Reject invalid agent_source_tenant_id instead of silently falling back, fix a resolveAgent nil panic when a shared selector fails, and add regression tests for wiki write-tool filtering and agent resolution.
The engine registry is per-process, so a store's engine is missing
whenever this process did not create it. That happens three ways:
- Startup skips a store whose engine fails to construct and never
retries it, so that instance cannot serve the store again -- not
even after a restart, if the backend is slow to come up.
- A store registered on one instance is absent on every other until
they restart, because registration is not broadcast.
- Registration can fail on the instance handling the create, leaving
the store missing even there.
Each one surfaces as "vector store is currently unavailable" and is
cleared today only by restarting the process.
Add GetOrLoadByStoreID, which falls back to building the engine from the
store row when the lookup misses, and route the two bound-store lookups
through it. The method goes on RetrieveEngineRegistry rather than
StoreRegistry because that is the interface the factory functions
receive; the two declare GetByStoreID separately and neither embeds the
other.
Concurrent callers collapse onto one build. That build runs on a context
detached from the caller that started it: callers share the result, so
letting the first one's cancellation abort it would fail everyone else
waiting. A panic inside the build is contained, because singleflight
re-raises it on a goroutine the HTTP recovery middleware cannot reach
and the build calls third-party client constructors.
Publishing is guarded by a per-store generation sampled before the
build, so an engine that finishes after the entry was registered or
removed by someone else does not overwrite it and leave that engine
orphaned with its connections open. A failed build starts a cooldown:
collapsing only helps concurrent callers, so without it a backend that
stays down would cost a full build timeout on every sequential request.
Separate "the store is gone" from "its engine cannot be built right
now". Async delete handlers answer ErrVectorStoreNotFound with
asynq.SkipRetry, so reporting a database blip or a backend outage that
way would permanently discard knowledge-base and index delete tasks and
orphan the vector data they exist to remove. ErrVectorStoreUnavailable
covers the retryable cases; the repository already distinguishes them,
returning (nil, nil) for a missing row and an error for a failure.
For the same reason a caller's own cancellation is never folded into a
store sentinel. The ownership check runs first and queries the database
with that context, so on a shutdown it is usually where cancellation is
noticed.
A search resolves one engine per distinct store, sequentially, and this
server sets no HTTP read or write timeout, so the resolution loop gets a
budget above a single build. An exhausted budget still leaves the
engines warming, since the build is detached, and is reported as an
unavailable store because retrying is likely to succeed.
Build failures are logged where they happen, with the panic stack. The
cause cannot travel back to the caller -- factory errors embed endpoints
and credentials -- so without this the one thing an operator needs when
self-healing fails to heal would be discarded with the error.
Note a deliberate contract change: an ownership lookup failure used to
be reported as not-found, and a test asserted that on purpose. A
database failure says nothing about whether the store exists, and
treating it as permanent is what discards the task, so that assertion is
inverted rather than preserved.
Normalize Excel file types when resolving parser rules and applying the
xlsx_first_row_as_header override for both xlsx and xls. Skip the override
for non-builtin engines, unify legacy XLS parsing with the XLSX default path,
and extend tests for case handling and explicit false overrides.
Discard stale async refreshes without clobbering newer summaries, queue manual refreshes instead of blocking HTTP, sort edited chunks by index for summaries, escape document metadata in chat headers, and distinguish lookup failures from true staleness.
Co-authored-by: Cursor <cursoragent@cursor.com>
Rework the Feishu datasource connector from whole-document export to the docx
blocks API, so embedded content that export silently dropped is now ingested,
and add the multi-item ingestion infrastructure it needs.
Feishu connector:
- Embedded online sheets and bitables are read via their APIs and inlined as
Markdown tables — searchable text, instead of being lost or (post-export)
rendered as unsearchable images. Bitable date/datetime cells render in the
table's timezone (configurable via Settings["timezone"], default GMT+8) rather
than a raw epoch-ms integer; a datetime formatter keeps the time component.
- Uploaded attachments (>2KB, parseable types) become standalone knowledge
sub-items so their text is parsed and retrievable.
- Embedded images become standalone sub-items routed through the KB's VLM OCR
pipeline, gated by the KB's multimodal setting: the image's external_id always
stays in SubtreeKeep (toggling VLM off never sweeps a previously-OCR'd image)
but the bytes are only downloaded/ingested when multimodal is enabled.
- Sub-documents (child wiki nodes) sync as their own nodes.
- The three paginated Feishu loops (document blocks, bitable fields, bitable
records) break on an empty page so a malformed has_more response cannot spin
until the task deadline.
Datasource ingestion infrastructure:
- Multi-item subtree reconciliation: a source node fans out into a parent plus
"<parent>#<kind>#<token>" children (types.SubtreeChildID); on re-sync the
sweep deletes prior children absent from SubtreeKeep. A per-child keep-set
means a single failed attachment can't wipe its siblings, and dupIsSameNode
guards the sweep so a file_hash-only dedup collision against a different node
cannot delete this node's children.
- config.MultimodalEnabled mirrors the KB's VLM state to the connector per run
(never persisted); an image a non-VLM KB rejects is skipped, not failed.
- A composite partial expression index (text_pattern_ops) backs the sweep's
metadata external_id prefix scan.
Verified end-to-end against a live Feishu wiki: docx text, native tables,
embedded sheets/bitables (with timezone-correct dates), attachments, images
(VLM OCR), and sub-documents all sync, parse, and are retrievable.
Store superseded wiki page versions for diff and rollback, attribute edits
to user/agent/pipeline/revert sources, and expose revision APIs plus an
editor UI with optimistic locking on save.
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)
Weak LLMs routinely mangle high-entropy wiki slugs (especially UUID-based
summary/<id> slugs), producing dead cross-links and cross-item dedup
hallucinations. This tightens integrity end to end:
- Ingest dedup now groups each new item with only its own trigram
candidates and validates merges against that per-item scope, rejecting
cross-item mispairings; candidate top-K narrowed 20 -> 5.
- Editor path aliases out-link slugs behind low-entropy ref-N tokens and
translates them back after generation, removing the chance to mangle.
- Agent write tool normalizes/validates slugs, blocks manual creation of
system-owned summary pages, and auto-repairs dead [[slug]] links
(rewrite-only, never strips) via bigram/display-text resolution.
- llmresource registry aliases summary/<uuid> slugs the same way as
storage handles so they round-trip safely through chat.
Adds unit tests for slug validation, aliasing round-trip, dead-link
repair, and dedup rejection rules.
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.
Operators can clear all finally-failed tasks in one queue from the runtime
drawer, with audit logging and scoped API access. Also localize the
chat_attachment queue label in the runtime queue details table.
Prevent concurrent refresh-token rotation across instances, distinguish
expired-but-refreshable tokens from active authorization, and bind OAuth
popup polling to a per-attempt ID so stale tokens cannot satisfy new consent.
- Updated the listFAQEntries function to support multiple tag IDs for filtering.
- Modified the frontend components to accommodate the new tag_ids parameter.
- Enhanced the KbTagManageDrawer and FAQEntryManager components to reflect changes in tag handling.
- Improved API documentation to clarify the new tag filtering capabilities.
This commit enhances the flexibility of FAQ entry retrieval, allowing users to filter by multiple tags for more precise results.
- Updated the exportFAQEntries function to support exporting FAQ entries in both CSV and JSON formats.
- Modified the frontend to include options for selecting the export format.
- Enhanced localization files to reflect the new export options in English, Russian, and Chinese.
- Implemented new tests to validate the export functionality for both formats.
This commit improves the flexibility of FAQ exports, allowing users to choose their preferred format for better integration with other tools.
- Added new fields for tracking added, merged, and partially failed entries during FAQ imports.
- Updated the import result statistics to include detailed counts for added and merged entries.
- Enhanced the UI to display progress hints and improved import result summaries.
- Implemented a new method for saving chunks in a single transaction to improve performance.
This commit improves the visibility and accuracy of FAQ import results, providing better user feedback and tracking capabilities.
- Added new actions for FAQ import lifecycle: `faq.import_started`, `faq.import_completed`, and `faq.import_failed`.
- Updated activity details to include success, failure, and skipped counts for FAQ imports.
- Enhanced UI to display additional activity information, including filtering options and clearer summaries.
- Implemented tests to validate the new FAQ import outcomes and activity details.
This commit improves the tracking and visibility of FAQ import activities within the knowledge base, ensuring better audit trails and user feedback.
Automatically remove wiki folders that become empty when knowledge is
retracted, using the durable finalize lane so pruning waits until ingest
drains. Repository DeleteFolder now atomically checks emptiness to avoid
races with concurrent page or child-folder writes.
- Added AminusB method to the knowledgeRepository, which returns IDs of knowledge in A that have no counterpart in B, comparing by file_hash as a multiset.
- This implementation addresses issues with handling duplicate entries and NULL file_hash values, ensuring accurate synchronization between knowledge bases.
- Updated the KnowledgeRepository interface to reflect the new functionality and improved documentation for clarity.
Introduce platform API keys with capability-based access to tenant management,
system settings, runtime queues, and audit logs, plus admin UI for key lifecycle
and system audit log viewing.
Large Feishu wikis stopped syncing after the first batch (Tencent/WeKnora#2136:
5000+ docs, only ~150 ingested; #1262: a child-list failure aborted the run).
Three compounding causes, fixed in layers:
1. Rate-limit / transient failures had no retry. Feishu's wiki/drive/export
APIs are aggressively rate limited; a thousand-doc sync issues tens of
thousands of calls, and a single 429/5xx burst failed whole swathes of docs.
Add exponential backoff honouring Retry-After to doRequest (JSON) and
downloadRawBytes (files): 429 up to 3x, 5xx once, transport errors back off.
2. The batch path fetched the entire tree into memory before ingesting, so a
sync that hit the 2h task timeout lost all progress and restarted from
scratch, and a transient per-document export failure advanced the cursor and
skipped that doc permanently. Add an optional StreamingConnector interface,
implemented for Feishu: FetchStream interleaves fetch -> ingest -> checkpoint,
bounding memory to one item and persisting a per-node edit-time cursor at
page boundaries. The service dispatches to it via type assertion; connectors
without it fall back to FetchAll / FetchIncremental unchanged (non-breaking).
A resumed/incremental run skips nodes already recorded at their current edit
time (converges), a failed fetch does not advance the cursor (retried next
run instead of dropped), and partial listing failures emit error items and
carry prior edit times forward instead of aborting the sync.
3. Failures and skips were invisible. Unsupported node types (mindnote/slides/…)
were skipped with no item, error or log ("13 docs synced only 3"). FetchAll
now tallies discovered/fetched/failed/skipped per resource with a per-type
skip breakdown and a progress line every 100 nodes. The backend already
persisted per-item errors in SyncLog.result.errors but the frontend type
omitted the field; declare `result` on SyncLog and render up to 50 failed
documents in the sync-log drawer (#1262). i18n for zh-CN/en-US/ko-KR/ru-RU.
result.errors is a jsonb column shipped in every list response, so it is
capped at 100 samples server-side (recordSyncError); the exact failure total
comes from the bounded items_failed counter, which the drawer uses for its
count and "N more" line so a sync that fails thousands of docs stays bounded.
Each sample is a structured, localisable SyncItemError instead of a raw
"title: ...api error: status=500 body={json}" dump that repeated the title
and leaked status/body/log_id. The connector (feishuFailure) classifies the
error into a stable i18n code (+ the feishu code as a param); the service
builds {title, code, params, message}; the drawer localises via vue-i18n to
the viewer's language (zh/en/ko/ru), falling back to message. A custom
UnmarshalJSON keeps old string-form logs readable. The banner is localised
from items_failed too. Raw errors stay in the server logs. Follows the
two-tier norm of Airbyte/Fivetran/Onyx: human, actionable, localised
summaries in the UI; raw detail in logs. Transient errors say "will retry"
(the cursor is retained); auth/permission errors point at the fix instead.
Tests: client_retry_test (backoff), tally_test (accounting), connector_stream /
datasource_stream (streaming dispatch + checkpoint), connector_convergence
(timeout + transient-failure resume converges to full coverage; proven to catch
#2136 via bug injection), datasource_result_cap (error-sample cap + structured
localisable SyncItemError, no raw body), connector_error_reason (error→i18n-code
classification), sync_item_error (legacy-string + structured back-compat) — all
in the normal suite. connector_realapi / connector_seed are an opt-in real-Feishu
harness gated by the feishu_integration build tag AND env vars, never run in CI.
Multiple GORM tag options must be separated by ';', not ','. The comma was
parsed as part of the type value, producing malformed DDL that causes
AutoMigrate(&Message{}) to fail on SQLite with a syntax error.
Restrict channel-managed sessions (API keys, IM, embed) to tenant admin
in the Web console for list/read paths, hide channel sidebar buckets from
non-admins, and add CountSessionsBySource for aggregate stats without
exposing session rows to viewers.
Add GetOwnedSession method to enforce strict ownership checks for session mutations, ensuring tenant admins can read but not modify API-key sessions. Update relevant handlers to utilize this method for operations that alter session data, maintaining security and integrity of API-key sessions.
Let tenant admins list and open API-key sessions from the Web console while
keeping per-key isolation for everyone else. Pin the session source filter so
it no longer jumps when switching buckets, and resolve session origin on read
so deep links stay in sync with the active folder after refresh.
Migrate WeKnora's Langfuse tracing from the legacy /api/public/ingestion
batch API to the OpenTelemetry Go SDK emitting OTLP/HTTP to
/api/public/otel/v1/traces — the Langfuse v3+ standard ingestion protocol.
Why: Langfuse v3 deprecated the legacy ingestion API in favor of OTLP, and
OTel-only backends (e.g. the LiteFuse Doris fork) dropped it entirely, so the
legacy client was incompatible. Migrating to the OTel SDK also adds W3C
traceparent cross-service propagation: an upstream caller's trace id is
inherited so its run and the WeKnora call land under one trace in the UI.
What:
- Manager owns an OTel TracerProvider backed by an otlptracehttp exporter
(Basic auth + the x-langfuse-ingestion-version gate header). Spans are
buffered by BatchSpanProcessor and exported complete on End, so a long
trace no longer fragments across flush batches.
- *Trace/*Span/*Generation wrap OTel spans; the public API (StartTrace/
Span/Generation, handles, InjectTracing, asynq/gin middleware, Config,
TokenUsage) is unchanged, so callers (agent/chat/handler/scheduler/asynqdl)
need no edits.
- W3C traceparent: GinMiddleware extracts traceparent from incoming requests;
AsynqMiddleware + InjectTracing propagate it across the asynq payload.
types.TracingContext gains a LangfuseTraceparent field.
- logger.CloneContext preserves the active OTel span across context rebuilds
so child spans (agent engine, summary generation) stitch to the HTTP trace
instead of orphaning into separate traces.
- StartSpan/StartGeneration re-establish the parent from the *Trace handle
when the ctx lost the active OTel span, covering background/derived contexts.
Compatibility: works with Langfuse v3+ (cloud / self-hosted / the bundled
docker-compose profile) and LiteFuse. Drops Langfuse v2 (legacy-only) support.
Uses the OTel Go SDK already (indirectly) in go.mod, promoting
otlptracehttp to direct; no third-party langfuse-go library.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Chat attachment parser rules belong on agents; drop the duplicate workspace settings section while preserving legacy tenant rules when engine config is updated without that field.
Lark and Feishu are the same product deployed on two isolated clouds
(open.larksuite.com and open.feishu.cn). WeKnora touches both from two
places — the IM bot and the Wiki data-source connector — and this brings
Lark to parity across both. Neither API surface differs, so rather than
duplicate either implementation, each is parametrized by a Region and
registered twice.
IM (internal/im/feishu):
Every API call routes through a.api(), which prefixes the region's
host, and the WebSocket client passes larkws.WithDomain — the SDK
defaults to open.feishu.cn, so Lark long connections would otherwise
fail to authenticate.
Data source (internal/datasource/connector/feishu):
Lark was half-supported: the client honoured a base_url override and a
LarkBaseURL constant existed, but the connector registered only as
"feishu" and the UI never exposed base_url, so Lark wikis could not be
configured at all. Lark is now its own connector type. base_url is kept
as an explicit override so data sources that pointed a "feishu"
connector at open.larksuite.com keep working.
Bugs found and fixed along the way:
- Wiki resource links were hardcoded to https://feishu.cn/wiki/{token} in
four places, so a Lark wiki linked into the wrong cloud entirely.
- Bot identity is now platform-prefixed (feishu:cli_x vs lark:cli_x).
BotIdentity carries a unique index, so without this a Lark channel
would be rejected as a duplicate of an unrelated Feishu bot sharing an
app_id.
- The card image cache is now keyed by app_id. image_keys are issued per
app on one cloud; a Feishu key handed to a Lark app fails the card
update with code=200570.
- The datasource doc claimed Lark "adapts automatically" — it could not,
since the UI had no way to reach base_url.
Permissions are where the clouds genuinely diverge. Identical APIs do not
mean identical permission catalogues: aily:file:read, aily:file:write and
corehr:file:download exist on Feishu but not on Lark, so pasting Feishu's
scope JSON into Lark fails to import. That JSON is also over-broad (it
covers products WeKnora never calls) and simultaneously incomplete (it
omits the drive:*/docx:* scopes the connector needs). Lark therefore gets
its own list, derived from the 17 endpoints the code actually calls and
checked identifier-by-identifier against Lark's own API reference, plus a
full Feishu→Lark mapping table marking each scope verified, unverified,
or absent. Both features use tenant_access_token only — no user_access_
token exists anywhere in the repo — so the user dimension stays empty.
Known limitation: Region.ThinkingText makes Lark's initial card
placeholder English, but once streaming starts the think-block headers
come from the shared MarkdownThinkStyle and are Chinese on every platform
(Lark, Slack, Telegram, Mattermost alike). Fixing that means threading a
locale through FormatIMDisplayContent, which has no platform context — a
cross-platform change, not a Lark one. Documented as such.
The channel-creation 400 message is now derived from validIMPlatforms so
the list cannot drift as platforms are added.
Tests: region invariants and per-region URL construction on both sides,
the Feishu/Lark bot-identity split, platform-to-channel mapping, base_url
override precedence, an end-to-end SendReply against a fake Open Platform
(token fetch, reply, thread-unsupported fallback, hard-error surfacing),
and a ListResources drive asserting wiki links never leak the other
cloud's host.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Update the SuggestedQuestionsRequest to include TagScopes for better filtering of knowledge items based on their parent knowledge bases. Modify related API documentation and frontend interfaces to accommodate the new structure. Ensure backend services handle the new tag scope logic correctly, including tests for the updated functionality.
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.
Implement a maximum limit of 5 for the total number of images and attachments that can be uploaded in a chat session. Introduce a warning message for users when the limit is exceeded. Additionally, add functionality to clear pending uploads when switching sessions to prevent errors related to stale document IDs. Update localization files to include new warning messages in English, Korean, Russian, and Chinese.
Add functionality to normalize temporary attachment IDs, ensuring deduplication and enforcing a maximum limit of 5 attachments per message. Update related handlers and tests to support this new behavior, enhancing the robustness of attachment management in chat sessions.