* feat(models): expose embedding input modalities
Whether an embedding model maps images into the same space as text is
what will decide if a knowledge base can store image vectors. The
catalog already had the field, but most multimodal embedding models
were missing it and embedding rows never reported it.
- Catalog: declare input modalities for the Doubao and Tongyi vision
embeddings, the Aliyun multimodal / qwen-vl embedding families and
jina-embeddings-v4, as listed in each vendor's documentation. The
migration baseline changes for exactly these nine entries.
- Model responses resolve capabilities for embedding rows too, so
capabilities.input reflects the catalog merged with spec.input.
- Model editor: an "Image input" switch for embedding models. It follows
the catalog and writes spec.input only when the row differs from it;
picking another catalog model or switching vendor drops the
declaration. Model cards show the image badge for such models.
* feat(embedding): embed images with multimodal embedding models
Adds the image side of an embedder, for models that map images into the
same space as their text vectors so a text query can find an image.
- api: EmbedImage (bytes + MIME type, sent inline as base64 because the
object store is usually unreachable from the vendor) and an
ImageEmbedder implemented by all four protocols, each in the image
shape its reference documents: Ark image_url, DashScope contents[].image,
Gemini inlineData parts, and on the OpenAI shape an object keyed by the
vendor's image_field (Jina: "image"). A plain OpenAI-compatible
endpoint names no image field and takes no images.
- Settings: image_field, max_image_batch_size (default one per request),
max_image_bytes and image_mime_types, filled from vendor docs for the
DashScope multimodal models, gemini-embedding-2 and Jina.
- embedding: an optional ImageEmbedder interface reached through
AsImageEmbedder, so existing Embedder implementations and fakes are
untouched. An embedder takes images only when the catalog (or the
row's spec.input) lists image AND the endpoint can carry one. Images
are checked against the documented limits before any request, then
split at the documented images per request. The concurrency, debug
and Langfuse decorators forward it; logs and traces carry the MIME
type and size, never the bytes.
The migration baseline changes only for the entries whose resolved
embedding settings now carry image limits.
* feat(knowledge): index images by their own vector and recall them
When the knowledge base's embedding model embeds images, each image now
gets a vector of its own next to its caption and OCR chunks, so a query
can find an image by what it shows even where the caption left it out.
Ingestion
- After the caption/OCR chunks are indexed, the multimodal task embeds
the image (embedding.PrepareImage converts and shrinks it to the
vendor's documented limits) and stores it as a new image_vector chunk:
Content is the caption (OCR text when there is none) for reranking and
the answer context; its index row carries the image's vector with
SourceType ImageSourceType. Scanned PDF pages are skipped, their OCR
text serves better. A failure here is logged and never fails the task.
- The image vector gets its own chunk rather than a second index row on
the caption chunk: OpenSearch and Weaviate key documents by ChunkID,
so a second row would overwrite the caption's text vector, and copy,
move and enable/disable all follow chunk IDs.
- CompositeRetrieveEngine.BatchIndexVectors writes precomputed vectors
through each engine's normal BatchIndex, keeping row order (the
composite's de-duplication does not).
- Paths that rebuild vectors from text leave image_vector alone:
syncChunkIndex only syncs its enabled flag, updateChunkVector skips
it, and a caption edit updates its Content only.
- The multimodal task's caption/OCR index rows now set IsEnabled; only
backends with a column default were storing them as enabled.
Retrieval
- A text query scores lower against an image than against text, so when
the model embeds images the document vector search is widened by half
(capped at 500) and run at min(VectorThreshold, 0.1); each store
group's hits are then held to their own threshold before score
normalization: text at VectorThreshold, images at min(VectorThreshold,
0.1). Keyword hits on image rows are dropped, their Content being the
already-indexed caption.
- image_vector hits are searchable and merge like caption hits: parent
context is attached, and the caption is not repeated from ImageInfo.
Adds golang.org/x/image for scaling and WebP/BMP decoding.
* fix(models): report no thinking levels for embedding, rerank and ASR models
Embedding rows now carry capabilities, and the end-to-end run showed them
listing off/auto/minimal/low/medium/high: these references resolve no chat
protocol, and an empty level map reads as "every level supported". A model
without a chat protocol does not think, so its capabilities say so.
* fix(compaction): never read a cancelled turn's stream as a summary
streamSummary selects on the turn's context and the provider stream at
once. When the turn is already cancelled both are ready (providers close
the stream once their context ends), and select picks at random, so a
stopped turn could come back as a completed, partial summary with a nil
error. TestSummarizationIsNotRetriedOnceTheTurnIsCancelled caught it
intermittently in CI and reproduces under -race with several -cpu values.
Check the context before each receive and again when the stream closes.
* feat(knowledge): put image vectors behind a knowledge-base switch, off by default
Image vector indexing and recall used to turn on whenever the knowledge
base's embedding model declared image input. A model that takes images is
often chosen for text alone, so an upgrade would have added an embedding
call per image, widened vector recall 1.5x and lowered the image threshold
to 0.1 for those knowledge bases without anyone asking.
ImageProcessingConfig gains image_vector_enabled (default false; stored
JSON without the key reads as off, so existing knowledge bases stay off
with no migration). Both sides now need the switch and the model:
- indexImageVector returns before touching the model unless the KB opted
in (KnowledgeBase.IsImageVectorEnabled, which also needs vector indexing).
- applyImageRecall resolves the embedding model only when some KB in the
search opted in, and records the opted-in KBs per store group
(storeGroup.ImageKBIDs). Only those groups widen the pool and lower the
threshold; image hits of any other KB, including ones indexed before the
switch was turned off, are dropped.
The KB editor shows the switch in the image processing section, with a
warning when the selected embedding model does not declare image input;
docs describe the switch, its cost and what turning it off does.
* fix(knowledge): leave image vectors behind when cloning into a KB that is off
CloneChunk copied every chunk of the source document and then CopyIndices
for all of them, so copying a document from a knowledge base with image
vectors on into one with the switch off (same model and store) still
created image_vector chunks and their index rows, bypassing the switch
that indexImageVector honours.
Image vector chunks are now dropped from the copy when the target has not
opted in. Every engine's CopyIndices copies only the rows of chunks in the
mapping, so their index rows stay behind too.
* fix(search): refill the vector pool past stale image vectors
Image hits of a knowledge base whose switch is off were dropped only after
the engine had ranked and cut its results at TopK. Image vectors indexed
before the switch was turned off could fill that pool and push out valid
text hits: with MatchCount 5 the pool is 50, and 50 stale image rows
ranked above the text left the search empty.
No engine can filter rows by source type (Postgres and SQLite do not even
take ExcludeChunkIDs), so the exclusion cannot move into the query. When
the document vector pool comes back full and holds stale image rows, that
search alone runs again on a pool twice as large, until it holds TopK rows
that are not stale, the index runs out, or the pool reaches
maxRetrievalPoolSize (at most four extra calls from a pool of 50). The pool
is then cut back to TopK such rows, matching an engine-side exclusion
unless more than maxRetrievalPoolSize-TopK stale rows rank above the text.
* feat(i18n): translate the image vector switch and image input keys into zh-TW
The zh-TW locale arrived from main after this branch added five keys: the
knowledge base's image vector switch (label, description, unsupported-model
warning) and the embedding model's image input flag (label, description).
* fix(knowledge): leave image vectors behind when moving into a KB that is off
A reuse_vectors move relocates a document's chunks and index rows in place:
MoveKnowledgeIndices rebinds the rows by document ID and
MoveChunksByKnowledgeID rebinds the chunks. Moving a document from a
knowledge base with image vectors on into one with the switch off therefore
carried its image_vector chunks and their vectors along, the same bypass the
clone path had.
When the target has not opted in, the image_vector chunks are now deleted
during the move, their index rows first and then the chunks, so a retry
finds the chunks again rather than leaving orphaned rows. A target that
opted in still receives them unchanged.
* fix(retriever): keep image vectors out of keyword-only engines
BatchIndexVectors handed image rows to every engine of the composite,
including one that serves keyword retrieval alone (Elasticsearch v7 next to
a vector store). There the row holds nothing but the image's caption, which
the caption's own chunk already indexes, and retrieval discards every
keyword hit on it.
Image rows now go only to engines that serve vector retrieval. An engine
that serves both keeps one row for both kinds, so its keyword search still
meets them; the search side handles that.
* fix(search): refill the keyword pool past image rows
Every engine that serves both kinds of retrieval stores one row per index
entry and runs its keyword search over that row's content, so an image
vector row, whose content is its caption, is a keyword hit wherever the
caption's own chunk is. filterImageHits drops all of them, but only after
the engine has cut its ranking at TopK, so they crowded out text and caption
hits; a keyword ranking that put the image rows first came back empty.
The keyword pool is now refilled the way the vector pool is refilled past
stale image vectors: when it came back full and held image rows, the
keyword search runs again on a pool twice as large, up to
maxRetrievalPoolSize, and is cut back to TopK hits that are kept.
Leaving the rows out of the keyword index instead would mean writing them
without searchable content (Milvus and Tencent VectorDB derive a BM25
sparse vector from it, and how they take an empty one is unverified) or a
source-type filter in every engine's keyword query (ten engines, and
filterable fields that existing collections lack).
* revert: let image vectors move with the document into a KB that is off
This reverts commit e259977fe915f7d0bc30c4b02bc7af04da1c8dd4.
Deleting the image vector index rows by chunk ID before the move is not
reliable: the Weaviate and Elasticsearch v8 delete wrappers ignore partial
failures, and OpenSearch rejects more than 1000 IDs with ErrBatchTooLarge.
A partial delete followed by the chunk delete leaves index rows whose chunks
are gone, which nothing ever cleans up.
Without it, the image_vector chunks and their index rows move together
with the document. While the target's switch is off they are filtered at
query time (the pool refill keeps them from crowding out text), they are
removed with the document on delete, reparse or knowledge base delete, both
chunks and index rows by knowledge ID, and they become active again if the
target turns the switch on. The clone path keeps dropping them, since there
the rows are copied rather than moved.
* fix(search): keep the image row refill from losing hits it already had
refillPool replaced a pool with the answer to each larger search. A refill
that answered only in part (RetrieveResult.Error, swallowed into a warning
by retainPartialResults, as when some collections of the store did not
answer) could hold fewer text rows than the complete smaller pool, or none:
40 stale image rows above 10 text came back from a partial refill as the
40 image rows alone, and the search lost all 10 text hits. A refill that
failed outright failed the whole store group, discarding a first round that
had succeeded.
A refill is now best effort. A failed or incomplete refill search stops
the refill and leaves the last complete pool in place, cut back to TopK
kept rows as before, so a refill can no longer leave a search worse off
than it was without it. Merging the partial answer into the pool was not
done: its rows would need deduplicating and its ranking is cut at a
different TopK over a different set of collections, while keeping the
previous pool is exact.
* docs(knowledge): say what happens to image vectors in a KB that is off and what the refill costs
The image vector section now says that with the switch off, image vectors
that were indexed before it was turned off, or that came along when the
document was moved in from a KB that has it on, are kept but filtered at
query time, are removed when the document is deleted or reparsed or the
KB is deleted, and are recalled again once the switch is turned on; a copy
into such a KB leaves them behind.
It also states the cost of keeping dropped image rows from crowding text
out of the pools: the vector and keyword pools are searched again with
TopK doubled up to the 500 candidate cap, which adds queries for store
groups that hold such rows, and text can still be missed when more than
about 500 candidates are image rows. A failed or partial refill keeps the
previous complete pool.
* fix(search): log unexpected refill result sets, and describe the refill stop condition precisely
Assistant history messages that carry tool_calls but empty text previously
omitted the content field entirely on the wire. Strict local OpenAI-compatible
engines (vLLM / Xinference serving Qwen-family Jinja chat templates) treat the
missing field as undefined/null and crash while rendering the next turn,
surfacing as "I'm sorry, I was unable to generate a response".
Always serialize "content": "" for assistant tool-call turns. The empty
string is accepted by every standard OpenAI-compatible endpoint.
Fixes#1487
Co-authored-by: LuckVd <LuckVd@users.noreply.github.com>
* fix(chat_pipeline): treat non-JSON model refusals as skipped graph extraction
When a chunk carries nothing extractable (e.g. a table-of-contents page),
the LLM correctly answers with a prose refusal such as "抱歉…". parseOutput
then fed that prose to json.Unmarshal, failing with a confusing
"invalid character 'æ'" (the first byte of a CJK character), marking the
chunk task GRAPH_EXTRACT_FAILED and letting asynq replay the identical
doomed LLM call up to the retry limit.
Classify responses that contain no JSON structure at all (no '{' or '[')
as errModelDeclined and complete the extraction with an empty graph so the
task succeeds as an explicit skip. Malformed or truncated JSON (structure
present) still fails and retries as before.
Fixes#3600
* fix(chat_pipeline): keep empty model responses retriable in graph extraction
An empty response is usually transient (gateway returning 200 with no
content, a reasoning model spending its whole budget on reasoning,
content filtering), and a retry often succeeds. Only non-empty prose
with no JSON structure is treated as a declined extraction.
* fix(graph): skip prose refusals only in the async chunk task, keep truncation retriable
ParseGraph no longer turns a prose answer into an empty graph. It returns
an error wrapping ErrModelDeclined, and ChunkExtractService.Handle is the
only caller that treats it as a skip (span output skipped=model_declined).
The /initialization "try extract" endpoint and query-side entity extraction
report the refusal as a failure again, so the graph settings form no longer
gets its few-shot examples replaced with an empty result.
Extractor.Extract drops the ErrModelDeclined marker when finish_reason is
"length": prose with no JSON because the output budget ran out is
truncation, and asynq should retry it.
* fix(graph): retry every truncation spelling, not only finish_reason=length
Extract kept a prose answer retriable only when FinishReason was exactly
"length". max_tokens / max_output_tokens (and case/space variants that
IsLengthFinishReason already accepts), a stream that broke before the stop
event ("incomplete") and an adapter that reports no reason were still read
as a refusal, so the async chunk task skipped the chunk for good.
Only a reason that proves the model stopped on its own now counts as a
refusal. Ollama's non-streaming Chat also forwards done_reason, so a
num_predict cut-off is reported as "length" instead of an empty string.
* fix(graph): only treat a response with no JSON anywhere as a refusal
The refusal check looked at the first extracted fence, so a prose fence
followed by a JSON fence was skipped as a refusal instead of failing as a
retriable parse error.
---------
Co-authored-by: LuckVd <LuckVd@users.noreply.github.com>
Co-authored-by: wizardchen <wizardchen@tencent.com>
* feat(rerank): support Hugging Face TEI protocol
* style(rerank): satisfy TEI lint checks
* fix(rerank): cap Hugging Face TEI batches at 32 documents
TEI rejects requests above --max-client-batch-size, which defaults to 32.
Without a declared ceiling the shared batching layer sends the whole
candidate set in one request, so a default TEI deployment fails rerank
whenever hybrid retrieval returns more than 32 candidates.
* fix(rerank): send truncate and merge extra_body in the TEI client
TEI before 1.9 starts with --auto-truncate=false and rejects the whole
/rerank call when one passage is longer than the model's input limit, so
retrieval silently falls back to the original order. Ask for truncation
explicitly, as the NIM client does with truncate=END.
The TEI client also ignored compat extra_body, unlike the Cohere, NIM and
DashScope clients. Merge it the same way: query, texts and raw_scores stay
owned by the client, truncate is a default that extra_body may override.
---------
Co-authored-by: bigdu332 <bigdu332@users.noreply.github.com>
Co-authored-by: wizardchen <wizardchen@tencent.com>
* fix(frontend): 修复复制模型时模型名称被改写的问题
* fix(model): preserve model name and credentials when copying a config
Copying rewrote the upstream model name and posted redacted parameters, so the duplicate could not authenticate. Clone the stored row, cap the display name, and skip a new download when the local model is already active.
* fix(lint): clear new golangci-lint findings on model copy
Check c.Error results, wrap the long test line, and format the stub. The pull request lint range also includes the new HTTP error test, whose unused request parameter failed revive.
---------
Co-authored-by: wizardchen <wizardchen@tencent.com>
Endpoint.Do turns a non-2xx reply into &HTTPError{StatusCode, Body} and
drops the response headers. A caller that retries a 429 or 503 therefore
cannot see the provider's Retry-After and falls back to its own guess: the
agent retry loop waits 1s, 2s, 3s, which against a provider asking for, say,
20s spends every attempt inside the rate-limit window and then fails the
turn.
- HTTPError carries the reply's Header, and RetryAfter() parses the header
in both forms (delay-seconds and HTTP-date); absent, malformed or past
values give zero.
- The agent retry loop waits max(linear backoff, Retry-After), with the
vendor's value capped at 30s so a large Retry-After cannot hold a turn
indefinitely.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Reverts the eleven plugin system PRs merged on 2026-09-26 so they can be
reviewed before landing:
* a829d7cc9 agent tools served by plugins (#3733)
* 4f5d475c3 Jira example plugin (#3730)
* 01222aac8 dynamic form options, OAuth connections and resource limits (#3729)
* cf502ba46 event delivery and webhooks (#3728)
* 996696057 sandboxed plugin pages (#3727)
* da4ef68d9 standalone plugin hosts (#3725)
* 728007ada remote runtime, Python SDK (#3724)
* ff49f9491 Host API and document parsers (#3721)
* 4328ab26a code plugins (#3720)
* ed63e35cc declarative plugins (#3716)
* c91a8d11f plugin system foundation (#3711)
The tree is back to 47773b5bd, migrations included: no environment ran
the plugin migrations (versioned 000115-000119, sqlite 000034-000038).
* feat(plugin): install declarative plugin packages across nodes
* feat(plugin): serve installed plugins' MCP servers in each workspace
* feat(plugin): let installed plugins contribute skills and model vendors
* feat(plugin): register plugin skills into the workspace skill catalog
* feat(frontend): manage installed plugins and their configuration
* test(weaviate): read the class state before the probe barrier
* fix(plugin): keep MCP tool directories reachable behind the plugin wrapper
* test(plugin): wrap a long line
* feat(plugin): add plugin manifest, registry and builtin catalog
* feat(plugin): add config schemas with unified secret handling
* feat(frontend): render config schemas with SchemaForm
* feat(datasource): serve connector credential forms from the backend
* feat(im): describe IM platforms and credentials from the backend
* feat(frontend): render model vendor extra fields from the config schema
* feat(frontend): build the settings modal from an extension registry
* refactor(agent): register builtin tools and their chat views
* feat(plugin): let workspaces enable and disable plugins
* feat(plugin): add the runtime driver interface and gather builtins
* fix(models): stop advertising graded thinking levels the wire format drops
A generic-provider row (relays, self-hosted runtimes — the default for
uncatalogued models) resolves an empty ThinkingLevelMap, whose missing-key
semantics report every graded rung as supported. Capabilities() then hands
minimal/low/medium/high to the frontend, which renders a selector whose
choice applyThinking silently discards: the chat-template-kwargs branch
sends only the enable_thinking boolean and no reasoning_effort is computed.
Same defect class as the two NVIDIA NIM entries fixed in #3489, but there
is no catalog entry to repair here.
Collapse Capabilities() to the levels the wire can actually express for a
boolean-only switch (off/auto) unless the settings opt into
supports_reasoning_effort. The vendor map itself is untouched, so vendors
that do carry effort keep their ladders.
* fix(models): carry reasoning_effort through the chat-template switch when opted in
The Capabilities() guard kept a vendor's ladder when it opts into
supports_reasoning_effort, but applyThinking's chat-template-kwargs branch
wrote only the enable_thinking boolean and dropped the already-computed
effort — so the opt-in combination still sent byte-identical bodies for
every graded rung (#3551's defect, alive in exactly the shape the review
flagged). Mirror the enable_thinking dialect: write reasoning_effort on
top of the switch when one was computed. A golden case pins the opt-in
combination; the no-opt-in guard case keeps asserting the level is
dropped.
* style(models): wrap long Resolve call in thinking-levels test
---------
Co-authored-by: wizardchen <wizardchen@tencent.com>
Co-authored-by: lyingbug <xiayula2014@gmail.com>
A stream body that ran out without a finish reason (a proxy cutting the
connection) was closed as a clean stop on the Completions, Responses and
Gemini paths, so a half-written answer was stored as complete. It is now
marked incomplete, as the Anthropic path already did, and both the chat
pipeline and the agent treat an incomplete answer as a stream error.
Retry classification checks typed HTTP/transport errors first and matches
status codes as whole numbers, so "max_tokens 5000" no longer reads as a
500. The retry backoff returns as soon as the turn is cancelled.
* 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.
* feat(models): add system admin model catalog console
System administrators can now maintain the model catalog from
Settings → System administration → Model catalog instead of editing
models.json on every replica. Administrator changes are stored as a
versioned overlay in the database (new model_catalog_config table),
layered between the deployment file and explicit per-model settings,
and synchronised to other instances within about 5 seconds.
Backend
- GET/POST preview/PUT /system/admin/model-catalog with optimistic
version + deployment baseline checks (409 on conflict), 20-version
history and audit logging.
- Console overlays reject credentials, headers, env interpolation and
non-inline icons; those stay deployment-owned.
- Catalog views include each chat model's resolved thinking levels and
the vendor defaults, so the console never re-implements level
inheritance.
Frontend
- Model list in the shared data-table style with provider/type filters,
"modified only" filter and source tags (built-in / deployment / admin).
- Edit drawer saves and applies immediately: display name, context
window, max output, input modalities, reasoning, thinking levels,
embedding dimension and picker visibility, each with its inherited
default and a layer comparison table.
- Add-model drawer, JSON editor with server-side change check before
publishing, and version history with one-step restore.
- Model settings gets a system-admin shortcut to the catalog; model
editor reloads providers when opened so published catalogs apply.
* fix(models): harden the admin model catalog console
- Reject console base_url, base_urls, url_patterns and auth: moving a
vendor's default URL would send its deployment api_key (and keys of rows
without their own base URL) to another host. With every interpolated
field now deployment-owned, '$' in names and descriptions is plain text.
- Drop deployment base URLs from catalog views; they may carry
interpolated tokens.
- Remember a stored version this replica cannot compile and skip it until
the version changes, instead of re-reading and recompiling every poll.
- Resolve vendor default thinking levels with reasoning forced on, like
per-model levels.
- Compute builtin/deployment views once; preview returns only the
candidate catalog and normalized overlay.
- Share deployment file lookup via runtime.ReadDeploymentOverlay.
- The model editor force-refreshes its own type instead of resetting the
provider store for every type; the refreshed list wins in the id index.
The catalog rebuild marked qwen3-rerank unsupported because Alibaba also
documents a flat /compatible-api/v1/reranks shape for it. The native
text-rerank endpoint serves the same model with identical scores (#3558),
and v0.8.0 routed it there, so existing Aliyun rerank rows naming it
failed with "failed to get rerank model" after upgrading.
Drop the unsupported override and the deprecated flag so the model
resolves to dashscope-rerank again and shows up in the picker.
* fix(embedding): stop four embedders from returning (nil, nil) on a dead upstream
doRequestWithRetry keeps resp and err in the enclosing scope and returns
(nil, err) once the retries run out. Writing `req, err := ...` inside the
loop introduces a loop-local err, so `resp, err = client.Do(req)` updates
that copy and the enclosing err stays nil. Every retry failing therefore
returned (nil, nil): BatchEmbed's `if err != nil` did not fire and the next
line dereferenced resp.Body, a nil-pointer panic that took the process down
instead of surfacing "connection refused". Stopping the embedding upstream
and asking any RAG question reproduced it.
The bug was found, documented and fixed in openai.go, then re-fixed in
zhipu.go and again, differently, in azure_openai.go; jina.go, nvidia.go,
aliyun.go and volcengine.go kept the original spelling because the retry
loop exists in eight near-identical copies and a fix to one reaches none of
the others.
TestEmbeddersReportUnreachableUpstream now drives all eight against a
refused port with the retry budget zeroed, so the next copy of this loop is
covered the moment it appears.
* fix(embedding): read the documented index field of DashScope multimodal vectors
The multimodal embedding response carries each vector's position in `index`:
{"output": {"embeddings": [{"index": 0, "embedding": [...], "type": "text"},
{"index": 1, "embedding": [...], "type": "image"}]}}
The decoder read `text_index`, which is the field name of DashScope's *text*
embedding API one page over. A missing field decodes as 0, so every vector in
a batch was written to slot 0 and the rest of the slice stayed empty: with the
default BATCH_EMBED_SIZE of 5, four of every five chunks were indexed with an
empty vector. Only models routed to this client are affected — an Aliyun
embedding model whose name contains "vision" or "multimodal"; the text-only
models go through the OpenAI-compatible path.
An index outside the batch, or a response that covers fewer inputs than were
sent, is now an error. Handing back an empty vector stores it and poisons
retrieval without anything failing.
The regression test replays the response example from the reference page, so
the decoder is pinned against the vendor's own document rather than against
our structs.
https://help.aliyun.com/zh/model-studio/multimodal-embedding-api-reference
* feat(agent): persist compaction checkpoints across turns
Agent compaction only lived inside the turn that ran it: the summary was
discarded at turn end and LoadAgentHistory rebuilt raw history from the
messages table, so every later turn past the threshold paid for another
summarization of nearly the same history.
A compaction whose history summary ends exactly on a stored turn is now
written onto that turn's assistant message (messages.context_checkpoint).
LoadAgentHistory starts from the newest checkpoint: the summary replaces
that turn and everything before it, and later turns are replayed verbatim.
The compactor already treats a leading summary as the previous summary, so
checkpoints fold forward incrementally.
History messages carry their turn's assistant message ID (chat.Message
TurnID, engine-internal) so compaction can tell whether its cut lands on a
stored-turn boundary. Nothing is persisted when the cut ends inside a stored
turn, covers only the live turn, or the history summary degraded to a raw
archive; a split turn's prefix summary is never persisted because that turn
is replayed whole next time. Writes are column-scoped and best-effort, and a
failed lookup falls back to history without a checkpoint.
Storing the checkpoint on the covered turn keeps the boundary implicit, so
session forks copy it with the turn and deleting the turn drops it. The
HistoryTurns cap still applies, now to the turns after the checkpoint.
* test(skill): wait for the background install in the name-conflict test
InstallSkill queues the install on a goroutine, and the fixture row already
starts at installing, so reading the status right after the call raced the
mock install and failed whenever it landed on ready first. Wait for the run
to finish on the reused row instead, as the other install tests do.
* feat(agent): size agent history by tokens instead of turn count
Agent history was capped at history_turns (default 5) because every turn
used to rebuild raw history and compaction could not persist. With
compaction checkpoints persisted, the cap is no longer what keeps a session
inside the window, and it threw away context that would have fit.
LoadAgentHistory now loads the newest checkpoint summary plus the newest
turns after it that fit a token budget: the run's compaction threshold
(agent.HistoryTokenBudget, the window less the reply reserve). History
that large already triggers the first round's compaction, which persists a
new checkpoint, so compaction rather than a turn count bounds the session.
Loading more would only hand that compaction a summarizer input larger
than the model can read.
Rows are paged backwards by (created_at, id) and the read stops at the
checkpoint or once the budget is full, so a long session is never read
whole. Kept turns stay contiguous, the newest turn is kept even when it
alone exceeds the budget (compaction splits it), and a turn cut by the read
is left out rather than replayed without its original question.
Agent mode no longer reads history_turns: AgentConfig.HistoryTurns is
removed, the editor hides the input in smart-reasoning mode and explains
the automatic behavior, and the smart-reasoning builtins drop the setting.
Quick-answer (KnowledgeQA) keeps its turn cap since it has no compaction.
* Revert "feat(agent): size agent history by tokens instead of turn count"
This reverts commit d05fb6e5f1f161813a88e9127609f7a059f38f12.
Loading history up to the compaction threshold lets the loader drop the
oldest turns before compaction can trigger: whenever a single turn's
replay is larger than the system prompt plus the new question, the
trimmed history stays under the threshold, nothing is summarized, and the
session degrades into a large sliding window that never gets a
checkpoint. A turn-by-turn simulation of LoadAgentHistory with the real
compactor showed zero compactions over 60 turns in those cases.
Keep the persisted checkpoints with the HistoryTurns cap for now; token
budgeted history returns in a separate change with a budget that leaves
the overflow to compaction.
* feat(mcp)!: persist tool catalogs and add guided configuration
* fix(mcp): write caller-owned catalogs and require a real tool sync
OAuth viewers can now persist their own directory, and the setup wizard no longer treats an unsynced service as complete. Missing catalogs auto-fetch on the tools step; already-used MCP functions are restored across engine turns.
* fix(mcp): skip redundant list_servers and tighten catalog UI
When the advertised server directory already fits in discover_mcp_tools, models should list_tools or describe directly. Chat and settings now show server names, collapsed docs, and initialize debug metadata for empty optional fields.
* fix(mcp): surface catalog summaries and polish tool detail UI
List responses now include a saved catalog count without live MCP calls, so settings cards and @ mentions can show sync state. Tool details open in a popup instead of expanding the row.
* fix(mcp): drop incomplete Anthropic calls and gate shared catalog refresh
Unclosed streamed tool_use was emitted as {} and executed. list_tools refresh=true now re-lists upstream instead of returning the saved snapshot, and only Admin (or MCP-capable API keys) can overwrite tenant-wide static-auth catalogs.
* fix(mcp): wrap client refresh path to satisfy lll
The metadata refresh URL made doRequest exceed the 120-character limit in the client module CI lint.
* fix(chat): send one completion-token field per provider
Agent dual-wrote max_tokens and max_completion_tokens, which Volcengine Ark
rejects. Collapse them to one budget and pick the wire name the way Pi does.
Fixes#3014
* fix(chat): shrink max_tokens provider whitelist to documented hosts
Only DeepSeek, Zhipu, SiliconFlow, Moonshot, NVIDIA, and self-hosted
vLLM-class runtimes stay on the legacy field. Aliyun DashScope documents
max_completion_tokens for thinking models and must not be forced onto
max_tokens.
* fix(chat): send max_tokens for LKEAP completions
Tencent LKEAP documents max_tokens only. Keep Pi's default of
max_completion_tokens, and classify every AllProviders() host so a new
vendor cannot silently inherit the default.
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.
Large write/edit calls used to sit on a spinner until the full JSON
arrived. Emit path and running line counts from partial arguments so
the UI can update in place, without putting file bodies on the SSE wire.
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.
* feat(provider): add LiteLLM as an OpenAI-compatible provider
* docs: use revive-compliant comment for ProviderLiteLLM const
* fix(provider): stop defaulting LiteLLM to localhost under SSRF
The catalog URL localhost:4000 is blocked by default SSRF and does not
match DetectProvider, so UI/CLI could neither save nor auto-detect it.
Use a placeholder host, keep thinking aligned with generic, and sync
ModelSource across types, Swagger, and docs.
---------
Co-authored-by: wizardchen <wizardchen@tencent.com>
Accumulate token usage across every LLM round of an agent turn and
expose the aggregate everywhere a turn is observable:
- AgentState.TurnUsage accumulates each round's reported usage,
including the final-answer synthesis call; TokenUsage.Accumulate
keeps the cache counters as prompt subsets and recomputes the
combined cache status (unsupported survives only while every call
was itself unsupported).
- The agent completion event carries the aggregate, and the complete
stream event exposes it both as data.usage and as a typed
StreamEvent.Usage that buildStreamResponse promotes into the
previously-unused StreamResponse.Usage wire field.
- The assistant message persists it as a nullable usage column on
both migration targets (postgres 000085 jsonb, sqlite 000012 text),
for web and IM channels alike, so history reads still carry the
turn's cost after the live stream is gone.
- "[LLM Usage]" log lines gain ", session_id=..., principal=type:id"
when the context carries them; calls outside sessions stay
byte-identical.
Usage from failed retry attempts is not accumulated: the per-call
usage log remains the ground truth, while the turn aggregate serves
as the operational signal.
Persist extracted and explicit memories across sessions with topic
resolution, semantic recall, document affinity, consolidation, workspace
settings, and inbox/interests UI.
Reject out-of-range and duplicate response indexes before mapping them to batch inputs. Return an error when a requested input has no response item instead of silently passing an incomplete batch downstream.
Add focused coverage for malformed responses and valid out-of-order results.
Signed-off-by: PP1 <74917296+pengpengyi92@users.noreply.github.com>
* fix(security): harden outbound URL SSRF protection
* test(security): add SSRF validation tests for outbound URL caching
Introduce tests for outbound SSRF validation cache to ensure proper handling of URL validation, including cache reuse for the same origin, cache invalidation on whitelist changes, and expiration of cached entries. This enhances the robustness of SSRF protection mechanisms in the application.
* refactor: clean up whitespace in engine_factory.go and ssrf_outbound_cache.go
Removed unnecessary blank lines in engine_factory.go and adjusted variable declarations for consistency in ssrf_outbound_cache.go. This improves code readability without altering functionality.
The VLM path sent max_tokens and temperature, both rejected by reasoning
and GPT-5 models. go-openai's reasoning validator failed the request
client-side, so every OCR and caption call errored with "this model is
not supported MaxTokens, please use MaxCompletionTokens" and no image
chunk was ever created.
The chat path already solved this in shapeOpenAIReasoning (issue #1283);
the VLM path was never wired to it. Reuse the shared
provider.IsOpenAIReasoningOrGPT5Model predicate so both paths agree on
which models are affected. Migrating max_tokens alone is not enough --
the VLM default temperature of 0.1 is independently rejected.
Also report an empty completion with finish_reason=length as an error.
max_completion_tokens covers reasoning tokens, so an exhausted budget
returns empty content rather than an API error, which was previously
recorded as no_extracted_content and looked identical to an image with
no text.
Fixes#2537
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
If the embedding provider returns an empty slice of vectors (e.g. length 0)
due to api mismatch or empty input, BatchEmbed will return length 0.
Accessing embedding[i] directly in BatchEmbedWithPool results in an out of bounds panic.
Because the panic occurs inside a mu.Lock() block, the mutex is never unlocked,
leading to a permanent deadlock of the document processing queue.
This fix checks the returned embeddings slice length and returns an error
instead of panicking, releasing the mutex properly.
Unify request-local model-handle handling behind a single
modelcontext.Registry (source citations, durable resource handles, wiki
issue handles) with per-tool argument policies and unresolved-handle
rejection before tool execution.
Also fixes issues found in review:
- decode wiki_read_issue's iN issue_id argument (was never resolved)
- preserve stored wiki page aliases when wiki_write_page omits aliases
- validate http/https scheme before registering url fields as web sources
- gate tool-role MultiContent text behind the tool result policy
- dedup KB IDs in NewWikiScopesFromKBIDs to avoid false ambiguity
- remove dead codec facades (decodeJSON, EncodeMessagesWithPolicy,
UnresolvedToolHandles)
Normalize provider-specific cache counters into TokenUsage, propagate call
purpose and prefix fingerprints through context for logging and Langfuse,
and restructure wiki page-modify prompts with warmup/singleflight to improve
prefix cache hit rates during reduce batches.
Extract NewSSRFSafeTransport and NewSSRFSafeHTTPClientWithTransport so
embedding and rerank can reuse one connection pool without allocating a
throwaway transport per client. SSRF dial and redirect policies are
unchanged; each client still keeps its own timeout.
The generic OpenAI-compatible reranker hardcoded `truncate_prompt_tokens:
511` on every /rerank request. Providers that honor this vLLM-specific
field (e.g. SiliconFlow serving Qwen3-Reranker) keep only the last 511
tokens of the templated rerank prompt, cutting the instruction and query
off any document longer than ~511 tokens. Every relevance score then
collapses to near zero (~0.0067 instead of ~0.998), falls below the score
threshold, and knowledge-base search returns empty results. A direct curl
without the field returns the correct scores.
WeKnora already removed this exact hardcoded parameter from the embedding
path in #960; the rerank path was never fixed.
- Add `omitempty` to `truncate_prompt_tokens` and `additional_data` so
empty optional fields are not sent.
- Make truncation opt-in via `extra_config.truncate_prompt_tokens`
(validated > 0 at construction), preserving an escape hatch for
self-hosted vLLM deployments with small-context rerankers.
- Add regression tests reproducing the collapsed-score symptom and the
opt-in path.
Fixes#2143
Split oversized candidate sets into concurrent batches instead of failing when retrieval exceeds the VikingDB rerank per-request cap, and align ko/ru locale hints with zh-CN.
Introduce resource://<handle> as the stable application identity for stored
objects, decoupling API responses, persisted rich text, and LLM prompts from
physical storage paths. Adds a resource catalog service (register/resolve/bind,
revocable expiring access grants), a request-local alias registry that compacts
resource references while in the LLM context and restores them (including across
streaming chunks), plus router, IM, chat-pipeline, and multimodal integration.
Includes schema migrations (resources, resource_bindings, resource_access_grants)
for sqlite, and hardens URL/output security handling.