github.com/Tencent/WeKnora

New Features

• NEW: Official Documentation Site — the headline of this release. A complete VitePress product documentation site now lives in website-docs/, organized into six sections (Getting Started → Architecture → Features → API → Clients → Development) across ~50 pages: 21 feature guides, 13 API reference pages covering ~360 endpoints with permissions/parameters/curl examples, 7 client guides (frontend, CLI, Go SDK, mini program, desktop, Chrome extension, Claw Skill), a configuration reference for ~150 environment variables, and a database/migration guide covering 40+ tables. Ships a custom landing page, a <Screenshot> component with placeholder fallback, Mermaid zooming, and link/diagram validation scripts (scripts/check-links.mjs, scripts/check-mermaid.mjs). The site reads the repository-root VERSION file at build time so version labels never drift. Deployment is self-contained: website-docs/Dockerfile + nginx.conf (listening on 8081) + docker-entrypoint.sh. Quickstart sample data (4 Markdown documents + an FAQ import JSON) and a runnable local MCP demo (examples/mcp-demo/) are included so a first knowledge base can be built end to end without hunting for test files.
• NEW: Knowledge Base Folder Tree — folder uploads no longer smuggle their relative directory into file_name. The path now lives in a dedicated folder_path column with a backfill for existing rows (migration 000079_knowledge_folder_path), so the document list renders a real sidebar folder tree that can be browsed like a file manager, folders can be renamed in place, and documents can be re-filed into another folder. Individually uploaded documents get a visible home at the tree root instead of disappearing. Adds GET /knowledge-bases/{id}/knowledge/folders and PUT /knowledge-bases/{id}/knowledge/folders, plus a folder picker popup on each row.
• NEW: Editable Chunks, Revision History & Custom Metadata — retrieval chunks are now editable directly from the UI, with every superseded version snapshotted for diff and one-click revert, and the index rebuilt automatically after an edit (migration 000078_chunk_editing_and_custom_metadata adds chunks.source_content / content_revision / index_status / last_editor_id / context_header, the chunk_revisions table, and knowledges.custom_metadata). Generated questions can be added, edited, deleted, and regenerated per chunk, and survive content edits. Adds PUT /chunks/{knowledge_id}/{id}, GET /chunks/{knowledge_id}/{id}/revisions, POST /chunks/{knowledge_id}/{id}/revert, and the generated-question endpoints. Chunk details moved into header popups for a tighter layout.
• NEW: Wiki Page Revision History, Diff & Manual Editing — every wiki page version is preserved before being overwritten (migration 000075_wiki_page_revisions, plus wiki_pages.last_edit_source / last_editor_id recording whether the current version came from the pipeline, an agent, a user, or a revert). A revision drawer lists the full history with line-level diffs and one-click rollback, and pages can be edited by hand in the browser. The Wiki reader layout, source-document reference handling, and sidebar UX were reworked to match the session-list experience.
• NEW: Directly Loadable File URLs (resource_urls=public) — chat answers, references, knowledge search, and embed responses can return ready-to-load http(s) URLs for files and images instead of internal resource:// handles, so third-party apps no longer need a second authenticated call to the /files proxy. Opt in per request with ?resource_urls=public or per deployment with RESOURCE_URL_MODE=public. A new internal/storageurl package centralizes the rewriting, reusing a live access grant per resource to keep read endpoints from writing on every request. Anonymous embed channels and KB-restricted API keys always stay on handle. See API overview.
• NEW: Feishu Drive Data Source — a new Feishu Cloud Drive connector joins the existing Feishu wiki connector (#2466). New-format Feishu cloud documents (docx) are now synchronized through the blocks API with per-block-type drill-down (#2087), configurable via the new E9 docx parsing-mode section in .env.example.
• NEW: Batch Document Tagging — select documents in the knowledge base list and apply tags in bulk through a dedicated dialog that pre-selects the tags already common to the selection.
• NEW: MCP Server 1.1.x — the WeKnora MCP server migrated from the removed low-level Server decorator API to the high-level MCPServer API (mcp 2.x), restoring HTTP (stateless_http) and SSE (/sse/messages/) transport compatibility, and now publishes as the official PyPI package tencent-weknora-mcp via Trusted Publishing. Two new tools bring the total to 29: create_knowledge_from_text (create a knowledge entry from Markdown text) and list_shared_knowledge_bases (shared KBs are also folded into name resolution and tool hints).
• NEW: AWS S3 Default Credential Chain — leaving S3_ACCESS_KEY and S3_SECRET_KEY both empty now falls back to the AWS SDK default credential chain, supporting EC2/ECS/EKS IAM roles, IRSA / Web Identity, environment variables, and shared config files (#2008).
• NEW: Local HTML Upload Parsing — docreader gained a dedicated HTML parser, so .html files can be uploaded directly instead of only imported by URL. The supported-extension set is now the single gate for both direct upload and URL import (#2447).
• NEW: QQBot Markdown Replies — QQBot channels reply with markdown formatting like the other IM integrations.
• NEW: PR CI Workflows — dedicated GitHub Actions checks for the Go app, frontend, docreader, and MCP server, plus a scripts/verify_frontend_pr.sh helper for local pre-PR verification.

Improvements

• IMPROVED: Router and auth middleware split by domain — the 2390-line internal/router/router.go was split into routes_agent.go, routes_auth_tenant.go, routes_chat.go, routes_infra.go, routes_knowledge.go, files.go, and static.go, with the duplicated file proxies deduplicated. The Auth middleware was likewise split and session-context attachment unified in one place.
• IMPROVED: modelcontext package consolidation — llmreference and llmresource were folded into modelcontext as source and resource codecs, every handle space was rebuilt on one generic table, source-key gating moved behind a single dispatch table, and stream decoding now goes through one suffix-hold primitive. Handle terminology is consistent across the merged package, and several handle-codec bugs were fixed along the way.
• IMPROVED: Unified Wiki operation history — the redundant Wiki Browser operation log was removed and its storage/API retired (migration 000077_remove_wiki_log); wiki mutations are recorded only in the knowledge-base Activity view.
• IMPROVED: Rerank quality and throughput — reranker passage cleaning now preserves code and math passages, NVIDIA logit-style scores are normalized before fusion, and Tencent LKEAP requests are batched to stay within API limits.
• IMPROVED: Chinese query expansion — expanded queries are segmented with jieba so Chinese rewrites actually match the sparse index.
• IMPROVED: Chunking consistency — line endings are normalized before splitting, overlap boundaries preserve semantic sentence ends (English ?/! require a following space), the chunking preview matches parent-child splitting, and XLS header-override behavior is aligned with XLSX.
• IMPROVED: Document summaries — table chunks keep every row in the summary, failed summary generation retries before falling back, and stale summaries are re-enqueued for refresh with the tenant context restored for background refresh tasks.
• IMPROVED: Frontend i18n hygiene — locale files were pruned against zh-CN, with an audit harness (localeKeyAudit, audit-action registry, regeneration script) that keeps the other locales aligned and catches missing keys in CI.
• IMPROVED: UI polish — redesigned upload confirmation dialog (tags can be configured at upload time), refined user settings menu, unified organization settings modal scrolling, clearer document processing timeline statuses, and a persistent live region that announces RAG wait status for screen readers.
• IMPROVED: Deployment docs — docker compose pull guidance completed across all deployment steps so upgrades stop reusing stale cached images.

Bug Fixes

• FIXED: Registration passwords are no longer sanitized before hashing, which corrupted passwords containing special characters.
• FIXED: WeCom long-connection deadlock between closeConn and heartbeatLoop; connection close is now scoped to its own generation, and disabled channels are handled in the IM callback.
• FIXED: SQLite — DataSource deletion failed, a zero vector threshold was treated as a filter instead of "no filter", and vector candidates are now filtered before top-k.
• FIXED: Evaluation retrieval metrics were always zero; chunks are mapped back to passage IDs and passage indexing is synchronous.
• FIXED: Empty or mismatched embedding results could panic an ants worker and deadlock a mutex.
• FIXED: Retrieval merge output ordering is deterministic, and trusted overlaps are trimmed by exact position rather than text search (including after pipeline rewrites, #2558).
• FIXED: Archived wiki pages are excluded from stats queries and the lint cursor; wiki prompt language is always resolved; wiki state is reconciled when documents move across knowledge bases; batched wiki_read_page results stay within the output budget.
• FIXED: Retried FAQ creation no longer produces duplicate entries, and disabled FAQ entries are excluded from agent retrieval.
• FIXED: Knowledge base deletion cleans up bound vector stores in batch delete and retries when engine resolution is deferred; a missing vector store engine is rebuilt on demand.
• FIXED: MinerU automatic PDF parsing preserved; skipped stages are counted in parsing progress; stale error_message values are cleared on every reprocess and finalize path; batch reparse failures are surfaced.
• FIXED: The embed runtime can read its own session again, fixing follow-up-suggestion 500s and lost conversations on refresh.
• FIXED: Answer-generation status is shown after retrieval instead of a silent gap; duplicate terminal answer events are suppressed; retrieval fallback errors are preserved and web partial success / wiki-only embed fallback are covered.
• FIXED: Agent web evidence is preserved when page fetches fail, with a safe rendered web_fetch fallback.
• FIXED: Frontend — attachment upload status stays reactive, protected embed images route through the channel file proxy, and audit / KB-activity strings were restored after the locale prune.
• FIXED: Shared agent source selector validation and shared agent source permissions hardened; API-runtime admin-console bypass scoped to owned sessions; external users included in the API session group.
• FIXED: Remote images referenced by HTML img tags resolve during parsing (#2355); PaddleOCR-VL Cloud HTML tables are normalized; MCP server diagnostics route to stderr to fix a stdio protocol crash (#2371); non-UUID agent IDs resolve in MCP.
• FIXED: Duplicate detection distinguishes files by type and is scoped correctly; custom metadata defaults on create; image uploads can fall back to the file service; wiki folder path is preserved after creation.

Infrastructure & Build

• BUILD: Migrations 000075_wiki_page_revisions, 000076_knowledge_metadata_external_id_index, 000077_remove_wiki_log, 000078_chunk_editing_and_custom_metadata, 000079_knowledge_folder_path; SQLite migrations 000001_remove_wiki_log, 000002_knowledge_folder_path.
• BUILD: New internal/storageurl package for storage reference URL rewriting; Go client extended with resource_urls support.
• BUILD: GitHub Actions workflows added for app / frontend / docreader / mcp-server; gofmt check scoped to PR commits; App workflow Go dependency caching sped up.
• BUILD: Swagger / API docs regenerated for chunk revisions, knowledge folders, wiki revisions, and resource_urls.

Documentation

• DOC: website-docs/ official documentation site added (six sections, ~50 pages, ~360 documented endpoints), with quickstart sample data and a local MCP demo under examples/mcp-demo/.
• DOC: docs/api/README.md and docs/api/chat.md document resource_urls and RESOURCE_URL_MODE; Feishu Drive data source guide added under docs/wiki/集成扩展/.
• DOC: docs/QA.md extended for the documentation site, folder tree, chunk editing, wiki revisions, and public resource URLs.
• DOC: Architecture diagram updated for the folder tree, chunk editing, wiki revision history, and public file URLs.