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.
Introduce tenant-scoped storage backend registry so workspaces can register
multiple storage instances and bind each knowledge base to a specific backend,
with migration from legacy tenant-level storage config and new settings UI.
Added new tests for the ImageMultimodalService to verify the handling of orphan tasks and the finalization of pending counters. Improved runtime task inspection with better error handling and task cancellation logic, ensuring robust management of orphan tasks in the system.
Split the monolithic upstream worker budget into explicit core, post-process,
enrichment, maintenance, and shared elastic pools, expose live cluster capacity
on the runtime dashboard, and retire the legacy asynq.concurrency setting.
Route summary/sync/maintenance tasks to dedicated queues and servers so parsing, enrichment, and batch work no longer share one concurrency budget, while extending the runtime dashboard and settings validation to reflect the new topology.
- docreader: serialize all pdfium access behind a global lock. pypdfium2
is process-global and not thread-safe, so concurrent PDF uploads could
corrupt its shared state and deadlock the whole process (requests stuck
in "Parsing document" forever). Non-PDF parsers still run concurrently.
- logger: strip ANSI color escapes from file logs so on-disk logs stay
plain text while stdout keeps terminal colors.
- storage: log baseDir/resolvedPath on file-open failure and when the
multimodal service resolves a file service, so a LOCAL_STORAGE_BASE_DIR
mismatch between writer and reader is immediately visible.
- docker: set TZ (default Asia/Shanghai) for the docreader container.
- Updated the `CreateKnowledgeFromFile` function to accept an optional `processConfig` parameter for per-upload parse configuration overrides.
- Introduced the `KnowledgeProcessOverrides` type to define various parsing settings, including parser engine rules and chunking configurations.
- Modified the API client and related components to handle the new process configuration, improving flexibility in knowledge file uploads.
- Updated documentation and API specifications to reflect the new `process_config` parameter in the upload process.
- Added a new `UploadConfirmHost` component to facilitate user confirmation for uploads, integrating with the updated upload logic.
- Enhanced user experience by providing detailed configuration options during the upload process.
Lets users stop an in-flight document parse to free up LLM / worker
resources without losing the chunks and index already written. The
core insight is that the previous parse_status=completed flipped as
soon as primary chunks landed, while the most expensive subtasks
(graph extract = N LLM calls per chunk, plus summary, question
generation) were still running in the background — so "completed"
wasn't actually terminal from a resource standpoint.
State machine
pending -> processing -> finalizing -> completed
|
+-> cancelled (any of the three
in-flight states)
+-> failed
+-> deleting
`finalizing` is the new post-process fan-out window. parse_status
only promotes to `completed` once pending_subtasks_count (a new
column tracking summary + question + per-chunk graph extract)
drains to zero via atomic FinalizeSubtask. Wiki ingest is
intentionally excluded from the counter — it's a KB-scoped
debounced batch and would otherwise pin parse_status in
`finalizing` for the wiki batch window.
Backend
- New ParseStatusFinalizing + pending_subtasks_count column with
migration 000056.
- knowledgeRepository.SetFinalizing transitions processing -> finalizing
conditionally so a racing cancel cannot be clobbered.
- knowledgeRepository.FinalizeSubtask atomically decrements the
counter and self-promotes the row to completed when it hits zero.
- KnowledgePostProcess restructured to compute expected subtask
count up front, flip to finalizing (or completed when no
enrichment is enabled), and only then fan out subtasks. Subtask
handlers (summary, question, graph extract) defer-decrement on
terminal exit using the existing isFinalAsynqAttempt convention.
- New POST /api/v1/knowledge/{id}/cancel-parse handler accepting
pending / processing / finalizing. Marks the row cancelled,
zeroes the counter, best-effort dequeues asynq tasks via a new
TaskInspector abstraction (asynq-mode walks pending/scheduled/
retry queues; Lite-mode noop), and scrubs wiki ingest pending op.
- SpanTracker.AbortAttempt flat-sweeps every still-running span
for the attempt via a new repo.CancelAllOpenSpans helper so the
trace viewer's striped bars all flip to cancelled, even leaf
generations whose parent stage already EndSpan'd (multimodal
fan-out pattern). knowledge_post_process closes its postSpan
via SkipSpan on the cancel/deleting entry guard so a worker
that opens a span AFTER the cancel sweep doesn't leak it.
- Housekeeping and resetPendingTasks sweep finalizing rows
identically to processing so a crash/restart can't strand them.
- DeleteKnowledge/DeleteKnowledgeList proactively dequeue
downstream tasks via the same TaskInspector path.
- ChunkExtractService gets a cancel entry guard so the most
expensive enrichment (graph extract) bails immediately when the
parent knowledge is aborted.
Frontend
- New cancelKnowledgeParse API client + "Stop parsing" entry in
both list view and card view more menus, gated on
pending/processing/finalizing.
- Polling predicate refactored to a shared isParseInFlight helper
that recognises `finalizing` (previously the doc list silently
stopped polling once parse_status flipped from processing).
- Knowledge processing timeline: isPolling includes finalizing,
new isHardTerminal short-circuits LIVE for cancelled/failed/
completed so stranded child spans cannot pin LIVE on.
- DocumentListView.computeStatus distinguishes finalizing
("增强中") from completed and shows the previous "生成摘要中"
copy when summary_status is still pending under finalizing.
Added cancelled badge as well.
- i18n: statusFinalizing / statusCancelled / cancelParse* keys
across zh-CN, en-US, ko-KR, ru-RU.
Docs / SDK
- docs/api/knowledge.md: documents the new finalizing state,
cancel-parse semantics, and which statuses accept cancel.
- client (Go SDK): CancelKnowledgeParse with docstring listing
the cancellable statuses.
The per-image multimodal subspan only captured image_url / enable_ocr /
enable_caption on input and chunk_id on output, so the trace viewer
could not answer "what did THIS image actually produce?" without
joining back to the chunks table.
Adds to the per-image span output:
- vlm_model_id (or "legacy_inline" for inline-config KBs)
- image_bytes (read size)
- ocr_prompt: "default" | "scanned_pdf"
- ocr_chars + ocr_preview (sanitized text, capped at 200 runes)
- caption_chars + caption_preview
- chunks_created (count of OCR/caption child chunks)
- indexed (true after BatchIndex completes)
- per-step error fields (read_error / ocr_error / caption_error /
skipped reason) when something fails
Also adds parent_chunk_id to the span input so the trace links back to
the text chunk this image hangs off — useful when a doc has hundreds
of inline images and you need to know WHERE in the text this one came
from.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Addresses review feedback that the PR② design had four shortcomings:
1. The pipeline is a DAG, not a sequence — Embedding and Multimodal
are independent of each other, both downstream of Chunking, both
upstream of PostProcess. The flat (knowledge_id, stage) table
couldn't represent that, so a Chunking failure left dependents
stranded as "pending" forever instead of being marked as
impossible-to-run.
2. No history across attempts. A reparse erased the previous run's
status before the new run started, leaving operators with no way
to investigate "why did this fail twice?".
3. Stages had only status + duration. Operators want to know how big
the work was — pages parsed, chunks created, tokens embedded, VLM
calls made — to distinguish "slow because the file is huge" from
"slow because the docreader is wedged".
4. Multimodal fans out N image tasks; Embedding fans out M batches;
PostProcess fans out into Summary/Question/Wiki/Graph. Each unit
is interesting on its own (Langfuse already captures this for
LLM calls). The flat model couldn't express it.
The redesign mirrors Langfuse's trace/span/generation hierarchy:
* Migration 000053 supersedes 000052: knowledge_processing_spans
table with (knowledge_id, attempt, span_id) primary key, plus
parent_span_id, kind ∈ {root, stage, subspan, generation},
status ∈ {pending,running,done,failed,skipped,cancelled}, and
JSONB input/output/metadata fields.
* SpanTracker (replacing StageTracker) exposes OpenAttempt /
BeginStage / BeginSubSpan / EndSpan / FailSpan / SkipSpan /
LookupStage. Cross-process workers (image_multimodal) get the
parent's attempt + span via payload + LookupStage so subspans
attach correctly.
* StageDependencies declares the DAG; FailSpan now cascades —
descendants of the failed span and dependent stages are flipped
to "cancelled" with a UPSTREAM_FAILED code. The UI sees a clear
blast radius instead of orphan spinners.
* Reparse now calls OpenAttempt up front so the timeline reflects
"new attempt, all pending" instead of letting the previous run's
status linger until the worker picks up the task.
* Image_multimodal records each image as a generation subspan with
its own success/failure on the parent attempt's multimodal stage.
The finalize-on-last-attempt counter logic is preserved unchanged.
* GET /api/v1/knowledge/:id/spans (also kept /stages alias) returns
a tree shape with synthesized pending placeholders so the
frontend always renders five timeline segments. ?attempt=N
enables history navigation.
Several failure modes left Knowledge.parse_status pinned at "processing"
forever, with no signal to users beyond a permanent spinner. This commit
addresses the root causes and adds a safety net.
- Asynq worker pool: explicit Concurrency (default 16, env-tunable via
WEKNORA_ASYNQ_CONCURRENCY) so batch uploads don't queue behind a
CPU-count-sized worker pool. Redis op timeouts raised to 500ms/1000ms
(WEKNORA_REDIS_OP_TIMEOUT_MS) to absorb bursty multimodal counter ops.
- DocReader RPC: cap each call with WEKNORA_DOCREADER_CALL_TIMEOUT
(default 30m). Without this, a hung docreader pinned a worker for the
full DocumentProcessTimeout window.
- ImageMultimodal: finalize-on-last-attempt semantics. A permanently
failing single image no longer strands the parent — the asynq retry
is allowed to run, but on the final attempt we count the image
regardless of outcome. Redis DECR errors fall back to enqueuing the
post-process task instead of returning silently.
- Dead-letter callback: when DocumentProcess / KnowledgePostProcess /
ManualProcess exhausts retries, immediately mark the corresponding
Knowledge as failed with the last error. This surfaces the failure
in the UI without waiting for the housekeeping sweep.
- HousekeepingService: 5-minute cron that flips knowledge rows stuck
in "processing" past DocumentProcessTimeout + 10m to failed, plus
summary rows stuck > 1h. Catches anything the other safety nets
miss (worker SIGKILL mid-handler, etc.). Disable with
WEKNORA_HOUSEKEEPING_ENABLED=false.
- Distributed startup recovery: previously the post-restart sweep was
skipped whenever REDIS_ADDR was set, even though Asynq does not
reschedule the task that was actively running on the dead instance.
Now the sweep runs in distributed mode too, but only against rows
older than 30 minutes to avoid racing peer instances.
Extracts the 25 repeated NewCompositeRetrieveEngine call sites across
seven services into two factory functions with tenant-ownership
verification: CreateRetrieveEngineForKB (synchronous) and
CreateRetrieveEngineFromPayload (async task handlers). Promotes
GetByStoreID from concrete-only to interfaces.RetrieveEngineRegistry.
Extends KBDeletePayload / IndexDeletePayload with an omitempty
VectorStoreID snapshot.
Design highlights
- Factory verifies tenant ownership of the resolved store
(defense-in-depth) so a gap in PR3 validation or a tampered Asynq
payload cannot cross tenants. Cross-tenant attempts return
ErrVectorStoreForbidden with a structured log entry.
- Sentinel errors (ErrTenantInfoMissing, ErrVectorStoreNotFound,
ErrVectorStoreForbidden) let async handlers classify non-retryable
failures as asynq.SkipRetry.
- nil and empty-string pointers for vectorStoreID normalize to
"unbound" so callers never send an empty UUID to GetByStoreID.
- resolveBoundEngine constructs engineInfos directly (with
slices.Clone of Support()) instead of going through
NewCompositeRetrieveEngine. The latter only reads from the
byEngineType map and cannot reach DB stores in byStoreID -- multiple
stores can share the same engine type. Unbound fallbacks still reuse
NewCompositeRetrieveEngine so the tenant effective-engines path is
byte-for-byte unchanged.
- DeleteKnowledgeBase now reads the KB before soft-delete so the
enqueued KBDeletePayload carries a VectorStoreID snapshot; once the
row is soft-deleted GORM's default scope hides it. Cost: +1 SELECT
per KB-delete request (a rare admin operation).
- enqueueIndexDeleteTask gains a vectorStoreID *string parameter that
the single caller (tag.go) populates from the owning KB. The worker
(ProcessIndexDelete) validates ownership via the factory.
- cleanupKnowledgeResources loads the KB so bound-store cleanup routes
correctly. If the load fails it warns and falls back to tenant
effective engines; orphan vectors become observable in logs instead
of disappearing silently.
- DeleteKnowledges (batch) stays on the tenant effective-engines path
-- batches can span stores and multi-store fan-out is PR4 scope.
Noted inline.
Merge safety
All knowledge bases have VectorStoreID = NULL at merge time (introduced
nullable in PR1, Tencent/WeKnora#994). Every factory call falls back to
NewCompositeRetrieveEngine(registry, tenantInfo.GetEffectiveEngines()),
which is the existing code path for these KBs. Async payloads from
before this change decode to VectorStoreID = nil (omitempty) and take
the same fallback.
Testing
- go build ./...
- go vet ./...
- go test -race -count=1 ./internal/application/service/retriever/...
covers unbound (nil/empty), bound, cross-tenant, unregistered store,
ownership infra error, legacy payload shapes (missing field/null/
empty string/nil), tampered payload, and parallel invocation.
- grep -r NewCompositeRetrieveEngine internal/application/service
returns zero results outside factory.go/composite.go.
Refs: Part of Tencent/WeKnora#993. Depends on Tencent/WeKnora#994.
When image bytes for a multimodal task cannot be read via FileService
(e.g. tenant.StorageEngineConfig.MinIO is empty while the image was
saved using the global MINIO_* env vars), the previous code fell back
to the HTTP downloader, which rejected the provider:// URL with
"unsupported URL scheme". The asynq handler then returned an error,
asynq retried until exhaustion, and the per-knowledge "pending images"
counter was never decremented — leaving the document stuck in
"processing" indefinitely (issue #1282).
Changes:
- ImageMultimodalService now holds a default FileService and
resolveFileServiceForPayload falls back to it when the tenant-scoped
storage config cannot produce a usable service, mirroring the
write-side fallback in knowledgeService.resolveFileService.
- Extract readImageBytes: provider:// URLs are read exclusively via
FileService and never handed to the HTTP downloader.
- On unrecoverable read failure for a single image, log and skip that
image but still call checkAndFinalizeAllImages so the parent
knowledge can progress to post-processing.
Fixes#1282
Previously the Langfuse integration only traced in-process HTTP requests
(chat / search / eval), so file uploads and every downstream asynq task
(document parse, chunk embedding, OCR/VLM, summary / question gen, wiki
ingest, datasource sync, etc.) produced either disconnected shallow
traces or no observation at all.
This change threads one trace end-to-end:
- tracer: add SPAN observation type and StartSpan; add ResumeTrace so a
worker can attach to an upstream trace without emitting a duplicate
trace-create; StartGeneration now auto-picks parentObservationId from
ctx so nested trace -> span -> generation trees render correctly.
- types.TracingContext + LangfuseTracingCarrier: embed on all 17 asynq
payloads so trace_id / parent_obs_id / user_id / session_id serialise
into every job.
- langfuse.InjectTracing: injected at 28 enqueue sites before json.Marshal
so the HTTP-layer trace survives the Redis hop.
- langfuse.AsynqMiddleware: mux.Use hook that peeks the payload, either
resumes the upstream trace or opens a standalone asynq.<type> trace
for scheduled jobs, and wraps the handler in a SPAN with task metadata
(id / queue / retry / payload_bytes) plus ERROR level on failure.
- GinMiddleware.shouldTrace: whitelist ingestion / knowledge-mutation /
FAQ / wiki / datasource endpoints so the root trace actually starts.
- Tests: tracer_test.go covers span nesting, error status, and
ResumeTrace no-trace-create guarantee; asynq_test.go covers
InjectTracing round-trip, middleware resume path, and standalone
trace fallback.
- Docs: docs/Langfuse\u96c6\u6210.md now lists the covered task types
and documents the cross-process propagation model.
No behavioural change when Langfuse is disabled (all new code paths are
no-ops and carriers serialise to empty strings with omitempty).
- Added a check to skip vector and keyword indexing for knowledge bases that do not require an embedding model, preventing errors during the indexing process.
- Ensured that chunks are marked as indexed even when skipping the indexing step, maintaining a consistent state for downstream processes.
- Enhanced logging to provide clear information on the indexing status and any issues encountered while updating chunk statuses.
These changes improve the robustness of the indexing process in the ImageMultimodalService, ensuring better handling of knowledge bases without embedding capabilities.
- Added a new logger for LLM calls, enabling detailed logging of model interactions, including request and response data.
- Introduced configuration options for enabling and specifying the log directory for LLM debug logs.
- Implemented cleanup for old log files to manage disk space effectively.
- Wrapped existing chat implementations to log calls when debug logging is enabled, enhancing traceability for model interactions.
Introduce KnowledgePostProcess task to decouple image multimodal
processing from downstream operations like summary, question
generation, and knowledge graph extraction.
This allows unifying the processing logic across documents and
images and ensures all text content, including from images,
is properly extracted before post-processing.
- Changed 'Audio Processing' to 'Audio & Video' in English, Korean, Russian, and Chinese translations to reflect expanded functionality.
- Updated ASR configuration titles and descriptions to include video processing capabilities, enhancing user understanding of the feature.
- Renamed file type labels from 'Audio Files' to 'Audio & Video' across multiple languages for consistency and clarity.
- Enhanced document upload handling in the KnowledgeBase to filter unsupported file types, providing user feedback for filtered images, videos, and audio files.
This update improves the internationalization of the application and clarifies the capabilities of the ASR feature.
- Add DefaultLanguage() that reads WEKNORA_LANGUAGE env with zh-CN fallback
- Replace hardcoded zh-CN in LanguageNameFromContext with DefaultLanguage()
- Unify middleware/language.go to use types.DefaultLanguage()
- Add Language field to DocumentProcessPayload (8 enqueue sites + worker restore)
- Add Language field to ImageMultimodalPayload (enqueue + worker restore)
- Propagate language from ImageMultimodalPayload to QuestionGenerationPayload
- Introduced a new task enqueuer in the ImageMultimodalService to handle question generation for image knowledge.
- Implemented logic to enqueue question generation tasks when valid captions or OCR content are available, enhancing the knowledge base's interactivity.
- Updated image reference handling to ensure markdown URLs are correctly formatted, accommodating filenames with spaces.
- Enhanced regex pattern in image resolver to support filenames with spaces, improving image processing accuracy.
Convert all LLM-facing prompts, templates, and tool output labels to
language-neutral English so the model responds in the user's input
language. Update frontend relevance/match-type mappings to match.
Output language strategy:
- User-facing: reply in user's input language
- Vector-indexed: match source document language
- LLM-internal: fixed English
- Removed the image storage handling from the `_resolve_images` function, which now only decodes images and returns them as inline bytes.
- Updated the function's docstring to clarify that image persistence is managed by the Go App, and the return value for `image_dir_path` is always empty.
- Adjusted related gRPC and protobuf files to remove the `ConvertToPDF` method and its associated types, streamlining the API.
- Enhanced security measures in the frontend by implementing a placeholder for provider images and ensuring proper hydration of protected file images.
These changes improve the clarity and efficiency of image processing and API interactions within the application.
- Introduced a loading state (`loadingChunks`) to manage document chunk loading more effectively.
- Updated the scroll handling logic to prevent multiple requests while loading chunks.
- Adjusted the pagination logic to ensure correct page increments based on the total document count.
This update improves the user experience by ensuring smoother document loading and better state management during scrolling interactions.
- Updated the `KnowledgeBase` struct to rename `cos_config` to `storage_config` for consistency across the application.
- Implemented a custom `UnmarshalJSON` method to maintain backward compatibility with legacy responses that still use `cos_config`.
- Modified various frontend components to reflect the new `storage_config` naming, ensuring seamless integration with the updated backend structure.
- Enhanced Dockerfiles and service configurations to support the new storage engine setup.
This update improves the clarity of storage configurations and ensures compatibility with existing data structures.
- Added a new `.env.lite.example` file for the Lite version, providing a minimal configuration template.
- Updated `.env.example` to remove deprecated variables and include new Docreader settings.
- Enhanced Docker configurations to support the Lite version, including a new Dockerfile for the Docreader service.
- Introduced a Makefile target for building and running the Lite version, along with packaging capabilities.
- Created GitHub workflows for building and releasing Lite binaries, including Homebrew formula support.
- Implemented a new service file for managing the Lite version as a system service.
This update enables a streamlined, single-binary deployment of WeKnora, reducing external dependencies and simplifying setup.