UnisKB/apps/system_manage/ai_hub_assets.py

214 lines
8.4 KiB
Python

"""Private, durable image files. References live in existing task/artifact records.
AI_HUB_ASSET_DIR must be on persistent shared storage when Web/worker are on
different hosts. No migration, public static route, or Redis image payloads.
"""
import base64
import hashlib
import io
import json
import os
from pathlib import Path
import re
import time
import uuid
from contextlib import contextmanager
from urllib.parse import quote
from django.conf import settings
from rest_framework.exceptions import NotFound, ValidationError
DATA_IMAGE = re.compile(r'data:image/(?:png|jpe?g|gif|webp|bmp);base64,[A-Za-z0-9+/=\r\n]+', re.I)
ASSET = re.compile(r'/admin/api/workspace/([^/\s"<>]+)/ai_hub/assets/([a-f0-9-]{36})/content')
MAX_IMAGE = 20 * 1024 * 1024
MAX_IMAGES_BYTES = 100 * 1024 * 1024
TTL = 7 * 86400
def root():
return Path(os.environ.get('AI_HUB_ASSET_DIR') or getattr(settings, 'AI_HUB_ASSET_DIR', '')
or Path(__file__).resolve().parents[2] / 'data' / 'ai-hub-assets').resolve()
@contextmanager
def asset_lock():
directory = root()
directory.mkdir(parents=True, exist_ok=True)
lock = directory / '.lock'
stream = lock.open('a+b')
stream.seek(0); stream.write(b'0'); stream.flush(); stream.seek(0)
deadline = time.monotonic() + 30
while True:
try:
if os.name == 'nt':
import msvcrt
msvcrt.locking(stream.fileno(), msvcrt.LK_NBLCK, 1)
else:
import fcntl
fcntl.flock(stream.fileno(), fcntl.LOCK_EX | fcntl.LOCK_NB)
break
except OSError:
if time.monotonic() >= deadline:
stream.close()
raise ValidationError('图片存储繁忙,请稍后重试')
time.sleep(.05)
try:
yield
finally:
if os.name == 'nt':
stream.seek(0)
msvcrt.locking(stream.fileno(), msvcrt.LK_UNLCK, 1)
else:
fcntl.flock(stream.fileno(), fcntl.LOCK_UN)
stream.close()
def _write(path, content):
temporary = path.with_name(path.name + '.' + uuid.uuid4().hex + '.tmp')
try:
with temporary.open('wb') as stream:
stream.write(content)
os.replace(temporary, path)
finally:
temporary.unlink(missing_ok=True)
def _metadata(asset_id):
try:
asset_id = str(uuid.UUID(str(asset_id)))
return json.loads((root() / (asset_id + '.json')).read_text(encoding='utf-8'))
except (OSError, ValueError, TypeError):
raise NotFound('图片不存在或已过期')
def authorize(workspace, user, asset_id):
from system_manage.ai_hub import admin
meta = _metadata(asset_id)
if meta['workspace_id'] != workspace or (meta['owner_id'] != str(user.id) and not admin(user)):
raise NotFound('图片不存在或无权访问')
return meta
def image_bytes(workspace, user, asset_id):
meta = authorize(workspace, user, asset_id)
try:
content = (root() / (str(uuid.UUID(str(asset_id))) + '.bin')).read_bytes()
except OSError:
raise NotFound('图片不存在或已过期')
return content, meta
def _store(uri, workspace, user):
from PIL import Image, UnidentifiedImageError
try:
encoded = uri.split(',', 1)[1]
if len(encoded) > MAX_IMAGE * 4 // 3 + 16:
raise ValueError()
content = base64.b64decode(re.sub(r'\s', '', encoded), validate=True)
if not content or len(content) > MAX_IMAGE:
raise ValueError()
with Image.open(io.BytesIO(content)) as image:
if image.width * image.height > 40_000_000:
raise ValueError()
fmt = image.format
image.verify()
mime = {'PNG': 'image/png', 'JPEG': 'image/jpeg', 'GIF': 'image/gif',
'WEBP': 'image/webp', 'BMP': 'image/bmp'}[fmt]
except (ValueError, KeyError, OSError, UnidentifiedImageError, Image.DecompressionBombError):
raise ValidationError('图片格式无效或超过大小限制')
digest = hashlib.sha256(content).hexdigest()
asset_id = str(uuid.uuid5(uuid.NAMESPACE_URL, f'ai-hub:{workspace}:{user.id}:{digest}'))
meta = {'id': asset_id, 'workspace_id': workspace, 'owner_id': str(user.id),
'sha256': digest, 'mime': mime, 'size': len(content), 'touched_at': time.time()}
if not (root() / (asset_id + '.bin')).exists():
_write(root() / (asset_id + '.bin'), content)
_write(root() / (asset_id + '.json'), json.dumps(meta).encode())
return f'/admin/api/workspace/{quote(workspace, safe="")}/ai_hub/assets/{asset_id}/content'
def normalize(markdown, workspace, user, images=None):
"""Caller holds asset_lock until its database reference is saved."""
replacements = {}
total = 0
def store(uri):
nonlocal total
if uri not in replacements:
total += len(uri) * 3 // 4
if total > MAX_IMAGES_BYTES:
raise ValidationError('文档图片总大小超过限制')
replacements[uri] = _store(uri, workspace, user)
return replacements[uri]
if images:
if not isinstance(images, dict):
raise ValidationError('解析图片格式无效')
# Only replace destinations in Markdown/HTML, never arbitrary words.
for name, uri in images.items():
if not isinstance(name, str) or not isinstance(uri, str) or not DATA_IMAGE.fullmatch(uri):
continue
pattern = re.compile(r'(?P<prefix>\]\(<?|src=["\x27]|^\s*\[[^\]\n]+\]:\s*<?)(?:\./)?(?:images/)?'
+ re.escape(name.removeprefix('images/')) + r'(?=[\s)>"\x27])', re.M)
if pattern.search(markdown):
url = store(uri)
markdown = pattern.sub(lambda m: m['prefix'] + url, markdown)
markdown = DATA_IMAGE.sub(lambda m: store(m[0]), markdown)
for match in ASSET.finditer(markdown):
if match[1] != quote(workspace, safe=''):
raise ValidationError('不能引用其他工作空间的图片')
meta = authorize(workspace, user, match[2])
meta['touched_at'] = time.time()
_write(root() / (match[2] + '.json'), json.dumps(meta).encode())
return markdown
def protect_images(markdown):
"""Keep image destinations out of model rewriting; reject lost/forged tokens."""
images = []
pattern = re.compile(r'!\[[^\]\n]*\]\([^\n]*?\)|<img\b[^>]*>', re.I)
def replace(match):
images.append(match[0])
return f'AIHUBIMAGE{len(images):04d}TOKEN'
return pattern.sub(replace, 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)}
restored = set()
def replace(match):
index = match[1].lstrip('0') or '0'
if index not in known or index in restored:
return ''
restored.add(index)
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 and append_missing:
markdown = markdown.rstrip() + '\n\n## 原文图片\n\n' + '\n\n'.join(missing)
return markdown
def cleanup(apply=False, now=None):
"""Only expired, unreferenced Hub images; saved originals/history count as refs."""
from system_manage.models.ai_hub import AIArtifact, AITask
cutoff = (now or time.time()) - TTL
expired = []
with asset_lock():
protected = set()
for artifact in AIArtifact.objects.only('markdown', 'original', 'history').iterator():
content = artifact.markdown + '\n' + artifact.original + '\n' + json.dumps(artifact.history)
protected.update(m[2] for m in ASSET.finditer(content))
for task in AITask.objects.filter(status__in=['PENDING', 'RUNNING']).only('input_data', 'result').iterator():
protected.update(m[2] for m in ASSET.finditer(json.dumps([task.input_data, task.result])))
for path in root().glob('*.json'):
meta = _metadata(path.stem)
if meta['touched_at'] >= cutoff or path.stem in protected:
continue
expired.append(path.stem)
if apply:
(root() / (path.stem + '.bin')).unlink(missing_ok=True)
path.unlink()
return len(expired)