting = ref(false)
const exportStage = ref('')
const totalNodeCount = ref(0)
const hiddenNodeCount = ref(0)
@ -471,28 +475,18 @@ async function customMap(root: GraphNode, maxWidth: number): Promise<any> {
}
// 监听内容变化自动重绘
watch([graphContent, showImages, nodeWidth, layout, () => matchesOnly.value ? search.value : '', matchesOnly, depthLimit, branch, () => props.documentKey], () => {
watch([graphContent, showImages, nodeWidth, layout, () => matchesOnly.value ? search.value : '', matchesOnly, depthLimit, branch, () => props.documentKey, () => props.excludedLines], () => {
window.clearTimeout(renderTimer)
renderTimer = window.setTimeout(() => {
renderTimer = undefined
renderQueue = renderQueue.catch(() => undefined).then(initMarkmap)
}, 160)
})
watch(viewportSize, () => {
if (nodeWidth.value !== -1) return
window.clearTimeout(renderTimer)
renderTimer = window.setTimeout(() => {
renderTimer = undefined
renderQueue = renderQueue.catch(() => undefined).then(initMarkmap)
}, 160)
})
onMounted(() => {
renderQueue = renderQueue.then(initMarkmap)
viewportObserver = new ResizeObserver(() => {
const bounds = svgRef.value?.getBoundingClientRect()
if (!bounds || bounds.width <= 0 || bounds.height <= 0) return
viewportSize.value = `${Math.round(bounds.width)}x${Math.round(bounds.height)}`
if (needsVisibleRender) {
needsVisibleRender = false
renderQueue = renderQueue.catch(() => undefined).then(initMarkmap)
@ -504,14 +498,11 @@ onMounted(() => {
}
})
if (svgRef.value) viewportObserver.observe(svgRef.value)
// 窗口缩放时自动调整导图大小
window.addEventListener('resize', handleResize)
})
onUnmounted(() => {
viewportObserver?.disconnect()
window.clearTimeout(renderTimer)
window.removeEventListener('resize', handleResize)
if (mmInstance && typeof mmInstance.destroy === 'function') {
mmInstance.destroy()
} else if (mmInstance) {
@ -519,10 +510,6 @@ onUnmounted(() => {
}
})
const handleResize = () => {
mmInstance?.fit()
}
function downloadBlob(blob: Blob, fileName: string) {
const url = URL.createObjectURL(blob)
const a = document.createElement('a')
@ -609,6 +596,9 @@ function createFullExportSvg() {
const downloadMindMap = async (format: ExportFormat = 'svg') => {
if (isExporting.value) return
isExporting.value = true
exportStage.value = '正在准备导图'
const started = performance.now()
let layoutMs = 0, imagesMs = 0, encodeMs = 0
const fileName = graphFilename(format)
try {
// Export the whole visible tree at its native layout size, independently of pan/zoom.
@ -636,7 +626,12 @@ const downloadMindMap = async (format: ExportFormat = 'svg') => {
if (exportScope.value === 'full') { rememberView(); skipRemember = true; exportingFull = true; await initMarkmap() }
const exportSvg = createFullExportSvg()
if (!exportSvg) return
exportSvg.data = await embedImages(exportSvg.data, true)
layoutMs = Math.round(performance.now() - started)
const imagesStarted = performance.now()
exportSvg.data = await embedImages(exportSvg.data, true, (done, total) => {
exportStage.value = `正在整理图片 ${done}/${total}`
})
imagesMs = Math.round(performance.now() - imagesStarted)
if (format === 'svg') {
downloadBlob(new Blob([exportSvg.data], { type: 'image/svg+xml;charset=utf-8' }), fileName)
@ -659,6 +654,8 @@ const downloadMindMap = async (format: ExportFormat = 'svg') => {
ctx.scale(scaleFactor, scaleFactor)
const img = new Image()
exportStage.value = '正在生成图片'
const encodeStarted = performance.now()
img.src = 'data:image/svg+xml;charset=utf-8,' + encodeURIComponent(exportSvg.data)
await img.decode()
ctx.drawImage(img, 0, 0, exportSvg.width, exportSvg.height)
@ -666,13 +663,16 @@ const downloadMindMap = async (format: ExportFormat = 'svg') => {
const blob = await new Promise<Blob>((resolve, reject) => {
canvas.toBlob(result => result ? resolve(result) : reject(new Error('导出尺寸过大,请选择 SVG')), mimeType, 0.96)
})
encodeMs = Math.round(performance.now() - encodeStarted)
downloadBlob(blob, fileName)
if (scaleFactor < requestedScale * 0.999) {
ElMessage.warning(`图片过大,已自动缩小为 ${canvas.width} × ${canvas.height} 像素;需无损放大请下载 SVG`)
}
} finally {
if (exportingFull) { exportingFull = false; try { await initMarkmap() } finally { skipRemember = false } }
isExporting.value = false
console.info('[mindmap-export]', { format, layoutMs, imagesMs, encodeMs, totalMs: Math.round(performance.now() - started) })
try {
if (exportingFull) { exportingFull = false; try { await initMarkmap() } finally { skipRemember = false } }
} finally { isExporting.value = false; exportStage.value = '' }
}
}

View File

@ -1,7 +1,9 @@
import { ref, reactive, watch } from 'vue'
import { ref, reactive } from 'vue'
import type { Ref } from 'vue'
import { documentApi, type ParseParams } from '@/views/mindmap/workbench/api/document'
import { documentApi, type ParseParams, type SummaryOptions } from '@/views/mindmap/workbench/api/document'
import { buildMindmapMarkdown, documentTitleFromName } from '@/views/mindmap/workbench/utils/mindmapMarkdown'
import { hubCall } from '@/views/mindmap/workbench/api/hub'
import useStore from '@/stores'
// 根据上一级标题自动补全下一级标题
export function autoPromoteParagraphsToSubheading(text: string): string {
@ -16,7 +18,9 @@ export interface DocumentConfig {
mindmapPrompt: string
parserResourceId: string
mindmapModelId: string
headingModelRepair: boolean
mindmapLlmTemperature: number | null
summaryOptions: SummaryOptions
tableEnable: boolean
formulaEnable: boolean
language: string
@ -24,6 +28,10 @@ export interface DocumentConfig {
}
export interface ProcessResult {
original?: string
headingStructure?: any
headingMarkdown?: string
taskId?: string
markdown: string
source: string
mindmap: string
@ -32,7 +40,7 @@ export interface ProcessResult {
export const DOCUMENT_CONFIG_STORAGE_KEY = 'mineru.documentProcessor.config.v5'
export const DEFAULT_MINDMAP_PROMPT = `你是文档结构整理助手。请基于用户提供的 Markdown 生成适合思维导图展示的 Markdown。
export const DEFAULT_MINDMAP_PROMPT = `你是文档总结助手。请将提供的正文提炼为适合思维导图展示的 Markdown。
要求:
1. 保留原文标题结构,不要重写或打乱主要标题层级。
@ -40,11 +48,12 @@ export const DEFAULT_MINDMAP_PROMPT = `你是文档结构整理助手。请基
3. 保留原文语言:英文内容输出英文,中文内容输出中文,多语言内容按原文语言分别保留。
4. 不要编造原文没有的信息。
5. 保留关键数字、公式、专有名词、步骤和结论。
6. 最大层级不超过 4 层。
7. 每个父节点下最多 8 个子节点。
6. 按本次提供的粒度、层级和节点参数整理;保留完整步骤、条件、例外和否定关系。
7. 若当前正文是分块,结合批次范围和章节路径,仅总结本批内容,不推测其他批次内容。
8. 节点标题尽量简短,正文说明使用短句列表。
9. 只输出 Markdown,不要输出解释、代码块围栏或额外说明。`
export const DEFAULT_DOCUMENT_CONFIG: DocumentConfig = {
maxPages: 1000,
backend: 'hybrid-auto-engine',
@ -53,29 +62,34 @@ export const DEFAULT_DOCUMENT_CONFIG: DocumentConfig = {
mindmapPrompt: DEFAULT_MINDMAP_PROMPT,
parserResourceId: '',
mindmapModelId: '',
headingModelRepair: false,
mindmapLlmTemperature: null,
summaryOptions: {
context_tokens: null, max_input_tokens: null, max_output_tokens: null, input_tokens: null,
output_tokens: null, safety_tokens: null, top_p: null,
timeout: 300, retries: 2, concurrency: 1, max_depth: 6, max_children: 8,
node_max_chars: 60, detail: 'standard', structure_mode: 'preserve', parallel: true, auto_split: true,
},
tableEnable: true,
formulaEnable: true,
language: 'ch',
forceOcr: false
}
function loadCachedConfig(): Partial<DocumentConfig> {
try {
const raw = window.localStorage.getItem(DOCUMENT_CONFIG_STORAGE_KEY)
const cached = raw ? JSON.parse(raw) : {}
const safe = Object.fromEntries(Object.keys(DEFAULT_DOCUMENT_CONFIG).filter(key => key in cached).map(key => [key, cached[key]]))
if ('exportResolution' in safe) {
const legacySizes: Record<number, number> = { 1080: 1024, 2560: 2048, 3840: 4096, 7680: 8192, 16384: 8192 }
const size = Number(safe.exportResolution)
safe.exportResolution = legacySizes[size] ?? ([1024, 2048, 4096, 8192].includes(size) ? size : 4096)
}
window.localStorage.setItem(DOCUMENT_CONFIG_STORAGE_KEY, JSON.stringify(safe))
return safe
} catch {
try { window.localStorage.removeItem(DOCUMENT_CONFIG_STORAGE_KEY) } catch { /* Storage unavailable. */ }
return {}
}
type WorkbenchFields = Pick<DocumentConfig, 'parserResourceId' | 'mindmapModelId' | 'mindmapLlmTemperature' | 'mindmapPrompt' | 'summaryOptions' | 'exportResolution'>
type WorkbenchOverride = Partial<Omit<WorkbenchFields, 'summaryOptions'>> & { summaryOptions?: Partial<SummaryOptions> }
// Browser-tab memory only: SPA navigation retains changes, any reload clears them.
const personalOverrides = new Map<string, WorkbenchOverride>()
export function workbenchFields(config: DocumentConfig): WorkbenchFields {
const { parserResourceId, mindmapModelId, mindmapLlmTemperature, mindmapPrompt, exportResolution } = config
return { parserResourceId, mindmapModelId, mindmapLlmTemperature, mindmapPrompt, exportResolution, summaryOptions: { ...config.summaryOptions } }
}
export function mergeWorkbenchDefaults(base: DocumentConfig, value: WorkbenchOverride | null | undefined): DocumentConfig {
return { ...base, ...value, summaryOptions: { ...base.summaryOptions, ...(value?.summaryOptions || {}) } }
}
function identity() {
const user = useStore().user
return user.userInfo?.id ? `${user.userInfo.id}:${user.getWorkspaceId()}` : ''
}
export function useDocumentProcessor() {
@ -85,22 +99,49 @@ export function useDocumentProcessor() {
const isUploading = ref(false)
// 配置相关
const config = reactive<DocumentConfig>({
...DEFAULT_DOCUMENT_CONFIG,
...loadCachedConfig()
})
watch(
config,
(value) => {
try {
window.localStorage.setItem(DOCUMENT_CONFIG_STORAGE_KEY, JSON.stringify(value))
} catch {
// 浏览器禁用本地存储时不影响当前页面使用。
}
},
{ deep: true }
)
const config = reactive<DocumentConfig>(mergeWorkbenchDefaults(DEFAULT_DOCUMENT_CONFIG, {}))
const configReady = ref(false)
let platformDefaults = mergeWorkbenchDefaults(DEFAULT_DOCUMENT_CONFIG, {})
async function loadWorkbenchDefaults() {
// Remove old persistent per-user preferences, which predate the new policy.
try { window.localStorage.removeItem(DOCUMENT_CONFIG_STORAGE_KEY) } catch { /* Storage unavailable. */ }
const [settings, resources] = await Promise.allSettled([hubCall('workbench_defaults'), hubCall('resources')])
platformDefaults = mergeWorkbenchDefaults(DEFAULT_DOCUMENT_CONFIG, settings.status === 'fulfilled' ? settings.value?.config : {})
const available = resources.status === 'fulfilled' ? resources.value : null
if (available) {
const models = (available.models || []).filter((m: any) => m.model_type === 'LLM')
const parsers = available.parsers || []
if (!models.some((m: any) => m.id === platformDefaults.mindmapModelId)) platformDefaults.mindmapModelId = settings.status === 'fulfilled' && settings.value?.model_unavailable ? '' : (models[0]?.id || '')
if (platformDefaults.parserResourceId !== 'builtin' && !parsers.some((p: any) => p.id === platformDefaults.parserResourceId)) platformDefaults.parserResourceId = settings.status === 'fulfilled' && settings.value?.parser_unavailable ? '' : (parsers[0]?.id || '')
}
const key = identity()
const effective = mergeWorkbenchDefaults(platformDefaults, key ? personalOverrides.get(key) : {})
if (available) {
if (!(available.models || []).some((m: any) => m.model_type === 'LLM' && m.id === effective.mindmapModelId)) effective.mindmapModelId = platformDefaults.mindmapModelId
if (effective.parserResourceId !== 'builtin' && !(available.parsers || []).some((p: any) => p.id === effective.parserResourceId)) effective.parserResourceId = platformDefaults.parserResourceId
}
Object.assign(config, effective)
configReady.value = true
return settings.status === 'fulfilled' && resources.status === 'fulfilled'
}
function setPersonalConfig(next: DocumentConfig) {
const key = identity()
const value = workbenchFields(next), baseline = workbenchFields(platformDefaults)
const override: WorkbenchOverride = {}
for (const field of ['parserResourceId', 'mindmapModelId', 'mindmapLlmTemperature', 'mindmapPrompt', 'exportResolution'] as const) {
if (value[field] !== baseline[field]) (override as any)[field] = value[field]
}
const options: Partial<SummaryOptions> = {}
for (const field of Object.keys(value.summaryOptions) as (keyof SummaryOptions)[]) {
if (value.summaryOptions[field] !== baseline.summaryOptions[field]) (options as any)[field] = value.summaryOptions[field]
}
if (Object.keys(options).length) override.summaryOptions = options
if (key) {
if (Object.keys(override).length) personalOverrides.set(key, override)
else personalOverrides.delete(key)
}
Object.assign(config, mergeWorkbenchDefaults(platformDefaults, override))
}
// 结果相关
const results = ref<ProcessResult | null>(null)
@ -178,6 +219,7 @@ export function useDocumentProcessor() {
validFiles.push(file)
displayFiles.push(file)
results.value = {
original: markdown,
markdown,
source: markdown,
mindmap: buildMindmapMarkdown(markdown, documentTitleFromName(file.name))
@ -286,6 +328,7 @@ export function useDocumentProcessor() {
// 处理文档转换
const processDocument = async () => {
error.value = null
if (uploadedFiles.value.length === 0) {
error.value = '请先上传文件'
return
@ -344,9 +387,13 @@ export function useDocumentProcessor() {
const mdContent = resultData.md_content || ''
const mindmapContent = buildMindmapMarkdown(mdContent, documentTitleFromName(uploadedDisplayFiles.value[0]?.name || ''))
results.value = {
original: resultData.original_md || mdContent,
headingStructure: resultData.heading_structure,
headingMarkdown: mdContent,
markdown: mdContent,
source: mdContent,
mindmap: mindmapContent
mindmap: mindmapContent,
taskId: response.task_id
}
progressPercent.value = 100
}
@ -389,6 +436,9 @@ export function useDocumentProcessor() {
uploadedFiles,
uploadedDisplayFiles,
config,
configReady,
loadWorkbenchDefaults,
setPersonalConfig,
results,
isUploading,
isProcessing,

View File

@ -30,7 +30,7 @@ markdownEngine.renderer.rules.workbench_math = (tokens: any, index: number) => {
return `<span class="export-math" data-formula="${markdownEngine.utils.escapeHtml(token.content)}">${katex.renderToString(token.content, { throwOnError: false, displayMode: token.block, trust: false })}</span>`
}
export function graphFromMarkdown(source: string, title: string): GraphNode {
export function graphFromMarkdown(source: string, title: string, excludedLines: number[] = []): GraphNode {
const root: GraphNode = { id: 'root', text: title, content: markdownEngine.utils.escapeHtml(title), children: [], payload: {} }
const lines = source.split('\n'), offsets: number[] = [0]
lines.forEach(line => offsets.push(offsets[offsets.length - 1] + line.length + 1))
@ -38,6 +38,7 @@ export function graphFromMarkdown(source: string, title: string): GraphNode {
const headings: { level: number; node: GraphNode }[] = [{ level: 0, node: root }]
const items: (GraphNode | undefined)[] = []
let depth = 0
const excluded = new Set(excludedLines)
for (let i = 0; i < tokens.length; i++) {
const token = tokens[i]
if (['bullet_list_open', 'ordered_list_open'].includes(token.type)) { depth++; continue }
@ -47,6 +48,7 @@ export function graphFromMarkdown(source: string, title: string): GraphNode {
if (!heading && !['paragraph_open', 'fence', 'code_block', 'table_open', 'html_block'].includes(token.type)) continue
const map = token.map
if (!map) continue
if (excluded.has(map[0]) && lines.slice(map[0], map[1]).every((line, n) => !line.trim() || excluded.has(map[0] + n))) continue
const start = offsets[map[0]], end = Math.min(source.length, offsets[map[1]] - 1)
const raw = source.slice(start, end)
const inline = tokens[i + 1]?.type === 'inline' ? tokens[i + 1].content : token.content

View File

@ -1,5 +1,6 @@
import { onBeforeUnmount, ref, watch, type Ref } from 'vue'
import { hubCall } from '@/views/mindmap/workbench/api/hub'
import { createExportImageCache, runLimited } from './exportPreparation'
const pattern = () => /\/admin\/api\/workspace\/[^/\s"<>]+\/ai_hub\/assets\/([a-f0-9-]{36})\/content/g
const cache = new Map<string, Blob>()
@ -79,16 +80,31 @@ export function blobDataUrl(blob: Blob): Promise<string> {
}
// Standalone HTML and SVG must not contain browser-session Blob URLs.
export async function embedImages(html: string, svg = false) {
const embeddedImages = createExportImageCache()
export async function embedImages(html: string, svg = false, progress?: (done: number, total: number) => void) {
const document = new DOMParser().parseFromString(html, svg ? 'image/svg+xml' : 'text/html')
if (document.querySelector('parsererror')) throw new Error('导出内容格式错误')
for (const image of Array.from(document.querySelectorAll('img'))) {
const src = image.getAttribute('src') || ''
if (src && !src.startsWith('data:')) {
const response = await fetch(src)
if (!response.ok) throw new Error('导出图片加载失败')
image.setAttribute('src', await blobDataUrl(await response.blob()))
const images = Array.from(document.querySelectorAll('img'))
const sources = [...new Set(images.map(image => image.getAttribute('src') || '').filter(src => src && !src.startsWith('data:')))]
const replacements = new Map<string, string>()
let completed = 0
progress?.(0, sources.length)
await runLimited(sources, 4, async src => {
const load = async () => {
const controller = new AbortController(), timeout = setTimeout(() => controller.abort(), 30000)
try {
const response = await fetch(src, { signal: controller.signal })
if (!response.ok) throw new Error('导出图片加载失败')
return await blobDataUrl(await response.blob())
} finally { clearTimeout(timeout) }
}
// Blob URLs identify immutable local images; remote URLs may change contents.
replacements.set(src, await (src.startsWith('blob:') ? embeddedImages.get(src, load) : load()))
progress?.(++completed, sources.length)
})
for (const image of images) {
const replacement = replacements.get(image.getAttribute('src') || '')
if (replacement) image.setAttribute('src', replacement)
}
return svg ? new XMLSerializer().serializeToString(document) : document.body.innerHTML
}

View File

@ -1,10 +1,19 @@
<template>
<div class="document-processor">
<MyMindmapsDialog v-model="myMindmapsVisible" :before-open="confirmOpenSavedArtifact" @open="openSavedArtifact" @renamed="renameSavedArtifact" @deleted="deleteSavedArtifact" />
<el-dialog v-model="summarySourcesVisible" title="总结与原文" width="80%" destroy-on-close>
<el-collapse>
<el-collapse-item v-for="item in smartSources" :key="item.id" :title="`${item.path.join(' / ') || '正文'} · 原文 ${item.start_line}–${item.end_line} 行${item.fallback ? ' · 保留原文' : item.review ? ' · 条件待核对' : ''}`" :name="item.id">
<div class="summary-source-columns"><div><strong>总结</strong><pre>{{ item.summary }}</pre></div><div><strong>原文</strong><pre>{{ item.source }}</pre></div></div>
</el-collapse-item>
</el-collapse>
</el-dialog>
<header class="top-header">
<h2 class="ellipsis" title="思维导图">思维导图</h2>
<h2 class="ellipsis" title="思维导图工作台">思维导图工作台</h2>
<div class="header-actions">
<el-button :disabled="!results" @click="saveArtifact">保存图谱</el-button>
<el-button type="text" @click="toggleSettings" class="settings-button">
<el-button :disabled="isProcessing || isSmartOrganizing || savingArtifact" @click="myMindmapsVisible = true">我的导图</el-button>
<el-button :loading="savingArtifact" :disabled="!results || isProcessing || isSmartOrganizing" @click="saveArtifact">保存导图</el-button>
<el-button type="text" :disabled="!configReady" @click="toggleSettings" class="settings-button">
设置
</el-button>
</div>
@ -40,7 +49,7 @@
/>
<el-button
type="primary"
:loading="isProcessing || isUploading"
:loading="isProcessing || isUploading || isSmartOrganizing" :disabled="!configReady"
@click.stop="handleProcessDocument"
>
开始转换
@ -149,22 +158,26 @@
v-model="sourceContentMode"
:options="sourceContentModeOptions"
size="small"
:disabled="sourceViewMode === 'ast' || isSmartOrganizing"
:disabled="sourceViewMode === 'ast' || isSmartOrganizing || !configReady"
/>
<span v-if="isSmartOrganizing" class="smart-organize-status">
{{ smartOrganizeStage || '智能整理中' }} {{ smartOrganizeProgress === null ? '' : `${smartOrganizeProgress}%` }}
{{ smartOrganizeStage || '智能总结中' }} {{ smartOrganizeProgress === null ? '' : `${smartOrganizeProgress}%` }}
</span>
<span v-else-if="smartOrganizeError" class="template-error-inline">
{{ smartOrganizeError }}
</span>
<span v-else-if="sourceContentMode === 'smart' && smartMarkdownContent" class="smart-organize-status">
智能整理结果
{{ smartResultStale ? '原文或配置已变更,请重新总结' : smartPlan?.fallback_sections || smartPlan?.review_sections ? '部分内容待核对' : '智能总结结果' }}
</span>
<span v-else-if="templateRenderError" class="template-error-inline">
{{ templateRenderError }}
</span>
</div>
<div class="content-toolbar-actions">
<el-button v-if="canShowSmartSources" @click="summarySourcesVisible = true">原文对照</el-button>
<el-button v-if="smartMarkdownContent || smartOrganizeError" :disabled="isSmartOrganizing" @click="startSmartOrganize(true)">重新总结</el-button>
<el-button v-if="smartTaskFailed && !smartResultStale" :disabled="isSmartOrganizing" @click="retrySmartOrganize">继续处理</el-button>
<el-button :disabled="headingBusyDisabled" @click="headingPanelVisible = true">标题层级</el-button>
<el-button
:type="hasActiveReplacementRules ? 'primary' : 'default'"
class="replacement-config-button"
@ -174,7 +187,6 @@
</el-button>
</div>
</div>
<div class="source-panel">
<el-input
v-model="sourcePanelContent"
@ -195,9 +207,11 @@
:source-name="exportSourceName"
v-model:layout="graphLayout"
:editable="!hasActiveReplacementRules"
:excluded-lines="activeHeadingStructure?.excluded_lines || []"
:document-key="mindmapDocumentKey"
:export-resolution="config.exportResolution"
@node-edit="handleMindmapNodeEdit"
@heading-edit="headingPanelVisible = true"
/>
</div>
</div>
@ -249,19 +263,28 @@
</div>
</template>
</el-dialog>
<HeadingStructurePanel v-model="headingPanelVisible"
:markdown="activeSourceMarkdownContent" :original="originalRecognized || results?.original || manualMarkdownContent"
:structure="activeHeadingStructure" :task-id="results?.taskId" :artifact-id="openedArtifact?.id"
:document-key="mindmapDocumentKey"
@inspect="inspectHeadingStructure" @apply="applyHeadingStructure" />
</div>
</template>
<script setup lang="ts">
import { computed, onMounted, onUnmounted, ref, watch } from 'vue'
import { useRouter } from 'vue-router'
import useStore from '@/stores'
import { Delete, Document, Folder, Close, Loading, ArrowDown } from '@element-plus/icons-vue'
import { ElMessage } from 'element-plus'
import { ElMessage, ElMessageBox } from 'element-plus'
import ConfigPanel from '@/views/mindmap/workbench/components/ConfigPanel.vue'
import MyMindmapsDialog from '@/views/mindmap/workbench/components/MyMindmapsDialog.vue'
import MarkdownRenderer from '@/views/mindmap/workbench/components/MarkdownRenderer.vue'
import MindMapRenderer from '@/views/mindmap/workbench/components/MindMapRenderer.vue'
import HeadingStructurePanel from '@/views/mindmap/workbench/components/HeadingStructurePanel.vue'
import { hubCall, onOpenArtifact } from '@/views/mindmap/workbench/api/hub'
import { documentApi } from '@/views/mindmap/workbench/api/document'
import { DOCUMENT_CONFIG_STORAGE_KEY, useDocumentProcessor } from '@/views/mindmap/workbench/composables/useDocumentProcessor'
import { DEFAULT_DOCUMENT_CONFIG, useDocumentProcessor } from '@/views/mindmap/workbench/composables/useDocumentProcessor'
import { markdownToAstString } from '@/views/mindmap/workbench/utils/markdownAst'
import { buildMindmapMarkdown, documentTitleFromName } from '@/views/mindmap/workbench/utils/mindmapMarkdown'
import { createEmptyReplacementRule, renderMarkdownTemplate, type ReplacementRule } from '@/views/mindmap/workbench/utils/markdownTemplate'
@ -300,6 +323,9 @@ const {
uploadedFiles,
uploadedDisplayFiles,
config,
configReady,
loadWorkbenchDefaults,
setPersonalConfig,
results,
isUploading,
isProcessing,
@ -316,6 +342,7 @@ const {
} = useDocumentProcessor()
const showSettings = ref(false)
const myMindmapsVisible = ref(false), savingArtifact = ref(false)
const fileInput = ref<HTMLInputElement | null>(null)
const isDragging = ref(false)
const activeTab = ref('markdown')
@ -330,30 +357,76 @@ const sourceViewMode = ref<'markdown' | 'ast'>(
const sourceContentMode = ref<'source' | 'smart'>('source')
const sourceContentModeOptions = [
{ label: '源码', value: 'source' },
{ label: '智能整理', value: 'smart' }
{ label: '智能总结', value: 'smart' }
]
const showReplacementDialog = ref(false)
const isMobileDialog = ref(false)
const replacementRules = ref<ReplacementRule[]>([])
const draftReplacementRules = ref<ReplacementRule[]>([])
const smartMarkdownContent = ref('')
const smartHeadingStructure = ref<any>()
const smartHeadingMarkdown = ref('')
const headingPanelVisible = ref(false)
const smartOrganizeTaskId = ref('')
const isSmartOrganizing = ref(false)
const smartOrganizeProgress = ref<number | null>(null)
const smartOrganizeStage = ref('')
const smartOrganizeError = ref('')
const smartPlan = ref<Record<string, any> | null>(null)
const smartSources = ref<any[]>([])
const smartSourceSnapshot = ref('')
const summarySourcesVisible = ref(false)
const smartRequestSnapshot = ref('')
const smartTaskFailed = ref(false)
let smartPollingGeneration = 0
let applyingSmartResult = false
let smartOrganizeTimer: ReturnType<typeof setInterval> | null = null
const summaryRecoveryKey = () => `mindmap.summary.task.v1:${useStore().user.userInfo?.id || ''}:${useStore().user.getWorkspaceId()}`
function rememberSummaryTask(taskId: string) {
try { sessionStorage.setItem(summaryRecoveryKey(), taskId) } catch { /* Optional recovery storage. */ }
}
function forgetSummaryTask() {
try { sessionStorage.removeItem(summaryRecoveryKey()) } catch { /* Optional recovery storage. */ }
}
function showSmartResult() {
if (sourceContentMode.value !== 'smart') {
applyingSmartResult = true
sourceContentMode.value = 'smart'
}
}
watch(markdownRenderMode, (value) => storeValue(MARKDOWN_RENDER_MODE_KEY, value))
watch(sourceViewMode, (value) => storeValue(SOURCE_VIEW_MODE_KEY, value))
const manualMarkdownContent = computed(() => results.value?.source || '')
const currentSmartRequest = computed(() => JSON.stringify([manualMarkdownContent.value,
config.mindmapModelId, config.mindmapPrompt, config.mindmapLlmTemperature,
config.summaryOptions, results.value?.headingStructure]))
const smartResultStale = computed(() => !!smartRequestSnapshot.value && smartRequestSnapshot.value !== currentSmartRequest.value)
const activeHeadingStructure = computed(() => {
const state = sourceContentMode.value === 'smart' ? smartHeadingStructure.value : results.value?.headingStructure
const snapshot = sourceContentMode.value === 'smart' ? smartHeadingMarkdown.value : results.value?.headingMarkdown
return state && snapshot !== activeSourceMarkdownContent.value ? { ...state, excluded_lines: [] } : state
})
const headingBusyDisabled = computed(() => !activeSourceMarkdownContent.value.trim() || isProcessing.value || isSmartOrganizing.value || hasActiveReplacementRules.value)
function inspectHeadingStructure(value: any) {
if (sourceContentMode.value === 'smart') { smartHeadingStructure.value = value.heading_structure; smartHeadingMarkdown.value = activeSourceMarkdownContent.value }
else if (results.value) { results.value.headingStructure = value.heading_structure; results.value.headingMarkdown = activeSourceMarkdownContent.value }
}
function applyHeadingStructure(value: any) {
if (!results.value) return
if (sourceContentMode.value === 'smart') smartMarkdownContent.value = value.markdown
else { results.value.source = value.markdown; results.value.markdown = value.markdown }
if (value.task_id) results.value.taskId = value.task_id
inspectHeadingStructure(value)
}
const activeSourceMarkdownContent = computed(() => {
if (sourceContentMode.value === 'smart' && smartMarkdownContent.value) {
return smartMarkdownContent.value
}
return manualMarkdownContent.value
})
const canShowSmartSources = computed(() => smartSources.value.length > 0 && !!smartSourceSnapshot.value && smartSourceSnapshot.value === activeSourceMarkdownContent.value)
const templateRenderResult = computed(() => renderMarkdownTemplate(activeSourceMarkdownContent.value, replacementRules.value))
const renderedMarkdownContent = computed(() => templateRenderResult.value.markdown)
const openedArtifact = ref<any>(null)
@ -403,6 +476,7 @@ const sourcePanelContent = computed({
})
const stopSmartOrganizePolling = () => {
smartPollingGeneration += 1
if (smartOrganizeTimer) {
clearInterval(smartOrganizeTimer)
smartOrganizeTimer = null
@ -413,78 +487,170 @@ const resetSmartOrganizeState = () => {
stopSmartOrganizePolling()
sourceContentMode.value = 'source'
smartMarkdownContent.value = ''
smartHeadingStructure.value = undefined
smartHeadingMarkdown.value = ''
headingPanelVisible.value = false
smartOrganizeTaskId.value = ''
isSmartOrganizing.value = false
smartOrganizeProgress.value = 0
smartOrganizeStage.value = ''
smartOrganizeError.value = ''
smartPlan.value = null
smartSources.value = []
smartSourceSnapshot.value = ''
forgetSummaryTask()
smartRequestSnapshot.value = ''
smartTaskFailed.value = false
}
const startSmartOrganizePolling = (taskId: string) => {
stopSmartOrganizePolling()
const generation = smartPollingGeneration
let polling = false
let failures = 0
smartOrganizeTimer = setInterval(async () => {
if (polling) return
polling = true
try {
const data = await documentApi.getMindmapProgress(taskId)
if (generation !== smartPollingGeneration) return
failures = 0
if (data.plan) smartPlan.value = data.plan
smartOrganizeProgress.value = data.progress
smartOrganizeStage.value = data.stage
if (data.status === 'completed') {
stopSmartOrganizePolling()
isSmartOrganizing.value = false
smartMarkdownContent.value = data.result_md || ''
smartHeadingStructure.value = data.heading_structure
smartHeadingMarkdown.value = smartMarkdownContent.value
smartSources.value = data.sources || []
smartSourceSnapshot.value = smartMarkdownContent.value
if (results.value && data.task_id) results.value.taskId = data.task_id
smartOrganizeError.value = ''
sourceContentMode.value = 'smart'
ElMessage.success('智能整理完成')
showSmartResult()
if (data.warnings?.length) ElMessage.warning('总结完成,部分章节待核对')
else ElMessage.success('智能总结完成')
} else if (data.status === 'failed') {
stopSmartOrganizePolling()
isSmartOrganizing.value = false
sourceContentMode.value = 'source'
smartOrganizeError.value = data.error || '智能整理失败'
smartOrganizeError.value = data.error || '智能总结失败'
smartTaskFailed.value = true
ElMessage.error(smartOrganizeError.value)
}
} catch (err) {
stopSmartOrganizePolling()
isSmartOrganizing.value = false
sourceContentMode.value = 'source'
smartOrganizeError.value = (err as Error).message || '智能整理进度查询失败'
ElMessage.error(smartOrganizeError.value)
if (generation !== smartPollingGeneration) return
failures += 1
smartOrganizeStage.value = failures >= 3 ? '进度查询暂不可用,正在重试' : '正在查询进度'
} finally {
polling = false
}
}, 1500)
}
const startSmartOrganize = async () => {
const startSmartOrganize = async (force = false) => {
if (isSmartOrganizing.value) return
const markdown = manualMarkdownContent.value.trim()
if (!markdown) {
sourceContentMode.value = 'source'
ElMessage.warning('暂无 Markdown 源码可整理')
return
}
if (smartMarkdownContent.value) {
if (!force && smartMarkdownContent.value && !smartResultStale.value) {
sourceContentMode.value = 'smart'
return
}
const taskId = `mindmap-${Date.now()}-${Math.random().toString(36).slice(2, 8)}`
smartOrganizeTaskId.value = taskId
smartRequestSnapshot.value = currentSmartRequest.value
smartPlan.value = null
smartTaskFailed.value = false
isSmartOrganizing.value = true
smartOrganizeProgress.value = 0
smartOrganizeStage.value = '创建智能整理任务'
smartOrganizeStage.value = '创建智能总结任务'
smartOrganizeError.value = ''
try {
await documentApi.createMindmapTask(taskId, markdown, 'hybrid', config.mindmapPrompt, {
const options = Object.fromEntries(Object.entries(config.summaryOptions).filter(([, value]) => value != null))
const node: Record<string, unknown> = {type: 'organize', model_id: config.mindmapModelId, prompt: config.mindmapPrompt, ...options}
if (config.mindmapLlmTemperature != null) node.temperature = config.mindmapLlmTemperature
smartPlan.value = await documentApi.previewSummary(markdown, node)
const created = await documentApi.createMindmapTask(taskId, markdown, 'hybrid', config.mindmapPrompt, {
model_id: config.mindmapModelId,
temperature: config.mindmapLlmTemperature
})
startSmartOrganizePolling(taskId)
temperature: config.mindmapLlmTemperature,
options: config.summaryOptions,
}, { heading_structure: results.value?.headingStructure, original: originalRecognized.value || results.value?.original || markdown })
smartOrganizeTaskId.value = created.task_id || taskId
if (created.task_id) rememberSummaryTask(created.task_id)
startSmartOrganizePolling(smartOrganizeTaskId.value)
} catch (err) {
isSmartOrganizing.value = false
sourceContentMode.value = 'source'
smartOrganizeError.value = (err as Error).message || '智能整理任务创建失败'
smartOrganizeError.value = (err as Error).message || '智能总结任务创建失败'
ElMessage.error(smartOrganizeError.value)
}
}
async function restoreSummaryTask() {
if (router.currentRoute.value.query.artifact) return
let taskId = ''
try { taskId = sessionStorage.getItem(summaryRecoveryKey()) || '' } catch { return }
if (!/^[a-f0-9-]{36}$/i.test(taskId) || results.value || isProcessing.value) return
try {
const data = await documentApi.getMindmapProgress(taskId)
if (results.value || isProcessing.value || router.currentRoute.value.query.artifact) return
const source = data.summary_input?.markdown
const node = data.summary_node
if (!source || !node) { forgetSummaryTask(); return }
const taskOptions = { ...DEFAULT_DOCUMENT_CONFIG.summaryOptions }
for (const key of Object.keys(taskOptions)) {
if (node[key] != null) (taskOptions as any)[key] = node[key]
else if (['context_tokens', 'max_input_tokens', 'max_output_tokens', 'input_tokens', 'output_tokens', 'safety_tokens', 'top_p'].includes(key)) (taskOptions as any)[key] = null
}
originalRecognized.value = data.original_md || source
results.value = { markdown: source, source, original: data.original_md || source,
headingStructure: data.summary_input?.heading_structure, headingMarkdown: source,
mindmap: buildMindmapMarkdown(source), taskId }
smartOrganizeTaskId.value = taskId
smartRequestSnapshot.value = JSON.stringify([source, node.model_id, node.prompt, node.temperature ?? null,
taskOptions, data.summary_input?.heading_structure])
smartPlan.value = data.plan || null
smartSources.value = data.sources || []
activeTab.value = 'source'
if (data.status === 'completed') {
smartMarkdownContent.value = data.result_md || ''
smartSourceSnapshot.value = smartMarkdownContent.value
smartHeadingStructure.value = data.heading_structure
smartHeadingMarkdown.value = smartMarkdownContent.value
showSmartResult()
} else if (data.status === 'failed') {
smartTaskFailed.value = true
smartOrganizeError.value = data.error || '总结失败,可继续处理'
} else {
isSmartOrganizing.value = true
startSmartOrganizePolling(taskId)
}
} catch { forgetSummaryTask() }
}
const retrySmartOrganize = async () => {
if (isSmartOrganizing.value || smartResultStale.value) return
isSmartOrganizing.value = true
smartOrganizeError.value = ''
try {
await documentApi.retryMindmapTask(smartOrganizeTaskId.value)
smartTaskFailed.value = false
startSmartOrganizePolling(smartOrganizeTaskId.value)
} catch (err) {
isSmartOrganizing.value = false
smartOrganizeError.value = (err as Error).message
}
}
watch(sourceContentMode, (value) => {
if (applyingSmartResult) { applyingSmartResult = false; return }
if (value === 'smart') {
startSmartOrganize()
}
@ -545,15 +711,12 @@ const handleBackendChange = (backend: string) => {
}
const applyConfig = (nextConfig: typeof config) => {
Object.assign(config, nextConfig)
try {
window.localStorage.setItem(DOCUMENT_CONFIG_STORAGE_KEY, JSON.stringify(nextConfig))
} catch {
// 浏览器禁用本地存储时不影响当前页面使用。
}
setPersonalConfig(nextConfig)
}
const router = useRouter()
const handleProcessDocument = async () => {
if (isProcessing.value || isSmartOrganizing.value) return
mindmapDocumentKey.value = `parse-${Date.now()}`
openedArtifact.value = null
originalRecognized.value = ''
@ -564,7 +727,7 @@ const handleProcessDocument = async () => {
return
}
if (results.value) {
activeTab.value = 'markdown'
activeTab.value = 'mindmap'
sourceViewMode.value = 'markdown'
ElMessage.success('转换完成')
}
@ -629,6 +792,7 @@ const handleMindmapNodeEdit = ({ start, end, original, newText }: { start: numbe
}
const triggerUpload = () => {
if (isProcessing.value || isSmartOrganizing.value) return
fileInput.value?.click()
}
@ -656,6 +820,7 @@ const handleFileInputChange = async (event: Event) => {
const handleDrop = async (event: DragEvent) => {
event.preventDefault()
isDragging.value = false
if (isProcessing.value || isSmartOrganizing.value) return
if (event.dataTransfer?.files) {
mindmapDocumentKey.value = `upload-${Date.now()}`
@ -676,6 +841,7 @@ const handleDrop = async (event: DragEvent) => {
}
const removeFile = (index: number) => {
if (isProcessing.value || isSmartOrganizing.value) return
uploadedFiles.value.splice(index, 1)
uploadedDisplayFiles.value.splice(index, 1)
if (uploadedFiles.value.length === 0) {
@ -690,41 +856,100 @@ const clearAllFiles = () => {
}
const originalRecognized = ref('')
const stopArtifactListener = onOpenArtifact((artifact) => {
let artifactLoadGeneration = 0
function applySavedArtifact(artifact: any) {
const generation = ++artifactLoadGeneration
resetSmartOrganizeState()
replacementRules.value = []
draftReplacementRules.value = []
showReplacementDialog.value = false
openedArtifact.value = artifact
graphLayout.value = artifact.layout || 'tree'
mindmapDocumentKey.value = `artifact-${artifact.id}`
uploadedFiles.value = []
uploadedDisplayFiles.value = []
originalRecognized.value = artifact.original || ''
results.value = { markdown: artifact.markdown, source: artifact.markdown, mindmap: buildMindmapMarkdown(artifact.markdown) }
const savedSummary = artifact.summary_state || {}
const source = savedSummary.source_markdown || artifact.markdown
smartMarkdownContent.value = savedSummary.smart_markdown || ''
const restoreSmart = savedSummary.mode === 'smart' && !!smartMarkdownContent.value
if (restoreSmart) {
sourceContentMode.value = 'smart'
smartHeadingStructure.value = artifact.heading_structure
smartHeadingMarkdown.value = smartMarkdownContent.value
}
results.value = { markdown: source, source, original: artifact.original, headingStructure: sourceContentMode.value === 'source' ? artifact.heading_structure : undefined, headingMarkdown: source, mindmap: buildMindmapMarkdown(artifact.markdown), taskId: artifact.task_id || undefined }
if (artifact.task_id) documentApi.getMindmapProgress(artifact.task_id).then(data => {
if (generation === artifactLoadGeneration && openedArtifact.value?.id === artifact.id && data.status === 'completed' && data.result_md === artifact.markdown) {
smartSources.value = data.sources || []
smartSourceSnapshot.value = data.result_md || ''
smartPlan.value = data.plan || null
}
}).catch(() => { /* The original task may have been deleted. */ })
activeTab.value = 'mindmap'
}
const stopArtifactListener = onOpenArtifact((artifact) => {
if (openedArtifact.value?.id !== artifact.id || openedArtifact.value?.revision !== artifact.revision) applySavedArtifact(artifact)
})
async function openSavedArtifact(artifact: any) {
applySavedArtifact(artifact)
await router.replace({path: '/mindmap', query: {...router.currentRoute.value.query, artifact: artifact.id}})
}
async function confirmOpenSavedArtifact() {
const current = openedArtifact.value
const changed = !!activeSourceMarkdownContent.value.trim() && (!current || current.markdown !== activeSourceMarkdownContent.value || current.layout !== graphLayout.value)
if (!changed) return true
return ElMessageBox.confirm('当前内容尚未保存,是否打开这张导图?', '打开导图', {type: 'warning', confirmButtonText: '打开', cancelButtonText: '取消'}).then(() => true, () => false)
}
function renameSavedArtifact(artifact: any) {
if (openedArtifact.value?.id === artifact.id) openedArtifact.value = {...openedArtifact.value, name: artifact.name, updated_at: artifact.updated_at}
}
async function deleteSavedArtifact(id: string) {
if (openedArtifact.value?.id !== id) return
openedArtifact.value = null
const query = {...router.currentRoute.value.query}
delete query.artifact
await router.replace({query})
}
watch(() => results.value?.markdown, (value) => {
if (!openedArtifact.value && value) originalRecognized.value = value
if (!openedArtifact.value && value) originalRecognized.value = results.value?.original || originalRecognized.value || value
})
async function saveArtifact() {
if (savingArtifact.value) return
savingArtifact.value = true
try {
const current = openedArtifact.value
const saved = await hubCall('artifacts', 'post', {
name: current?.name || uploadedDisplayFiles.value[0]?.name || '思维导图',
revision: current?.revision, markdown: activeSourceMarkdownContent.value,
task_id: results.value?.taskId,
original: originalRecognized.value || manualMarkdownContent.value,
summary_state: { source_markdown: manualMarkdownContent.value, smart_markdown: smartMarkdownContent.value, mode: sourceContentMode.value },
heading_structure: activeHeadingStructure.value,
layout: graphLayout.value
}, current?.id)
openedArtifact.value = saved
originalRecognized.value = saved.original || ''
results.value = { markdown: saved.markdown, source: saved.markdown, mindmap: buildMindmapMarkdown(saved.markdown) }
const source = saved.summary_state?.source_markdown || saved.markdown
results.value = { markdown: source, source, original: saved.original, headingStructure: sourceContentMode.value === 'source' ? saved.heading_structure : undefined, headingMarkdown: source, mindmap: buildMindmapMarkdown(saved.markdown), taskId: saved.task_id || undefined }
await router.replace({path: '/mindmap', query: {...router.currentRoute.value.query, artifact: saved.id}})
ElMessage.success('已保存')
} catch (e: any) { ElMessage.error(e.message || '保存失败') }
finally { savingArtifact.value = false }
}
onMounted(() => {
let componentActive = true
onMounted(async () => {
updateViewportState()
window.addEventListener('resize', updateViewportState)
const loaded = await loadWorkbenchDefaults()
if (!componentActive) return
if (!loaded) ElMessage.warning('工作台默认设置加载失败,请检查网络')
await restoreSummaryTask()
})
onUnmounted(() => {
componentActive = false
stopArtifactListener()
stopSmartOrganizePolling()
window.removeEventListener('resize', updateViewportState)
@ -732,6 +957,9 @@ onUnmounted(() => {
</script>
<style scoped>
.summary-source-columns { display: grid; grid-template-columns: 1fr 1fr; gap: 20px; }
.summary-source-columns > div { min-width: 0; }
.summary-source-columns pre { white-space: pre-wrap; overflow-wrap: anywhere; max-height: 55vh; overflow: auto; font: inherit; }
.document-processor {
position: relative;
height: 100%;
@ -788,6 +1016,7 @@ onUnmounted(() => {
flex-shrink: 0;
}
.drag-upload-area {
border: 2px dashed #CED4DA;
border-radius: 8px;

View File

@ -49,7 +49,7 @@ async function importFiles() {
}
try {
selected.value = (await post('/local_model/imports', body, undefined, loading)).data
MsgSuccess('已导入'); path.value = ''; files.value = []; await load()
MsgSuccess(selected.value?.already_imported ? '模型已存在' : '已导入'); path.value = ''; files.value = []; await load()
} catch { /* request interceptor displays the server error; retain the form */ }
}
onMounted(load)

View File

@ -142,6 +142,7 @@
<script setup lang="ts">
import type { Provider, Model } from '@/api/type/model'
import { computed, ref, onMounted, onBeforeUnmount } from 'vue'
import { Download } from '@element-plus/icons-vue'
import EditModel from '@/views/model/component/EditModel.vue'
import ModelLoadingDialog from './ModelLoadingDialog.vue'
import useStore from '@/stores'

View File

@ -10,7 +10,7 @@
<el-card shadow="never" class="policy-page__card">
<el-tabs v-model="activeTab">
<el-tab-pane :label="t('views.system.policy.configTab')" name="policy">
<div class="policy-page__toolbar mb-16">
<div class="policy-page__toolbar policy-page__toolbar--actions mb-16">
<div class="policy-page__filters">
<el-input
v-model="policyQuery.name"
@ -32,7 +32,7 @@
</el-button>
</div>
<el-table :data="policyList" border v-loading="policyLoading">
<el-table :data="pagedPolicies" border v-loading="policyLoading">
<el-table-column
prop="name"
:label="t('views.system.policy.policyName')"
@ -81,6 +81,7 @@
</template>
</el-table-column>
</el-table>
<SystemPagination v-model:current-page="policyPage" v-model:page-size="policyPageSize" :total="policyList.length" />
</el-tab-pane>
<el-tab-pane :label="t('views.system.policy.recordTab')" name="record">
@ -105,7 +106,7 @@
</div>
</div>
<el-table :data="recordList" border v-loading="recordLoading">
<el-table :data="pagedRecords" border v-loading="recordLoading">
<el-table-column
prop="policy_name"
:label="t('views.system.policy.policyName')"
@ -147,10 +148,11 @@
show-overflow-tooltip
/>
</el-table>
<SystemPagination v-model:current-page="recordPage" v-model:page-size="recordPageSize" :total="recordList.length" />
</el-tab-pane>
<el-tab-pane label="当前绑定" name="bindings">
<el-table :data="runtime.bindings" border>
<el-table :data="pagedBindings" border>
<el-table-column prop="policy_name" label="策略名称" min-width="180" />
<el-table-column prop="target_name" label="应用对象" min-width="220" />
<el-table-column prop="target_identifier" label="对象标识" min-width="250" />
@ -161,13 +163,16 @@
</template></el-table-column
>
</el-table>
<SystemPagination v-model:current-page="bindingPage" v-model:page-size="bindingPageSize" :total="runtime.bindings.length" />
<p class="mt-12 color-secondary">
多个策略同时绑定时,须全部满足。停用或生效期外的策略不限制访问。
</p>
</el-tab-pane>
<el-tab-pane label="访问记录" name="decisions">
<div class="policy-page__filters decision-filters mb-16">
<div class="policy-page__toolbar policy-page__toolbar--actions mb-16">
<el-button :disabled="decisionLoading" @click="loadDecisions">刷新</el-button>
<div class="policy-page__filters decision-filters">
<el-date-picker v-model="decisionQuery.time" type="datetimerange"
value-format="YYYY-MM-DDTHH:mm:ss" start-placeholder="开始时间" end-placeholder="结束时间"
range-separator="至" aria-label="访问时间(北京时间)" :default-time="decisionDefaultTime" />
@ -181,7 +186,7 @@
<el-input v-model="decisionQuery.reason" placeholder="拒绝原因" aria-label="拒绝原因" clearable maxlength="200" @keyup.enter="searchDecisions" />
<el-button type="primary" :loading="decisionLoading" @click="searchDecisions">查询</el-button>
<el-button @click="resetDecisionFilters">重置</el-button>
<el-button :disabled="decisionLoading" @click="loadDecisions">刷新</el-button>
</div>
</div>
<el-table v-loading="decisionLoading" :data="decisions" border>
<el-table-column label="时间" min-width="180"
@ -201,9 +206,8 @@
<el-table-column prop="target" label="服务标识" min-width="250" />
<el-table-column prop="reason" label="拒绝原因" min-width="220" />
</el-table>
<el-pagination class="mt-16" v-model:current-page="decisionPage.current"
v-model:page-size="decisionPage.page_size" :total="decisionPage.total" :page-sizes="[10, 20, 50, 100]"
layout="total, sizes, prev, pager, next, jumper" @change="changeDecisionPage" />
<SystemPagination v-model:current-page="decisionPage.current"
v-model:page-size="decisionPage.page_size" :total="decisionPage.total" @change="changeDecisionPage" />
</el-tab-pane>
<el-tab-pane label="区域网段" name="regions">
@ -461,6 +465,7 @@
</template>
<script setup lang="ts">
import SystemPagination from '@/components/system-pagination/index.vue'
import { computed, onMounted, reactive, ref, watch } from 'vue'
import type { FormInstance, FormRules } from 'element-plus'
import accessControlPolicyApi from '@/api/system-settings/access-control-policy'
@ -484,6 +489,14 @@ const applyFormRef = ref<FormInstance>()
const policyList = ref<any[]>([])
const recordList = ref<any[]>([])
const runtime = reactive<any>({ bindings: [], targets: [], region_networks: [] })
const policyPage = ref(1), policyPageSize = ref(10), recordPage = ref(1), recordPageSize = ref(10), bindingPage = ref(1), bindingPageSize = ref(10)
const pageRows = (rows: any[], page: number, size: number) => rows.slice((page - 1) * size, page * size)
const pagedPolicies = computed(() => pageRows(policyList.value, policyPage.value, policyPageSize.value))
const pagedRecords = computed(() => pageRows(recordList.value, recordPage.value, recordPageSize.value))
const pagedBindings = computed(() => pageRows(runtime.bindings, bindingPage.value, bindingPageSize.value))
watch([() => policyList.value.length, policyPageSize], ([total, size]) => { policyPage.value = Math.min(policyPage.value, Math.max(1, Math.ceil(total / size))) })
watch([() => recordList.value.length, recordPageSize], ([total, size]) => { recordPage.value = Math.min(recordPage.value, Math.max(1, Math.ceil(total / size))) })
watch([() => runtime.bindings.length, bindingPageSize], ([total, size]) => { bindingPage.value = Math.min(bindingPage.value, Math.max(1, Math.ceil(total / size))) })
const decisions = ref<any[]>([])
const decisionLoading = ref(false)
const decisionQuery = reactive({ time: [] as string[] | null, result: '', ip: '', device_type: '', reason: '' })
@ -616,11 +629,13 @@ async function loadPolicies() {
policyLoading,
)
policyList.value = response.data || []
policyPage.value = 1
}
async function loadRecords() {
const response = await accessControlPolicyApi.getApplicationRecordList(recordQuery, recordLoading)
recordList.value = response.data || []
recordPage.value = 1
}
function resetPolicyFilters() {
@ -827,10 +842,11 @@ onMounted(async () => {
&__toolbar {
display: flex;
align-items: center;
justify-content: space-between;
justify-content: flex-end;
gap: 12px;
flex-wrap: wrap;
}
&__toolbar--actions { justify-content: space-between; .policy-page__filters { order: 2; } }
&__filters {
display: flex;

View File

@ -1,17 +1,42 @@
<template>
<div class="mindmap-management p-16-24">
<h2 class="mb-16">思维导图管理</h2>
<el-card :body-style="{ padding: '0' }">
<HubManager @open-artifact="openArtifact" />
<el-breadcrumb separator-icon="ArrowRight" class="mb-16"><el-breadcrumb-item>服务管理</el-breadcrumb-item><el-breadcrumb-item><h5 class="ml-4 color-text-primary">导图设置</h5></el-breadcrumb-item></el-breadcrumb>
<el-card>
<el-tabs v-model="tab" @tab-change="changeTab">
<el-tab-pane label="处理流程" name="flows" />
<el-tab-pane label="导图服务" name="services" />
<el-tab-pane label="运行记录" name="tasks" />
<el-tab-pane label="导图管理" name="artifacts" />
<el-tab-pane v-if="isAdmin" label="工作台设置" name="workbench" />
</el-tabs>
<WorkbenchSettings v-if="tab === 'workbench' && isAdmin" />
<HubManager v-else :key="tab" :initial-tab="hubTab" hide-tabs @open-artifact="openArtifact" @tab-change="changeTab" />
</el-card>
</div>
</template>
<script setup lang="ts">
import { useRouter } from 'vue-router'
import { computed, ref, watch, onMounted } from 'vue'
import { useRouter, useRoute } from 'vue-router'
import HubManager from '@/views/mindmap/HubManager.vue'
import WorkbenchSettings from './WorkbenchSettings.vue'
import useStore from '@/stores'
const router = useRouter()
const route = useRoute()
const isAdmin = computed(() => useStore().user.is_admin())
const tabs = ['flows', 'services', 'tasks', 'artifacts', 'workbench']
function selectedTab(value: unknown) { return typeof value === 'string' && tabs.includes(value) && (value !== 'workbench' || isAdmin.value) ? value : 'services' }
const tab = ref(selectedTab(route.query.tab))
const hubTab = computed(() => tab.value as 'services' | 'flows' | 'tasks' | 'artifacts')
function syncTab(value: unknown) {
if (value === 'advanced') { router.replace({ path: '/system/advanced-flows' }); return }
tab.value = selectedTab(value)
}
watch(() => route.query.tab, syncTab)
watch(isAdmin, () => syncTab(route.query.tab))
onMounted(() => syncTab(route.query.tab))
function changeTab(value: string | number) { router.replace({ query: { ...route.query, tab: String(value) } }) }
function openArtifact(artifact: { id: string }) {
router.push({ path: '/mindmap', query: { artifact: artifact.id } })
}

View File

@ -1,12 +1,13 @@
<template>
<div class="p-16-24 access-page" v-loading="loading">
<h2 class="mb-16">角色与授权</h2>
<el-breadcrumb separator-icon="ArrowRight" class="mb-16"><el-breadcrumb-item><h5 class="color-text-primary">角色与授权</h5></el-breadcrumb-item></el-breadcrumb>
<el-card shadow="never">
<el-tabs v-model="tab">
<el-tab-pane label="角色管理" name="roles">
<div class="flex-between mb-16">
<el-button type="primary" @click="editRole()">新建角色</el-button>
</div>
<el-table :data="config.roles" border>
<el-table :data="pagedRoles" border>
<el-table-column prop="name" label="角色名称" min-width="140"
><template #default="{ row }"
>{{ row.name }} <el-tag v-if="row.is_builtin" size="small">内置角色</el-tag
@ -55,9 +56,16 @@
></el-table-column
>
</el-table>
<SystemPagination v-model:current-page="rolePage" v-model:page-size="rolePageSize" :total="config.roles.length" />
</el-tab-pane>
<el-tab-pane label="授权管理" name="grants">
<div class="user-roles mb-16">
<div class="grant-toolbar mb-16">
<el-button type="primary" @click="editGrant()">新增授权</el-button>
<div class="user-roles">
<el-radio-group v-model="grantView"
><el-radio-button value="resources">授权项</el-radio-button
><el-radio-button value="subjects">授权明细</el-radio-button></el-radio-group
>
<el-radio-group v-model="grantCategory"
><el-radio-button value="">全部</el-radio-button
><el-radio-button value="MODULE">菜单</el-radio-button
@ -83,13 +91,9 @@
:label="userLabel(u)" /></el-option-group
></el-select>
<el-input v-model="keyword" placeholder="搜索菜单或资源" clearable style="width: 240px" />
<el-button type="primary" @click="editGrant()">新增授权</el-button
><el-button @click="clearFilters">重置</el-button>
<el-button @click="clearFilters">重置</el-button>
</div>
</div>
<el-radio-group v-model="grantView" class="mb-16"
><el-radio-button value="resources">按授权项查看</el-radio-button
><el-radio-button value="subjects">授权明细</el-radio-button></el-radio-group
>
<el-table key="overview" v-if="grantView === 'resources'" :data="pagedOverview" border>
<el-table-column label="菜单 / 资源" min-width="220"
><template #default="{ row }">{{ resourceName(row) }}</template></el-table-column
@ -159,15 +163,11 @@
></el-table-column
>
</el-table>
<div class="app-table__pagination mt-16">
<el-pagination
<SystemPagination
v-model:current-page="grantPage"
v-model:page-size="grantPageSize"
:page-sizes="[10, 20, 50, 100]"
:total="grantTotal"
layout="total, prev, pager, next, sizes"
/>
</div>
</el-tab-pane>
<el-tab-pane label="用户权限" name="users">
<el-select
@ -232,6 +232,7 @@
</template>
</el-tab-pane>
</el-tabs>
</el-card>
<el-dialog
v-model="roleDialog"
:title="roleForm.id ? '编辑角色' : '新建角色'"
@ -371,6 +372,7 @@ import { computed, onMounted, ref, watch } from 'vue'
import { useRoute } from 'vue-router'
import { ElMessage, ElMessageBox } from 'element-plus'
import api from '@/api/system/platform-access'
import SystemPagination from '@/components/system-pagination/index.vue'
const tab = ref('roles'),
loading = ref(false),
saving = ref(false),
@ -383,6 +385,9 @@ const config = ref<any>({
modules: {},
actions: {},
})
const rolePage = ref(1), rolePageSize = ref(10)
const pagedRoles = computed(() => config.value.roles.slice((rolePage.value - 1) * rolePageSize.value, rolePage.value * rolePageSize.value))
watch([() => config.value.roles.length, rolePageSize], ([total, size]) => { rolePage.value = Math.min(rolePage.value, Math.max(1, Math.ceil(total / size))) })
const roleDialog = ref(false),
grantDialog = ref(false),
roleForm = ref<any>({}),
@ -762,9 +767,10 @@ onMounted(async () => {
})
</script>
<style scoped>
.grant-toolbar { display:flex; justify-content:space-between; align-items:center; gap:12px; flex-wrap:wrap; }
.access-page {
padding-top: 40px;
height: 100%;
box-sizing: border-box;
overflow: auto;
background: var(--app-layout-bg-color);
}

View File

@ -1,6 +1,6 @@
<template>
<div class="user-manage p-16-24">
<h2 class="mb-16">{{ $t('views.userManage.title') }}</h2>
<el-breadcrumb separator-icon="ArrowRight" class="mb-16"><el-breadcrumb-item><h5 class="color-text-primary">{{ $t('views.userManage.title') }}</h5></el-breadcrumb-item></el-breadcrumb>
<el-card class="user-manage__card" style="--el-card-padding: 0">
<div class="user-manage__body p-24">
<div class="user-manage__toolbar mb-16">
@ -264,17 +264,13 @@
</el-table>
</div>
<div class="app-table__pagination mt-16">
<el-pagination
<SystemPagination
v-model:current-page="paginationConfig.current_page"
v-model:page-size="paginationConfig.page_size"
:page-sizes="[10, 20, 50, 100]"
:total="paginationConfig.total"
layout="total, prev, pager, next, sizes"
@size-change="handleSizeChange"
@current-change="getList"
/>
</div>
</div>
</el-card>
<UserDrawer :title="title" ref="UserDrawerRef" @refresh="refresh" />
@ -285,6 +281,7 @@
</template>
<script lang="ts" setup>
import SystemPagination from '@/components/system-pagination/index.vue'
import { onMounted, reactive, ref } from 'vue'
import UserDrawer from './component/UserDrawer.vue'
import UserPwdDialog from './component/UserPwdDialog.vue'

View File

@ -362,6 +362,15 @@ export const variableAssignNode = {
},
}
export const componentNode = {
type: WorkflowType.ComponentNode,
text: '调用文件读取、Markdown、模型、导图或 MCP 组件', label: '组件调用', height: 260,
properties: { stepName: '组件调用', config: { fields: [
{ label: '文本结果', value: 'text' }, { label: '结构化结果', value: 'structured' },
{ label: '任务编号', value: 'task_id' },
] } },
}
export const mcpNode = {
type: WorkflowType.McpNode,
text: t('views.applicationWorkflow.nodes.mcpNode.text'),
@ -375,6 +384,7 @@ export const mcpNode = {
label: t('common.result'),
value: 'result',
},
{ label: '文本结果', value: 'text' }, { label: '结构化结果', value: 'structured' }, { label: '任务编号', value: 'task_id' },
],
},
},
@ -653,7 +663,7 @@ export const menuNodes = [
},
{
label: t('views.applicationWorkflow.nodes.classify.other'),
list: [mcpNode, toolNode],
list: [componentNode, mcpNode, toolNode],
},
]
export const applicationLoopMenuNodes = [
@ -691,7 +701,7 @@ export const applicationLoopMenuNodes = [
},
{
label: t('views.applicationWorkflow.nodes.classify.other'),
list: [mcpNode, toolNode],
list: [componentNode, mcpNode, toolNode],
},
]
@ -785,6 +795,7 @@ export const nodeDict: any = {
[WorkflowType.ImageGenerateNode]: imageGenerateNode,
[WorkflowType.VariableAssignNode]: variableAssignNode,
[WorkflowType.McpNode]: mcpNode,
[WorkflowType.ComponentNode]: componentNode,
[WorkflowType.TextToVideoGenerateNode]: textToVideoNode,
[WorkflowType.ImageToVideoGenerateNode]: imageToVideoNode,
[WorkflowType.IntentNode]: intentNode,

View File

@ -250,6 +250,12 @@
</div>
</template>
</div>
<el-form-item label="平台能力">
<el-select v-model="chat_data.model_setting.engine_tools" multiple placeholder="按需启用" class="w-full">
<el-option label="知识库检索" value="knowledge-search" />
<el-option label="生成并保存导图" value="mindmap-save" />
</el-select>
</el-form-item>
<el-form-item @click.prevent v-if="chat_data.mcp_enable || chat_data.tool_enable">
<template #label>
<div class="flex-between">

View File

@ -82,12 +82,12 @@
</el-button>
</div>
</template>
<el-select v-model="form_data.mcp_tool" @change="changeTool" filterable>
<el-select v-model="toolSelection" @change="changeTool" filterable>
<el-option
v-for="item in form_data.mcp_tools"
:key="item.value"
:key="JSON.stringify([item.server, item.name])"
:label="item.name"
:value="item.name"
:value="JSON.stringify([item.server, item.name])"
class="flex align-center"
>
<el-tooltip
@ -99,7 +99,7 @@
<AppIcon iconName="app-warning" class="app-warning-icon"></AppIcon>
</el-tooltip>
<span class="ml-4">{{ item.name }}</span>
<span class="ml-4">{{ item.server }} / {{ item.name }}</span>
</el-option>
</el-select>
</el-form-item>
@ -248,6 +248,11 @@
</el-form>
</div>
</template>
<el-form label-position="top" class="p-8-12" @submit.prevent>
<el-form-item label="调用超时(秒)"><el-input-number v-model="form_data.timeout" :min="1" :max="600" /></el-form-item>
<el-form-item label="网络重试次数"><el-input-number v-model="form_data.retries" :min="0" :max="2" /><el-text size="small" type="info">仅为可重复调用的工具开启重试。</el-text></el-form-item>
<el-form-item label="优先级(0~9)"><el-input-number v-model="form_data.priority" :min="0" :max="9" /></el-form-item>
</el-form>
<McpServerInputDialog ref="mcpServerInputDialogRef" @refresh="handleMcpVariables" />
</NodeContainer>
</template>
@ -302,6 +307,7 @@ const wheel = (e: any) => {
}
}
const form = {
timeout: 60, retries: 0, priority: 5,
mcp_tool: '',
mcp_tools: [],
mcp_servers: '',
@ -365,7 +371,7 @@ function _getTools(mcp_servers: any) {
MsgSuccess(t('views.applicationWorkflow.nodes.mcpNode.getToolsSuccess'))
// 修改了json,刷新mcp_server
form_data.value.mcp_server = form_data.value.mcp_tools.find(
(item: any) => item.name === form_data.value.mcp_tool,
(item: any) => item.name === form_data.value.mcp_tool && (!form_data.value.mcp_server || item.server === form_data.value.mcp_server),
)?.server
})
}
@ -411,13 +417,21 @@ function handleMcpVariables(vars: any) {
_getTools(mcp_servers)
}
const toolSelection = computed({
get: () => form_data.value.mcp_tool ? JSON.stringify([form_data.value.mcp_server, form_data.value.mcp_tool]) : '',
set: (value: string) => {
const [server, name] = JSON.parse(value)
form_data.value.mcp_server = server; form_data.value.mcp_tool = name
},
})
function changeTool() {
form_data.value.mcp_server = form_data.value.mcp_tools.find(
(item: any) => item.name === form_data.value.mcp_tool,
(item: any) => item.name === form_data.value.mcp_tool && (!form_data.value.mcp_server || item.server === form_data.value.mcp_server),
)?.server
const args_schema = form_data.value.mcp_tools.find(
(item: any) => item.name === form_data.value.mcp_tool,
(item: any) => item.name === form_data.value.mcp_tool && (!form_data.value.mcp_server || item.server === form_data.value.mcp_server),
)?.args_schema
form_data.value.tool_form_field = []
for (const item in args_schema?.properties) {
@ -428,7 +442,7 @@ function changeTool() {
let input_type = 'TextInput'
if (params[item2].type === 'string') {
input_type = 'TextInput'
} else if (params[item2].type === 'number') {
} else if (['number', 'integer'].includes(params[item2].type)) {
input_type = 'NumberInput'
} else if (params[item2].type === 'boolean') {
input_type = 'SwitchInput'
@ -447,11 +461,11 @@ function changeTool() {
},
input_type: input_type,
source: 'referencing',
required: args_schema.properties[item].required?.indexOf(item2) !== -1,
required: (args_schema.properties[item].required?.includes(item2) ?? false),
props_info: {
rules: [
{
required: args_schema.properties[item].required?.indexOf(item2) !== -1,
required: (args_schema.properties[item].required?.includes(item2) ?? false),
message: t('dynamicsForm.tip.requiredMessage'),
trigger: 'blur',
},
@ -464,7 +478,7 @@ function changeTool() {
let input_type = 'TextInput'
if (args_schema.properties[item].type === 'string') {
input_type = 'TextInput'
} else if (args_schema.properties[item].type === 'number') {
} else if (['number', 'integer'].includes(args_schema.properties[item].type)) {
input_type = 'NumberInput'
} else if (args_schema.properties[item].type === 'boolean') {
input_type = 'SwitchInput'
@ -483,11 +497,11 @@ function changeTool() {
},
input_type: input_type,
source: 'referencing',
required: args_schema.required?.indexOf(item) !== -1,
required: (args_schema.required?.includes(item) ?? false),
props_info: {
rules: [
{
required: args_schema.required?.indexOf(item) !== -1,
required: (args_schema.required?.includes(item) ?? false),
message: t('dynamicsForm.tip.requiredMessage'),
trigger: 'blur',
},
@ -520,6 +534,7 @@ const form_data = computed({
const replyNodeFormRef = ref()
const isEmpty = (value: any) => value === undefined || value === null || value === '' || (Array.isArray(value) && !value.length)
const validate = async () => {
// 对动态表单,只验证必填字段
if (dynamicsFormRef.value) {
@ -530,7 +545,7 @@ const validate = async () => {
if (requiredFields.length > 0) {
for (const item of requiredFields) {
if (form_data.value.params_nested) {
if (!form_data.value.tool_params[form_data.value.params_nested][item]) {
if (isEmpty(form_data.value.tool_params[form_data.value.params_nested][item])) {
return Promise.reject({
node: props.nodeModel,
errMessage: item + t('dynamicsForm.tip.requiredMessage'),
@ -538,7 +553,7 @@ const validate = async () => {
}
} else {
// 这里是没有嵌套的情况
if (!form_data.value.tool_params[item]) {
if (isEmpty(form_data.value.tool_params[item])) {
return Promise.reject({
node: props.nodeModel,
errMessage: item + t('dynamicsForm.tip.requiredMessage'),
@ -588,6 +603,7 @@ function getMcpToolSelectOptions() {
}
onMounted(() => {
form_data.value.timeout ??= 60; form_data.value.retries ??= 0; form_data.value.priority ??= 5
if (typeof props.nodeModel.properties.node_data?.is_result === 'undefined') {
if (isLastNode(props.nodeModel)) {
set(props.nodeModel.properties.node_data, 'is_result', true)