* fix(retrieval): keep index writes and shared-KB lookups correct
Index writes
- Image OCR/caption chunks were indexed without IsEnabled, so every engine
but Postgres stored them disabled and retrieval never returned them.
The SQLite engine re-enables the affected rows at startup.
- FAQ edits, similar-question additions and import merges re-index under
the old SourceID and relied on the engine overwriting it. Postgres and
SQLite kept the stale row (ON CONFLICT DO NOTHING); ES v8 and Qdrant
mint a new document ID and kept both. FAQ re-indexing now deletes the
old rows first, and Postgres/SQLite BatchSave upserts by
(source_id, source_type).
- Postgres: drop the GORM default tag on is_enabled, which replaced false
with true on insert and indexed disabled chunks as enabled.
- SQLite: BatchSave no longer attaches vectors and FTS rows to the wrong
rows when part of a batch already exists, and vecTables is guarded
against concurrent map writes.
- CopyIndices: Postgres filters to the copied knowledge, pages by id and
keeps is_enabled; SQLite copies every row of a chunk (generated
questions included) with stable source IDs and enabled state.
Shared knowledge bases
- Chat merge: FAQ answer population and neighbor expansion look chunks up
without the caller-tenant filter, like parent resolution (#3342).
- Neighbor expansion never cuts the retrieval hit; the length budget goes
to the text adjacent to it.
- FAQ negative questions are applied whenever an FAQ KB is in scope and
to chunks of shared FAQ KBs.
- Search results get image_info for shared-KB chunks.
- search_knowledge reads FAQ metadata from the search result instead of
one tenant-scoped query per hit, so shared FAQs render their answers.
* fix(retrieval): put every retrieval score on one scale
Scores from different retrieval paths were ranked together while living on
different scales, so ordering depended on which path produced a hit.
Fusion
- RRF ranked a chunk by its position in the concatenation of every
vector (or keyword) list, appended in goroutine completion order. The
second list's best hit ranked behind every hit of the first, and the
order changed between identical requests. Each list is now ranked on
its own by score and a chunk takes its best rank.
- RRF is divided by its maximum, so hybrid scores land in [0, 1] like
vector-only (cosine) and keyword-only (rescaled BM25) results. Raw RRF
topped out near 0.016 and lost to any single-retriever hit it was ranked
against across searches (document targets, model groups, FAQ calls);
rerank's base-score term and MMR's relevance term had nothing to use.
- Keyword-only BM25 is rescaled per list; engines' BM25 is not comparable.
- FAQ iterative retrieval fuses each round like the first pass instead of
sorting raw BM25 and cosine together, starts at twice the depth already
searched, triggers when any single vector list is full (the old summed
count never matched with more than one list) and replaces results per
round.
Rerank and chat pipeline
- The FAQ boost is no longer capped at 1; capping made strong FAQs tie at
1.0 and fall to the tie-breaker. The composite score before boosts is
kept as composite_score.
- The FAQ direct-answer threshold is compared against that pre-boost
score, and the FAQ that cleared it is the one marked match="exact"
rather than the first FAQ.
- Graph hits no longer carry a fabricated retrieval score of 1.0, which
ranked them above every scored hit when rerank did not run; the
composite score uses the model score in its place.
Engines
- OpenSearch and Weaviate return (1 + cos) / 2; their drivers now report
cosine similarity and map thresholds onto the engine scale, so a
threshold means the same on every engine.
- Milvus L2 returned squared distance, ranked farthest-first downstream,
and took the threshold as a radius; scores and radius are converted.
- The multi-store normalizer no longer maps Milvus through (s + 1) / 2,
which put it on a different scale from every other engine and from its
own single-store results.
Docs updated for fusion, normalization, FAQ boost and iterative recall.
* perf(retrieval): bound rerank, agent and pipeline cost; stop failing on slow or missing models
Rerank
- Rerankers report the longest document a request can carry
(max_document_chars / max_request_chars minus the query) and the rerank
stage fits its own passages to it, trimming appended captions, OCR and
generated questions first. One chunk with a large screenshot's OCR used
to exceed the vendor limit and fail the whole batch, so every candidate
lost its score and the threshold stopped applying.
- Rerank HTTP requests default to a 60 s timeout; chat and agent contexts
carry no deadline, so a stalled self-hosted rerank server held the turn
open.
- OCR and caption chunks no longer repeat their own text from image_info in
the passage.
- The chat pipeline and the agent tool rerank at most 200 candidates, best
retrieval score first.
- A rerank model that cannot be loaded degrades to retrieval order instead
of failing every chat turn, as the search APIs already do; without rerank
results the merge stage keeps only RerankTopK hits instead of enriching
every one.
Retrieval APIs
- knowledge-search match_count and rerank.top_k are capped at 200
(unbounded values multiplied into tens of thousands of billed rerank
passages); normalizedMatchCount clamps to the pool size so the
over-retrieval arithmetic cannot overflow.
- hybrid-search with rerank recalls at least rerank.top_k deep.
Agent
- search_knowledge fits its rendered output to four fifths of the tool
output budget, dropping the lowest-ranked rows and telling the model how
many it omitted; the model's view is rebuilt from Data, which registry
truncation never reached.
- Agent searches skip neighbor/parent/related enrichment rows, which were
never retrieved yet tripled the rerank input.
- ParallelToolCalls is enabled. The model was always told it may call tools
in parallel while the engine ran them one by one; only read-only tools
overlap, anything else stays a barrier.
Chat pipeline
- Query expansion searches keywords only (its variants are local rewrites
whose embeddings repeat the main search) and is skipped when keyword
matching is off.
- Entity extraction, a second LLM call, is skipped for intents that need no
KB retrieval and uses the rewritten query.
- Partial-overlap dedup tokenizes each text once instead of per pair, and
only compares token overlap between texts within 3x of each other's size.
Storage
- Migration 000112 adds btree indexes on embeddings(knowledge_id) and
(chunk_id); document deletes, chunk re-index and enable/tag updates
scanned the whole table.
- Shared-KB knowledge rows are fetched in one query instead of one per ID.
- Drop GORM Debug() from the ParadeDB keyword query and the knowledge batch
lookup, which logged SQL and user queries on every search.
Docs updated for the API limits, rerank timeout and passage fitting, agent
parallel tools and output budget, and pipeline changes.
* fix(retrieval): make agent tools and pipeline stages report what they did
Agent tools
- query_knowledge_graph now queries the graph store: entity names from the
query are matched in each KB's graph, relations are returned (and shown to
the model), and the entities' source chunks lead the results, with text
search as a supplement. It used to run text search only, so a graph-only
KB always answered "no relevant graph information".
- search_knowledge reports failures instead of "no matching chunks": when
every search fails the tool fails with the reason; partial failures are
listed. A failed query embedding falls back to keyword search for hybrid
KBs, as the chat pipeline does.
- Keyword mode uses rerank to order hits, not to reject them (identifiers
score low), and the empty-result hint no longer sends keyword mode back to
keyword mode.
- Dedup keeps siblings under one parent, no longer merges every image chunk
of a document through chunk_index 0, and keeps the better-scored copy.
- Search rows omit a <match> excerpt that only repeats the content; snippet
terms are segmented for Chinese queries and skip stopwords, and FAQ
matched questions are chosen by match strength.
- read_document query mode returns next_offset when truncated and resumes
from offset.
- wiki_search over-fetches under document/tag scopes before filtering and
says when every match was out of scope.
- Prompts point to read_document(id=cN, context=k) for surrounding text.
- The agent editor hides the retrieval-strategy section in agent mode,
where search_knowledge uses the global retrieval settings.
Chat pipeline
- Intent labels are normalized; an unknown label retrieves instead of
silently skipping the knowledge base, and web_search without web search
enabled searches the knowledge bases.
- Image turns get a 2048-token rewrite budget, and a reply cut off by the
cap is salvaged field by field instead of being discarded.
- History references are matched by how much of the query they cover;
Jaccard against chunk-sized text never reached its threshold.
- Entity search resolves chunks of org-shared KBs within the turn's KB
scope, skips chunks whose document is gone (it used to panic) and caps
graph chunks at 30.
- Empty history is not fetched twice; fetched pages are cut on a rune
boundary.
Retrieval
- A multi-KB search whose primary KB has no vectors embeds the query with a
vector KB's model instead of failing on an empty model ID.
- Rerank passages include the chunk's heading breadcrumb, which is part of
what was embedded.
- SQLite FTS searches the content column only and turns a lone Han
character into a prefix query; TencentVectorDB ef follows TopK.
Docs updated for the agent tools and pipeline stages.
* fix(retrieval): address review findings on the retrieval fixes
- Restore NeedsRetrieval for web_search without web search enabled: the
intent prompt "web_search" is how the user learns web search is off, and
running KB retrieval instead bypassed it.
- query_knowledge_graph keeps the case of entity terms; Neo4j matches node
names with case-sensitive CONTAINS, so lowercased terms never matched
English entities.
- An empty search_knowledge or query_knowledge_graph result states partial
failures in its text, which is all the model sees for an empty result.
- SQLite BatchSave replaces a source in one transaction, so a failed insert
keeps the old rows instead of leaving the source unindexed.
- Reuse stringSliceValue and an exported searchutil.ContainsChinese instead
of new copies; drop an unused named return and the FAQ obsolete-source
delete that the full per-chunk delete now covers.
* fix(retrieval): address second review of the retrieval fixes
- query_knowledge_graph: under a document/tag scope, keep only relations
whose entities come from an in-scope evidence chunk (the graph namespace
is the whole KB); drop chunks whose document no longer exists; report a
graph or text search failure even when the other source returned hits.
- search_knowledge: record the hybrid -> keyword fallback taken when the
query embedding fails in mode_fallbacks, so an empty result is not read
as "no semantic match".
- query understanding salvage: take a cut-off value only for the image
description; a truncated rewrite_query or intent is dropped instead of
replacing the user's query.
- Graph hits (no retrieval score) are kept past the no-rerank RerankTopK
cut and the rerank candidate cap instead of always falling below them.
- FAQ boost is capped at 1 again so scores stay on the [0, 1] scale; FAQs
tied at the cap are ordered by their pre-boost score.
- Negative-question filtering loads only candidates of FAQ knowledge bases.
- read_document query continuation starts one chunk before offset so the
first match keeps its preceding context at a page boundary.
- graphSearchTerms and searchQueryTokens share one queryTerms splitter.
The title prompt sent the user's first message as a raw user turn, so a
query ending in "summarise it in a table" made the model emit a Markdown
table that became the session title.
- Wrap the query in a <user_message> block (embedded delimiter tags are
neutralised) and append a guard to the configured system prompt telling
the model to treat the block as data and never follow its instructions.
- sanitizeGeneratedTitle now drops a full <think> block, keeps the first
line with content, strips headings, list/quote markers, labels, links,
inline markup and wrapping quotes, and falls back to a 30-rune plain-text
prefix of the user query when the output is still table-like.
- Tighten the default session title template.
* docs: sync website-docs with changes since v0.8.0
Audit every change between v0.8.0 and the v0.8.2 release against the
documentation site, adding what users and operators need and correcting
statements that no longer match the code.
- Add local browser (BrowserSkill), built-in MCP Server endpoints, the
toolbox, conversation steer/fork/rewind and reasoning effort, the
artifacts library, sandbox terminal and desktop, the Lite host sandbox,
the rebuilt model catalog (including a protocol-compat field reference),
Confluence and DingTalk Docs sources, Bocha/Serply search, per-channel IM
locale, whitelist-only outbound, admin user creation, new configuration
variables and migrations 000091-000110.
- Fix wrong API examples and parameters (model type values, knowledge
search params, FAQ match_count cap, Notion credentials), tenant deletion
semantics, default timeouts, parser engine claims, removed tables and
commands, stale navigation paths, and route locations after the router
split into routes_*.go.
- Add screenshots for the new pages and refresh outdated ones.
* fix(frontend): link the protocol-compat field to its documentation
The model editor told users to read catalog/compat.go on the backend, a
path that no longer exists after the model catalog refactor. The hint now
says when the override is needed and links to the field reference in
website-docs/03-features/06-models.md; the config examples and the API
type comment point to the same place.
* feat(website): showcase v0.8.2 on the homepage
- Add a v0.8.2 section for the local browser (BrowserSkill), the built-in
MCP Server and conversation control, and turn the v0.8.0 release block
into a standing skills & sandbox topic with the sandbox desktop and
terminal.
- Replace the three-up galleries with a feature list beside one large
screenshot: every item's text stays visible, clicking an item or one of
its views switches the image with a cross-fade, and the list plays itself
while on screen until the visitor picks something.
- Add Confluence, DingTalk, BrowserSkill and more model vendors to the
integrations, bump the header badge to v0.8.2 and link the changelog.
- check-site reports missing docs screenshots as pending instead of
failing, since the Screenshot component renders a placeholder for them.
* docs(changelog): add v0.8.2 release notes
* docs(website): add per-version release notes pages
Add a "版本发布" docs section with one page per release, starting with
v0.8.2 and a backfilled v0.8.0. Each page walks through the highlights
with the existing product screenshots, the other improvements and fixes,
and an upgrade checklist with the actions each breaking change needs.
The sidebar orders these pages by version, newest first; the docs header
gains a 版本 entry, and the homepage's v0.8.2 section now links to the
release notes page instead of the raw CHANGELOG.
* feat(website): let the site image listen on a configurable port
The Docker image hard-coded `listen 80`, so running it with host
networking or on a platform that assigns the container port required
editing the Nginx config and rebuilding. The image now renders its
config from the official nginx template directory at startup and reads
the listen port from `PORT` (default 80); deploy/nginx.conf stays a plain
config for static deployments.
test-docker starts a second container with PORT=8080 and checks the
rendered config and that the site answers on that port; the README
documents the variable.
* fix(website): namespace the nginx listen port variable
* fix(website): support bash-wrapped container startup
* fix(website): align nginx startup with frontend entrypoint
* fix(website): pin nginx runtime for older host compatibility
* fix(website): bake the configured nginx port into the image
* docs(ollama): document embedding model config and RAM guidance (#3480)
Local embedding and generation share the Ollama process at OLLAMA_BASE_URL.
The embedding model name is the local Embedding model's name; EMBEDDING_MODEL_NAME
is read only when builtin_models.yaml references it. The published 8 GB
starting point does not include Ollama weights or optional Neo4j.
Fixes#3480
* docs(ollama): shrink README/docs scope per maintainer review (#3480)
* feat(agent): size agent history by the window and calibrate token estimates
Agent history was capped at history_turns (default 5) because compaction
could not survive a turn. With checkpoints persisted, load history by
tokens instead, and let compaction bound the session.
Budget. The loader drops the oldest turns that do not fit, and what it
drops is lost: neither replayed nor summarized. A budget equal to the
compaction threshold (the earlier attempt, since reverted) let the loader
trim first, so whenever one turn was larger than the system prompt plus
the new question the request never crossed the threshold; nothing was
summarized and the session became a sliding window with no checkpoint.
agent.HistoryTokenBudget is now the whole window: overflow reaches the
first round's compaction, which summarizes it into a new checkpoint.
A turn-by-turn regression test drives the real loader, budget and
compactor and fails as soon as a stored turn after the checkpoint is
missing from the loaded history; restoring the old budget fails it.
Summarizer input. History can now reach the window, so the part being
summarized can outgrow one request. An overflowing request degrades to a
raw archive that is never persisted, so every later turn would fail the
same way. Prepare keeps the newest messages that fit, the prompt states
how many were left out, and file paths still come from all of them.
Calibration. cl100k counts Chinese at about a token a character where
Qwen and DeepSeek spend about 0.6. The estimator now carries a scale
(clamped to 0.5-1.5, text only), learned from consecutive requests in a
turn: they share tools, system prompt and earlier messages, so the
provider's growth between them prices exactly the appended messages,
without mixing in how a provider renders tool schemas. The turn's scale
is stored on its usage (context_token_scale); the next turn's loader
prices history with it and hands it to the engine. Loading, the first
compaction check and compaction itself share one estimator, so they
agree; calibrating only the trigger would let the loader drop turns
again, which the regression test covers.
Also: history is priced as the engine sends it (redacted KB/Wiki results
cost one line); a checkpoint older than the rows read is matched by the
assistant rows' (created_at, id) order; a read ending on a turn's
question keeps that turn; a compaction discarded for freeing under 5%
still persists its checkpoint; a checkpoint write matching no row is an
error; a partial index backs the newest-checkpoint lookup. Agent mode no
longer reads history_turns: the field leaves AgentConfig, the editor
hides it in smart-reasoning mode, and quick-answer keeps its turn cap.
* fix(agent): address review of window-sized history loading
Summarization. The summarizer call had a fixed 60s total timeout. With
history loaded up to the window, a request could carry ~175k tokens and be
cut off while prefilling or writing normally; the degraded result left no
checkpoint, so the next turn reloaded the same history and failed the same
way. Summarization now streams and is cancelled only after no output for the
engine's stall timeout (120s by default, LLMCallTimeout overrides), with the
overall ceiling left to the transport, as for the engine's own rounds and as
in Codex and Pi. A stopped turn is not retried. When the summarizer still
fails, the raw archive is persisted as a checkpoint marked degraded so the
loop is broken, and the next compaction refines it as the previous summary.
The archive is held to the summary budget, newest messages first, so a
degraded compaction always frees room.
Loading. The loader priced turns as sent but returned and held them as
stored, so wiki pages stored in full and sent as one line could pile up in
memory. It now returns the as-sent history and releases each turn's stored
rows once replayed, keeping about a page of rows at a time.
Index. The partial checkpoint index is replaced by a composite
(session_id, created_at DESC, id DESC) index serving the backwards page,
the newest-checkpoint lookup and the recent-messages reads. It is built with
CREATE INDEX CONCURRENTLY so writes to messages are not blocked; the
migration holds that single statement, since a multi-statement query runs as
an implicit transaction. A PostgreSQL integration test, skipped without
WEKNORA_MIGRATION_TEST_POSTGRES_DSN, applies the chain, runs 000106 down and
up through golang-migrate, and checks both queries use the index without a
sort; the SQLite migration test checks the same plans.
Calibration. It no longer waits for a total token count, and a usage with
only a scale is persisted. A turn that measures nothing carries forward the
scale it started from, so the newest turn always has the latest one.
Repeating the same query in another mode is usually wasted, but keyword
mode can still recall identifiers that hybrid never surfaced. Stop telling
the model that switching mode cannot help.
The agent searched seven times for "Echo 开放权重模型 降低成本". Hybrid
retrieval ranked the right chunk first (RRF rank 0) every time, but the
rerank model saw only the chunk text, which never names Echo or cost;
it scored 0.0025 and all 18 candidates fell below the 0.15 fallback
floor. The tool reported "No matching chunks", and the model cycled
hybrid -> semantic -> keyword, which cannot help since every mode shares
the same rerank check. Scoring the same chunk with its document title
("Show HN: Echo - ... at 1/3 the cost using open-weight models") gives
0.45, and 0.55 for the query written as a question.
- Rerank passages are title + enriched chunk (rerankPassage); FAQ chunks
keep their own question, and the pattern filter still sees chunk text
only.
- When retrieval found candidates but the rerank model rejected all of
them, the empty statement says so and that switching mode will not
help, instead of "No matching chunks".
- The tool description, query schema and agent system prompt ask for a
complete natural-language question in hybrid/semantic mode rather than
a keyword list, and drop "switch mode" as the retry advice. Every
keyword engine matches any token (OR + BM25), so a question costs the
keyword leg no recall; on the same 20 bases the Echo chunks rank the
same for both query styles.
* refactor(agent): consolidate knowledge retrieval tools into search_knowledge, read_document, list_documents
The agent's retrieval surface had grown to five overlapping tools whose
behaviour was scattered across long descriptions, interleaved "tips" in
tool output (most of which modelcontext discarded before the model saw
them), and a grep_chunks implementation that ran an unindexed regex scan
over the chunks table ordered by created_at, so large knowledge bases lost
recall and paid a full-table scan on every call.
Tool surface
- search_knowledge replaces knowledge_search and grep_chunks: one query,
mode = hybrid | semantic | keyword, optional knowledge_base_ids, limit.
Keyword mode runs through the retrieval engine's keyword index instead of
the DB regex scan. Modes unavailable for the scope are refused with a
reason. Rerank, MMR and de-duplication are kept; already_seen markers,
retrieval_statistics and the zero-result "CRITICAL" block are gone.
- read_document replaces list_knowledge_chunks, get_document_info and
wiki_read_source_doc: id accepts a dN or cN handle, always returns the
document header, pages by exact offset (fixing the old floor-to-page
bug), finds passages with query (literal by default, regex opt-in) and
reads neighbours around a chunk with context.
- list_documents pages through one knowledge base, honouring @file/@tag
scope.
- wiki_search takes a single query plus regex, knowledge_base_ids, limit;
literal queries are escaped so the model never reasons about JSON
escaping. Legacy queries / knowledge_base_id are still accepted.
- query_knowledge_graph is registered only when a graph-enabled KB is in
scope and its description/tips were trimmed.
Compatibility
- Retired names in stored allowlists, presets and API calls are mapped to
their successors by NormalizeAllowedTools at registration time; no data
migration. Stored histories with old names still decode, render and
redact. Calls to retired names return a replacement hint.
- modelcontext gains an "id" key space that decodes against both the chunk
and document handle tables (decode-only), plus rendering for the
document header, chunk roles, in-document matches and next_offset.
Prompts and surfaces
- The RAG, wiki researcher and hybrid templates now name the tools and the
search -> read -> cite order; tool descriptions are short and contain no
workflow instructions.
- Engine display names, history redaction, grounding prompt, IM rendering,
the built-in MCP server (search_knowledge gains mode/limit, grep_chunks
becomes keyword search, read_document gains query, wiki_search gains
regex), frontend editor/renderers/i18n, presets, website docs and the
changelog are updated accordingly.
Tests cover name mapping, mode validation and refusal, output shape,
read_document paging/context/query/scope, list_documents, wiki_search
escaping and legacy arguments, modelcontext decoding and rendering, and
registration-time aliasing and graph gating.
* fix(agent): address review findings on read_document, list_documents and model rendering
- read_document: fetch the second page whenever the first page did not
reach the end of the document, so an unaligned window near the end no
longer drops the trailing chunks (also fixes context windows around a
chunk near the end); a query passed together with a cN handle now
searches the owning document instead of being ignored; query mode stops
collecting once the output budget is nearly spent and reports truncated;
the document row is authorized in place instead of being fetched twice;
the context read takes the total from the window fetch instead of a
separate count query.
- list_documents: resolve the page's document scope in one batched tag
lookup instead of one call per row.
- modelcontext: render read_document output once and read next_offset
through the existing intValue helper.
- search_knowledge: point title/file-name lookups at list_documents.
- Tests for the boundary window, chunk context near the end, cN + query,
and budget-bound query results.
* fix(agent): close compatibility gaps in the consolidated retrieval tools
- Wiki-only scopes: read_document and list_documents are registered
whenever a wiki KB is in scope, since every KB stores chunks. Their
capability requirement includes wiki but is marked auxiliary, so they do
not widen a RAG agent's derived "all knowledge bases" filter to wiki-only
bases (Go and frontend derivations both honour it).
- search_knowledge resolves the retrieval mode per knowledge base instead
of refusing the call: FAQ and vector-only bases answer keyword requests
semantically, keyword-only bases answer semantic requests by keyword.
Results carry requested_mode and mode_fallbacks, the empty-result text
names the fallback, and the call is refused only when no base in scope
has a chunk index.
- The model rendering uses the mode the search actually ran with (it was
always labelled semantic) and emits <mode_fallback> per degraded base.
- list_documents applies the @file/@tag scope before paging: tags are
pushed into the database filter, pinned documents are loaded by ID, and
total/next_page count only readable documents.
- read_document resolves chunk context by chunk_index through a new
ChunkRepository.ListChunkNeighbors, so parent, summary and image chunks
sharing the index sequence no longer shift the neighbourhood; offset is
documented as a reading-order position, not a chunk index.
- wiki_search (agent and MCP) keeps its regular-expression default and
only matches literally when the text is not a valid expression;
regex=false forces a literal match and regex=true rejects invalid input.
- MCP grep_chunks keeps grep semantics: literal terms from the pattern
fetch candidates from the keyword index (semantic for bases without
one), and every returned chunk is verified against the pattern.
* docs(changelog): align the consolidation entry with regex wiki_search and grep_chunks semantics
* docs(mcp): note that grep_chunks recall is bounded by index candidates
* feat(knowledge): derive knowledge base descriptions from document profiles
Knowledge bases only had a hand-written description, which most users
leave empty, so agents bound to several knowledge bases had nothing to
route on and searched all of them. This adds a generated description
that is derived from per-document data instead of maintained by a model.
Document profile: the summary call now returns JSON with a 2-3 sentence
summary plus a gist, 3-5 topic keywords, a document type and one
typical question (knowledges.profile). Input is capped at 8k chars and
output at 1024 tokens; plain-text replies from custom templates still
work and simply carry no profile. The summary chunk embeds the gist and
short summary.
Knowledge base aggregate: a database-only fold over those rows (counts
per file type, tag, normalized topic and doc type; stratified samples of
titles and questions; a hash of the inputs). Deleting, moving or
re-parsing a document changes nothing but the rows, so the next rebuild
is exact without any model judgement.
Description text: one small model call over the aggregate yields a
gist, merged topics and typical questions (knowledge_bases.
generated_profile). The call is skipped when the aggregate hash is
unchanged; failures keep the previous text and record the error; an
empty knowledge base clears it. The manual description is never
touched, and the agent runtime context renders both so the model can
pass knowledge_base_ids instead of fanning out.
Refreshes are requested from summary task exit, post-process completion,
deletion, cross-KB moves and enabling the option, debounced per
knowledge base in 30 s windows via kb:profile on the summary queue.
POST /knowledge-bases/:id/profile/generate regenerates synchronously.
profile_config (enabled, model_id, custom_instructions) opts a document
knowledge base in; the settings dialog shows the generated card with
regenerate and adopt-as-description actions in five locales. Search
results now log hits per chunk type so the value of indexing summary
chunks can be measured.
Migrations 000101 (PostgreSQL) and 000021 (SQLite) add the columns; the
SQLite schema test expects the new version and columns.
* fix(knowledge): keep generated descriptions out of duplicates and refresh clones
DuplicateKnowledgeBase copies the source row through JSON, which carried
generated_profile onto a knowledge base that has no documents yet, so
an empty duplicate advertised the source's topics. The duplicate now
starts without a generated description and keeps profile_config so it
derives its own once documents arrive.
CopyKnowledgeBase builds the target field by field and skipped
profile_config, and the clone task never asked for a description after
copying the documents, so a clone stayed undescribed until its next
upload. The target now carries a copy of profile_config and both clone
paths request a refresh on completion.
* fix(knowledge): refresh the loaded clone target and carry document profiles
ProcessKBClone requested the description refresh with the reservation
stub it builds before CopyKnowledgeBase runs. For the main create path
that stub has neither a type nor profile_config, so the refresh helper
rejected it and nothing was enqueued. The refresh now uses the loaded
target returned by CopyKnowledgeBase.
cloneKnowledge copied the summary text but not the structured profile,
so even a scheduled refresh would have aggregated titles without topics
until every clone was re-summarized. Cloned documents now carry an
independent copy of the profile.
* refactor(wiki): consolidate source_refs parsing and resolution
source_refs entries ("uuid" or "uuid|title") were parsed by ten
independent implementations that disagreed on whitespace handling, on
whether a model-supplied title could be trusted, and on how a leading
pipe was treated. The scope-enforced and legacy agent resolvers also
duplicated the lookup and the routing step looked every document up a
second time.
- types.ParseWikiSourceRef / FormatWikiSourceRef / WikiSourceRefMatchesKnowledge
are now the only readers and writer of the format.
- resolveWikiSourceDocuments replaces resolveAuthorizedSourceRefs,
resolveSourceRefs and wikiKnowledgeBasesForSourceRefs: one lookup per
document, titles always rebuilt from server data, dedupe by canonical ID,
unknown documents fail closed in every mode.
- removeSourceRef now matches padded or titled refs of a deleted document.
* refactor(data_analysis): expose the selected document as a fixed table name
The tool required the document ID inside the SQL text and blindly
string-replaced it with the physical DuckDB table name, which forced the
model-context codec to rewrite handles inside quoted SQL and let the
UUID-derived k_<uuid> table name reach the model through stored schema
summaries.
The document is now selected only by knowledge_id. Every query runs on a
dedicated connection where "dataset" is a temporary view over the
physical table (DuckDB temporary objects are connection-local, so
concurrent sessions stay isolated). The physical name remains accepted
for internal callers; any other table fails validation with an error
naming the table the model should use. Schema summaries, the schema
tool, the analyst prompt and the tool schema all name the same table, so
data_analysis.sql no longer carries source handles and its text codec
policy is removed.
* feat(modelcontext): report raw identifiers that survive encoding
Text-level compaction exists because raw UUIDs reach model context through
producers the registry never sees: error strings, slugs, results of
unknown shape. Retiring it requires knowing where those producers are.
Registry.LeakedIdentifiers scans an encoded message list for UUID-shaped
values in every application-composed field (user messages are skipped),
classifies each as registered (codec missed a rewrite) or unregistered
(producer emitted a raw ID), and SummarizeLeaks renders one line per
field. The agent engine and the three RAG pipeline call sites log the
report as a warning after EncodeMessages.
* fix(chat_pipeline): align RAG data analysis with the fixed table contract
The RAG data-analysis plugin still asked the planning model to fill in
knowledge_id and to write SQL from a prompt that exposed the raw
knowledge UUID, then passed the model's JSON straight to Execute. Under
the new data_analysis contract a query that names anything but the
dataset table is rejected, and the plugin logged that rejection at error
level and moved on, so the analysis vanished without any signal. A
"no analysis needed" plan took the same error path.
The pipeline now owns document selection: the model returns only
needs_analysis and sql, the prompt names the dataset table and never
shows the knowledge ID or physical table, and the plugin builds the tool
input itself. A rejected query is fed back to the model once for repair;
a query that still fails is reported through pipelineWarn with the SQL
and the tool error, and "not needed" is logged as such instead of as a
failure.
Covered by unit tests with a scripted model and executor: not-needed,
pipeline-owned knowledge_id, one-shot repair, persistent failure
surfaced, unparseable plan rejected.
* feat(browser): enhance BrowserIcon component and update localization strings
- Updated BrowserIcon component to accept dynamic width and height props.
- Revised localization strings in English, Japanese, Korean, Russian, and Chinese for improved clarity and user guidance regarding browser connection and usage instructions.
- Enhanced BrowserConnectionSettings view with better structure and user feedback elements.
* Default BrowserSkill tasks to a separate window
* refactor(prompts): streamline agent system prompts and context templates
- Updated agent system prompts to enhance clarity and specificity in user interactions.
- Revised context templates to better separate metadata and user requests, improving overall structure.
- Adjusted fallback and intent prompts to ensure consistent language handling and response requirements.
- Introduced new localization strings for browser search instructions across multiple languages.
- Added tests for agent prompt serialization and hydration to ensure template integrity and functionality.
* chore: wrap prompt strings to satisfy golangci-lint
Keep the settings, separate-window, and prompt commits pushable together
without rewriting their history.
* fix(frontend): match agent editor save test to serialized payload
The prompt-inheritance save path still closes after update; the test
should assert that payload, not the unsanitized form object.
* feat(browser): implement browser connection management and related tests
- Introduced a new `browserConnection` store to manage the state of browser connections, including online status and device information.
- Updated the `Input-field` component to handle browser connection states, enabling/disabling features based on the connection status.
- Added tests for the `browserConnection` store to ensure correct behavior under various scenarios.
- Enhanced the `Input-field` and `BrowserConnectionSettings` components to provide user feedback based on the browser connection state.
- Adjusted related components and views to integrate the new browser connection logic, improving overall user experience.
* fix(frontend): update model-save assertion for serialized agent payload
The agent editor still persists the selected model ID via a spread of
formData; the source test should match that payload, not the raw form.
Second pass over the strings touched by the proofreading commit,
reverting over-corrections that read worse than the original:
- Progress messages go back to the 〜しています… sentence form; the
clipped 〜中… labels were unnatural in context
- Keep established katakana terms (プログレッシブエージェンティックRAG,
エージェンティックRAG, Wikiレビュー, ネガティブ例, 生の出力) instead
of kanji paraphrases; keep クエリ in the embed widget
- Remove commas placed inside noun phrases (〜のための、より〜) and fix
stacked-modifier sentences (共有解除の確認文, 権限不足の説明文,
並列実行数の説明)
- Restore sentence-form placeholders/toasts where the noun-only form
sounded cut off (モデルに直接質問できます, お待ちください…)
Follow-up to #3080. A full proofreading pass over the Japanese locale
against the en-US and zh-CN sources, fixing mistranslations, awkward
phrasing and inconsistent terminology:
- Meaning: "organization" contexts now consistently say 共有スペース
(not bare スペース); remove-from-space actions no longer read as
deleting the member/agent/KB; built-in model description says
"visible to" rather than "usable by" all workspaces; several
descriptions (parser features, xlsx header option, runtime capacity,
credentials invalidation) corrected against zh-CN
- Terminology: 改写 -> リライト, 机器人 -> ボット, 知识库 -> ナレッジベース
(not ナレッジ), ReRank -> リランク, サインイン -> ログイン,
制御プレーン -> コントロールプレーン, API-key capabilities -> 権限,
"Agentic RAG" kept in English, self-hosted -> セルフホスト
- Wording: fix broken particles/係り受け (JWTに署名, 〜に失敗), drop
unnatural potential forms (超えられません), align status chips
(〜済み), progress messages use the 〜中... form, placeholders follow
the English form, 未提供 -> なし, remove trailing ! on error toasts
- Conventions: keep ヘッダー / リンター with the long-vowel mark; remove
remaining Latin/Japanese spaces (チャンク#{index}, unix://を)
- Prompt template and agent preset labels get the same fixes
* feat(i18n): add Japanese (ja-JP) locale support
Add a complete Japanese UI locale for the admin frontend and the web
embed widget, and register it everywhere the supported-locale list is
enumerated.
- frontend/src/i18n/locales/ja-JP.ts: full bundle (5,600+ keys) aligned
with en-US/zh-CN key structure; verified by the locale key audit
- frontend/src/i18n/embed.ts: ja-JP block for the embed widget, plus
`ja*` normalisation in normalizeEmbedLocale
- Register ja-JP in i18n index, SUPPORTED_LOCALES, localeKeyAudit
bundles/order, TDesign locale map (ja_JP), login language picker,
general settings dropdown and embed default-locale options
- Add `language.jaJP` key to every locale so the settings dropdown can
use $t like its siblings
- Accept ja-JP as an embed channel default_locale on the backend and as
DEFAULT_LOCALE in the frontend docker entrypoint
- Extend locale-enumerating tests and update accepted-value lists in
.env.example, public/config.js, widget script and website-docs
* feat(i18n): add ja-JP names and descriptions for prompt templates
Add a ja-JP block to the i18n map of all 32 prompt templates
(system_prompt, context_template, rewrite, fallback, intent_prompts,
agent_system_prompt) so the template picker shows Japanese labels when
the UI locale is ja-JP instead of falling back to the Chinese names.
Prompt content is unchanged.
* feat(i18n): add ja-JP strings for the CLI integration page
main gained the CLI setup page (#3081) with 18 new integrations.cli.* /
integrations.tabs.cli keys; add their Japanese translations so the
locale key audit stays green after rebasing onto main.
* style(i18n): unify ja-JP spacing and long-vowel conventions
- Remove spaces between Latin tokens and adjacent Japanese text
(e.g. "MCP サービス" -> "MCPサービス"), matching the majority style
already used in the bundle
- Drop the trailing long-vowel mark on common technical katakana terms
(サーバ, ユーザ, フォルダ, プロバイダ, パラメータ, ヘッダ, クラスタ, ...)
so the whole locale follows one convention
* feat(i18n): add ja-JP labels for agent type presets
config/agent_type_presets.yaml carried default / zh-CN / zh-TW labels
only, so the agent editor showed English preset names in the Japanese
UI. Add a ja-JP label and description to all five presets.
* feat(sandbox): run sandboxed execs as root by default
Every chat session owns its own sandbox, so the in-container account is
not a tenant boundary and root is the least surprising default for an
agent that installs packages and writes wherever it needs. Host and
cross-tenant isolation stay at the container boundary.
DefaultSandboxExecUser becomes "root" (remote_client.go) and the sandbox
image ends with USER root. The three backends keep naming the account
explicitly on every exec so the account a call lands on does not depend
on which backend a workspace selected; the adapter default, the file
operations and the artifact-directory bootstrap all follow the constant.
The template still provisions a uid 1000 "user" account for interop with
E2B/Cube tooling that addresses it by name.
Root also makes the skill-install flow's account hand-over pointless, so
it goes: verification no longer runs after a chown/chmod lock-down and a
repair round no longer reopens the tree (normalizeSkillPermissions and
reopenSkillDirForRepair are gone), and the prepare and scratch commands
stop chowning. The scratch wipe now also clears /tmp, and the cache
budget guard sizes only root's caches, since no second account can grow
one. Verification is asserted to stay on the ordinary executor (no
AsRoot, no AllowSkillsRoot) rather than on which account it ran as.
* refactor(skills): install chat-time extras into the skill's own env
PYTHONPATH / NODE_PATH pointed at a per-session overlay
(/workspace/.skill-packages/<skill>) because the skill tree was frozen
after install: root-owned, mode 555, and often built by `uv venv` with no
pip, so `pip install --target` somewhere else was the only way a chat
session could add a package.
None of that holds now. The lock-down is gone and execs run as root, so a
missing package goes straight into the skill's own environment, and the
write is already session-local: a chat sandbox is built from the config's
image and its filesystem dies with the session, so nothing reaches the
image other sessions start from. The overlay bought nothing over that
while splitting one skill's packages across two locations.
SessionSkillPackageDir goes, and applySessionPackagePath becomes
applySkillNodePath: it appends the skill's own node_modules after
whatever the caller supplied, replacing a `+=` that prepended a bare
separator once the overlay stopped seeding the variable. No PYTHONPATH is
injected at all. Python dependencies are reached through the skill's venv
interpreter, which the shell wrapper already puts first on PATH; the
overlay-era value named the venv root, which is not an importable path.
PYTHONPATH stays on the injected-name blacklist so a stored value cannot
shadow the venv.
The runtime hints and every prompt that described the overlay now name
`<skillDir>/.venv/bin/python -m pip install` and `npm install` in the
skill directory. frozenSkillTreeGuidance / isFrozenSkillVenvFailure are
renamed for what they detect now that nothing is frozen.
* docs(prompts): reserve /workspace/output for finished deliverables
/workspace/output is the one directory an artifact collector reads, so
everything written there is offered to the user as a deliverable. The
prompts described it as the place for "deliverables" or "downloadable
artifacts" without saying it is exclusively that, which reads as "the
directory you write results to" — and a model doing multi-step work has
results at every step. Drafts, intermediate renders and debug dumps land
next to the finished file and the user is handed all of them.
The four places that describe the workspace now say the same thing: the
collector reads only /workspace/output, so finished deliverables go there
and everything in progress goes anywhere else under /workspace. The
attachment block gains the matching half for the other direction —
/workspace/input holds the user's own files and is not ours to write to,
which it previously implied by calling the paths "read-only inputs" while
the tools that enforce it say so plainly.
Only the /workspace/input half is enforced (write_sandbox_file refuses
that tree). Whether a file is finished is a judgement no tool can make,
so the output half stays guidance and is stated wherever the model is
already reading about the workspace rather than in one place it might
not see.
* fix(skills): restore executable bundles and package recovery
* style(sandbox): wrap prompt text for repository lint checks
* fix(sandbox): guard recovery paths and document root execution
---------
Co-authored-by: hailongzhao <hailongzhao@tencent.com>
Co-authored-by: wizardchen <wizardchen@tencent.com>
* feat(agent): unify sandbox tools and drop host preloaded skills
Read workspace and skill files through read_file, and run installed or staged skills through shell_exec(skill_name). Remove the unused skills/preloaded tree and keep generated-file previews stable across re-renders.
* fix(frontend): do not assume classList in markdown sanitize hooks
Node unit tests stub anchors without a DOM classList. Detect download cards from class/download attributes instead.
Provider prefix caches only hit when system/tools stay byte-stable and
cache routing is actually on the request. Stop rewriting those prefixes
mid-turn, attach prompt_cache_key / cache_control, and log hit rate
instead of showing it in the conversation tree.
Refuse truncated tool calls, compact-and-retry once on overflow, and let
sandbox files be appended, paged, and batch-edited instead of rewritten.
Models can set context_window (default 200K) so compaction sizes itself
from the real window; agent max_iterations of -1 means unlimited.
The installer agent was recently given its working directory up front and
stopped prefixing `cd <skill-dir> &&` onto every command. Session agents
never got the same treatment: shell_exec only mentioned /workspace as the
work_dir default, so the model could not tell it was already there, and
execute_skill_script never said scripts run with /workspace as their cwd —
a relative path inside a skill script silently resolves outside the skill.
Describe the workspace once, where each tool is described: shell_exec gains
a Working Directory section (already in /workspace, relative paths, no `cd`
prefix, and each call starts there again so a `cd` does not carry over),
execute_skill_script states its cwd and points at $WEKNORA_SKILL_DIR, and
the skills prompt gains a Workspace section covering input, output and the
per-session package overlay.
Also drop the installer prompt's paragraph about write_sandbox_file and
friends: those tools are not registered in install mode, so the model never
sees them in its schema and the note only spent tokens describing — and
misdescribing — a scope it could not reach.
- Added write_skill_file and edit_skill_file tools to facilitate writing and editing files directly within the skill directory during installation.
- Updated the agent service to register these tools exclusively for the built-in skill installer, ensuring they operate within the correct scope.
- Enhanced error handling and validation to prevent unauthorized file access outside the designated skill directory.
- Improved documentation and prompts to clarify usage and limitations of the new tools, emphasizing the importance of proving import resolutions by executing them.
- Added tests to validate the functionality and security of the new file management tools.
* perf(skills): batch install verification and keep caches under budget
The tail of an install (agent_done -> verified) issued one remote exec per
check: two for permission normalisation, one per bundled script for the tree
check, three for the scratch wipe. A skill with twenty scripts paid thirty-
odd round trips, each carrying a hidden bootstrap exec and a distributed-
lock resolve, while the console sat on a finished agent with no feedback.
- normalizeSkillPermissions sends chmod and chown as one command.
- verifySkillTree folds the SKILL.md probe and one test per script into a
single command that reports every missing file instead of stopping at the
first. Exit codes 0/1/3 carry the verdict, stderr lines are the final
messages, and the two existing error texts survive verbatim.
- cleanImageScratch merges the workspace reset, the directory restore and
the cache pass into one command whose exit code is still owned by the
strict workspace half. Package caches are no longer wiped wholesale: uv,
npm and pnpm trim their own entries, and the pip/uv/npm/pnpm totals of
both accounts are kept while under 256MB and wiped whole over it, so the
next install resolves a previous download instead of the network.
- The install prompt hands the agent absolute paths for uv, npm, pnpm, pip,
python3 and node from the same one-command probe that used to ask only
whether uv exists, so it neither burns a turn on `which` nor gambles on
the shell PATH.
Tests pin the new command shapes exactly, cover quoting of archive-chosen
file names, assert the tree check stays one command regardless of script
count, exercise the budget guard against a real shell, and keep the
workspace-restore failure semantics intact.
* fix(skills): credit a script's own sys.path bootstrap for shipped modules
The python check resolved every import against the script's directory, its
ancestors up to the skill root and the venv - and nothing else. A module the
skill ships in a sibling subtree (lib/image_video.py, imported by
scripts/generate.py) was invisible to that set, so a skill whose scripts
bootstrap their own library directory was refused for a dependency no
package provides, and the repair round that followed could not fix it
either: it installs packages and is forbidden from editing the tree. One
real install spent seven minutes discovering that, three LLM stream
timeouts included.
The verifier now scans what the tree ships and statically evaluates the
sys.path.insert/append calls each script raises before its imports - the
os.path and pathlib idioms, with variables tracked in source order. A
bootstrap proven to raise the directory that provides the module turns the
finding into a note; anything unprovable stays a refusal, now naming the
shipped file instead of pretending a dependency is missing. The refusal is
unrepairable on purpose: no installer round can fix a shipped module the
import cannot reach.
Tests pin the incident's exact shape, the pathlib idiom, and the guarantee
that makes the evaluator safe: a vendored module still cannot satisfy an
import unless the script's own bootstrap provably reaches it, and modules
pip placed under .venv never count as shipped.
* feat(skills): move the install progress bar while the installer works
The progress card sat at the seeded 35% for the whole agent phase and
only jumped at agent_done. Every tool call the transcript already sees
now advances an asymptotic curve 35 + 44·(1 − e^(−k/12)): monotonic for
any round count, converging below the agent_done anchor at 80, topping
out around 75 even when max_iterations is exhausted.
Each step publishes a one-line command summary (engine hint preferred,
80-rune cap) so the admin can see what the installer is doing. After
the first round ends the activity progress is muted — verification and
repair rounds own the stage anchors (80/82) from there, and a repair
round's tool calls must not drag the bar back under them.
* fix(skills): pin the installer agent's thinking mode off
installerAgentConfig left AgentConfig.Thinking nil, which most providers
read as "defer to the model's own default" — a thinking-capable model
would reason on every ReAct round of an install, burning latency and
completion tokens on what are short shell-command turns. Set an explicit
false so every provider path (Qwen enable_thinking, Ollama think, LKEAP
thinking type) receives a deterministic off instead of a default.
* style(skills): gofmt tenant_skill_verify_python_test.go
CI rejects aligned map keys in this table-driven test; gofmt drops the extra spaces.
* docs(skills): drop the duplicated exit 1 line in the python verifier
The shipped-module case was appended by repeating the whole exit 1 block
instead of extending the existing one.
---------
Co-authored-by: hailongzhao <hailongzhao@tencent.com>
Unset (0) now resolves at call time (2048 / 4096 / 24576 with sandbox) so the editor can offer Default vs Custom without silently rewriting a saved value.
Generated PPT/report scripts overflowed the shell_exec command cap, and
session agents kept mutating the frozen skill tree or failing SkillHub
installs on display titles. Add write/edit sandbox tools, attach the
skill venv to /workspace scripts, and accept slug-based install names.
Introduce the hidden skill-installer agent and TenantSkillService so a
sandbox config can install or remove skills into snapshot images under a
per-config lock, without yet exposing HTTP or UI entry points.
Co-authored-by: Cursor <cursoragent@cursor.com>
Move citation protocol injection to the system layer so custom prompts no
longer own citation syntax. Agents can disable inline citations while source
aliases still encode and decode tool arguments correctly.
Replace raw chunk/document/KB IDs in LLM context with short cN/dN/bN/wN
aliases, decode tool arguments on the way in, and expand compact <ref/>
citations to the public <kb>/<web> format before user-facing output.
Update the classification logic for follow-up questions to include reasoning about previously described images or documents. Ensure that follow-up queries referencing earlier attachments are correctly identified and processed without triggering unnecessary knowledge base searches. This change improves the model's ability to maintain context and respond accurately to user inquiries related to past interactions.
When retrieved context or tool results contain Markdown images, dynamically
append output requirements to system prompts and final-answer instructions.
Also normalize fullwidth image parentheses in the frontend renderer so
localized model output still renders correctly.
Chat and agent tool context now emits answer-ready Markdown images instead
of internal <image> XML so copied content renders safely. The frontend also
normalizes legacy XML blocks in persisted conversations during streaming.
Tighten per-turn @MCP/@Skill/@tag handling after review: enforce agent
MCP/Skill whitelist, ignore mentions when selection mode is none, tag-scope
grep_chunks, and simplified English must_use lines without tool dumps.
Skip persisting scope envelopes to history rendered_content.
Wire per-turn @mentions for skills, MCP services, tags, and files through
QA handlers into agent config, runtime_context/must_use user blocks, and
tool registration. Fix tag-only scopes dropping KB tools, refresh suggested
questions on mention changes, and strengthen MCP must_use so @selected
services are invoked before local KB search.
- Updated the agent system prompt to clarify mandatory deep read requirements for FAQ entries, ensuring accurate retrieval and response generation.
- Added new localization strings for FAQ-related labels in English and Chinese, improving user experience and consistency across the application.
- Refactored the document and grep results components to support distinct FAQ identification, enhancing clarity in the output.
- Introduced new tests for FAQ metadata handling, ensuring robust functionality and correctness in the FAQ processing logic.
- Eliminated all instances of the final_answer tool from the agent's configuration and logic, transitioning to a model where the agent concludes responses with plain text instead.
- Updated event handling and response analysis to reflect this change, ensuring that the agent's output is now directly delivered as text without relying on a dedicated final answer tool.
- Adjusted localization strings across multiple languages to remove references to final_answer, enhancing clarity and consistency in user-facing messages.
- Improved the overall structure of the agent's response handling to streamline the process and reduce potential confusion regarding tool usage.
refactor(agent): streamline answer handling and preamble logic
- Updated the agent's response handling to treat any text streamed before a tool call as a preamble, ensuring it is not included in the final answer.
- Removed the superseded marker logic from the event emission process, simplifying the flow of answer content and enhancing clarity in the UI.
- Adjusted the event handling to ensure that preamble segments are correctly retracted and relocated into the steps tree, improving the overall user experience.
- Enhanced the agent's internal logic to better manage answer segments and their states during streaming, leading to a more coherent response structure.
Allow built-in models to be declared in config/builtin_models.yaml
instead of inserting rows via SQL. On every startup the file is read
and each entry is UPSERT-ed into the models table (is_builtin=true)
by stable id.
Any string field may reference an environment variable with ${NAME}.
Unset variables are left as the literal placeholder so
misconfiguration surfaces clearly in provider calls rather than
failing silently with an empty token.
The file is optional: missing file, parse errors, and per-entry
upsert failures all log a warning without aborting startup.
docker-compose.yml adds env_file (.env, required:false) so
deployment-specific variables are passed through automatically.
When a custom agent has `MultiTurnEnabled=false`, `applyAgentOverridesToChatManage`
sets `chatManage.MaxRounds = 0` to signal "no history". Two pipeline plugins
mistreated this zero value as "use the global default" and silently re-loaded
session history into the LLM context:
- `PluginLoadHistory` fell back to `Conversation.MaxRounds` when
`chatManage.MaxRounds == 0`.
- `PluginQueryUnderstand.loadHistory` had the same fallback, and even when
`LOAD_HISTORY` was skipped it would re-populate `chatManage.History`,
leaking previous turns into rewrite, image analysis, and the final answer.
The RAG branch in `session_knowledge_qa` also added `LOAD_HISTORY`
unconditionally, unlike the pure-chat branch which guarded it with `hasHistory`.
This change:
- Treats `chatManage.MaxRounds <= 0` as an explicit disable in both plugins;
no fallback to global config.
- Makes the RAG pipeline consistent with the pure-chat path by gating
`LOAD_HISTORY` on `hasHistory`.
- Removes the duplicated `Current Time: {{current_time}}` line from
`agent_system_prompt.yaml`. The agent already receives a fresh
`<runtime_context><current_time>` block with each turn from
`observe.buildRuntimeContextBlock`, so the static placeholder was
redundant.
The ReAct agent path (`session_agent_qa`) already checked `MultiTurnEnabled`
directly and is not affected.
Closes#1479
Documents whose only payload is an embedded image (e.g. a docx with a
single picture) intermittently produced the refusal line "No textual
content was extractable from this document." even though the vision
model had successfully extracted a caption.
Three coordinated fixes:
- Clarify the summary prompt that text inside `<image_caption>` and
`<image_ocr>` is first-class extracted content, not an image
reference, so the model only triggers the empty-content branch when
the body is genuinely textless.
- For image-dominated documents (real text < 200 runes after stripping
image markup) include OCR alongside captions so screenshots and
scanned figures contribute their actual content; text-heavy
documents continue to use caption-only enrichment to avoid OCR
noise from incidental figures.
- Add `EnrichContentCaptionAndOCR` which embeds caption + OCR text
inline next to the original Markdown image link, deliberately
omitting the `<image url=...>` and `<image_original>` wrapper
blocks. Those wrappers carry only opaque export hashes that consume
tokens and have been observed to retrigger the LLM's "image
reference with no extracted text" heuristic.
The grep_chunks tool previously accepted an array of regex queries (1-5)
and an optional knowledge_base_ids filter and limit. In practice the LLM
either fired multiple near-duplicate calls or split synonyms across
entries instead of using POSIX alternation, and KB scoping plus result
limit are server-side concerns the model should not control.
Reshape the contract to match `grep -E -i` semantics:
- Schema accepts a single required `query` string. Combine concepts with
`|` alternation in one regex instead of multiple calls.
- Drop `knowledge_base_ids` and `limit` from the schema; the tool now
always searches the full agent scope and uses a fixed internal cap.
- Legacy `pattern`, `queries`, `patterns`, `max_results` keys are still
accepted and joined into a single alternation regex so older callers
and in-flight model outputs keep working.
- Update the agent system prompt template to document the new single
`query` field.
- Frontend tool title now reads `query`/`queries`/`pattern`/`patterns`
in that order so the search text is shown again under the new schema.
- Add a dedicated `grepSearch` / `grepSearchFailed` tool status (zh-CN,
en-US, ko-KR, ru-RU) and rename the zh-CN tool label to "搜索关键词"
so the UI no longer prefixes the call with a generic "调用 ..." label.
Scanned PDFs without a successful OCR pass (e.g. when no VLM is configured
on the KB, or when the VLM returned a "no text content" reply that the
sanitizer discarded) reach the summary and wiki-ingest pipelines with
nothing but markdown image references for content. The downstream LLM
prompts used to receive the filename, file type, and title alongside that
empty content, and the generate_summary template explicitly forbade the
model from refusing — so the model would invent a topic from the filename
alone, producing e.g. a "Canon scanner manual" summary for an MX5280.pdf
that is actually a scanned legal letter.
Two changes shut this down:
1. Drop filename / title / file-type from every LLM prompt input. The
wiki summary, knowledge extract, candidate slug, and chunk citation
prompts now see only the document content. The knowledge summary
path no longer prepends the "Document Type: X / File Name: Y" intro.
2. Add a content-sufficiency guard in front of the LLM call. A new
helper strips markdown image references and HTML image tags and
refuses to invoke the model when fewer than 50 real-text runes
remain. getSummary returns a sentinel error in that case and the
caller marks the knowledge as SummaryStatusFailed instead of
surfacing the first chunk's bare image reference. Wiki ingest skips
LLM extraction with a clear log message.
The generate_summary.yaml prompt is also updated: the "NEVER output
refusal phrases" clause is removed and replaced with an explicit
empty-content rule that returns a deterministic placeholder line.
Real OCR (Tesseract / PaddleOCR) and filename-pattern sanitisation are
intentionally left for separate PRs.
When users ask broad queries like "请整理知识库中的数据" in RAG mode,
vector/keyword search returns nothing because the query has no specific
content to match. The user needs to see what documents exist, not search
results.
- Add buildKBDocumentListing() to fetch document titles/filenames from
the knowledge base and inject them into the fallback prompt via
{{kb_documents}} placeholder
- Update fallback prompt templates (model_fallback, default_fallback_prompt)
to include document listing context so the LLM can guide users
- Improve rewrite prompt: tighten summarize vs kb_search classification,
expand kb_search to cover browse/organize/list operations, add examples
Closes#959
Revised the agent system prompt to clarify the synthesis process and the handling of factual errors. The instructions now emphasize the mandatory use of the `final_answer` tool for submitting responses and the need to flag issues before calling `final_answer`. This enhances the clarity and consistency of the agent's operational guidelines.
- Introduced a unified capability requirements system for agent tools, ensuring consistent filtering of knowledge bases based on their capabilities.
- Implemented derived `kb_filter` logic to streamline compatibility checks between agent tools and knowledge bases, reducing redundancy in configuration.
- Updated frontend components to reflect new filtering logic, including improved handling of empty states and user feedback when no compatible knowledge bases are available.
- Enhanced internationalization support by adding relevant messages for tool compatibility issues across multiple languages.
- Changed `kb_selection_mode` in agent type presets from "selected" to "all" to allow agents to access all knowledge bases by default.
- Revised prompt templates to clarify the handling of bound knowledge bases, replacing the `{{knowledge_bases}}` placeholder with a reference to the `<bound_knowledge_bases>` block in the user message's `<runtime_context>`.
- Enhanced internal logic to ensure that only searchable knowledge bases are considered during retrieval, improving the efficiency of knowledge searches.
- Added new functions for better formatting and handling of knowledge base metadata in the runtime context.
- Updated documentation and comments for clarity on the changes made to knowledge base interactions.