github.com/Tencent/WeKnora

What's Changed

• refactor(agent): consolidate model-handle codecs into internal/modelcontext by @lyingbug in #2330
• refactor(auth,router): unify auth session context and split router by domain by @lyingbug in #2331
• fix(knowledge): clean bound vector stores in batch delete by @BigFishDreamWater in #2329
• fix: support string args for skill script execution by @koishi514-Z in #2318
• fix(agent): keep batched wiki_read_page results within output budget by @lyingbug in #2333
• fix(pipeline): 修复合并阶段输出顺序的不确定性 by @mdrkrg in #2319
• fix(knowledge): reconcile wiki state when moving documents across KBs by @lyingbug in #2335
• feat(wiki): page revision history with diff, revert, and manual editing by @lyingbug in #2336
• style(wiki): polish revision drawer and sidebar to match session list UX by @lyingbug in #2338
• feat(feishu): sync docx via blocks API with per-type downdrill by @voidkey in #2334
• fix(i18n): remove duplicate 'members' key in organization locale objects by @hiasky in #2305
• fix(migrations): renumber knowledge metadata index migration to 000076 by @lyingbug in #2340
• fix(im): harden connection lifecycle across replicas by @lyingbug in #2349
• Refine user settings menu and access shortcuts by @lyingbug in #2368
• Remove redundant Wiki operation log by @lyingbug in #2367
• feat(knowledge): configure upload tags and processing options by @lyingbug in #2369
• fix: Remove SanitizeForDisplay from chunk handler responses by @mdrkrg in #2323
• feat(knowledge): support editable chunks and retrieval metadata by @lyingbug in #2370
• fix(mcp-server): correct SSE message route to prevent initialization … by @FincherYuan in #2375
• fix(im): recover channels after leader lock loss by @March-77 in #2365
• fix(knowledge): clear stale errors when reparse starts by @March-77 in #2363
• docs: clarify scope-aware contributor validation by @March-77 in #2366
• fix(frontend): distinguish main stages from postprocess tasks by @March-77 in #2364
• fix(wiki): prevent finalizing stalls on enqueue failure by @BigFishDreamWater in #2362
• fix(docparser): resolve remote images referenced by HTML img tags by @ochanism in #2355
• fix(mcp-server): route diagnostics to stderr to fix stdio protocol crash by @annopick in #2371
• fix(im): render resource:// images in IM channels (nginx /r/ + config guards) by @voidkey in #2352
• fix(chat): prevent duplicate terminal answer events by @koishi514-Z in #2353
• fix(docreader): restore optional XLSX header context by @WHUTcjh-2024 in #2314
• fix: 在chunk overlap中保留更细的语义边界 by @ttommybot in #2342
• [Bug] preserve web evidence when page fetches fail by @langcaiye in #2309
• fix(knowledge): default custom metadata on create by @lyingbug in #2379
• fix(knowledge): distinguish duplicate files by type by @koishi514-Z in #2384
• fix(storage): allow image uploads to use file service fallback by @lyingbug in #2389
• fix(retriever): rebuild a missing vector store engine on demand by @ochanism in #2357
• fix(agent): exclude disabled FAQ entries by @lyingbug in #2392
• fix(mcp-server): pin MCP Python SDK to v1.x by @lyingbug in #2395
• Redesign upload confirmation dialog by @lyingbug in #2397
• ci(mcp-server): remove duplicate workflow.yml (superseded by mcp-server.yml) by @annopick in #2398
• fix(knowledge): support local HTML file uploads by @BigFishDreamWater in #2393
• fix(session): include external users in API session group by @lyingbug in #2405
• fix(organization): pass avatar when creating shared space by @hiasky in #2408
• feat(knowledge): add batch tag dialog for selected documents by @hiasky in #2406
• fix: harden shared agent source permissions by @lyingbug in #2417
• ci: add PR checks for app, frontend, and DocReader by @lyingbug in #2416
• fix(mcp-server): migrate to MCPServer high-level API (mcp 2.x) [v1.1.0] by @annopick in #2420
• fix(docparser): normalize PaddleOCR-VL Cloud HTML tables by @zldkr in #2419
• test(ci): stabilize MCP session test and decouple frontend checks by @lyingbug in #2424
• fix(mcp): 修复非 UUID Agent ID 无法解析的问题 by @ZhuangShiHao1207 in #2422
• fix(sessions): 放行 embed 运行时读取自身会话，修复追问建议 500 与刷新丢对话 by @asmh1989 in #2421
• fix(frontend): resolve i18n @ syntax errors and TDesign drawer warnings by @lyingbug in #2426
• fix: preserve wiki folder path after creation by @ttommybot in #2429
• fix(sqlite): filter vector candidates before top-k by @BigFishDreamWater in #2440
• fix: 修复 SQLite 下删除 DataSource 失败 by @yuefanxiao in #2435
• fix(evaluation): use sync passage indexing to fix always-zero retrieval metrics by @lazyayuan in #2359
• fix(frontend): 编辑保存后自动关闭编辑器，创建后引导继续配置 by @lyingbug in #2448
• fix(frontend): restore audit and KB activity i18n after locale prune by @lyingbug in #2449
• fix(frontend): route embed protected images through channel file proxy by @lyingbug in #2454
• fix(knowledge): allow xlsx in URL import (#2447) by @lyingbug in #2452
• fix(chat): segment Chinese query expansions with jieba by @BigFishDreamWater in #2463
• fix(knowledge): surface batch reparse failures by @Myang-print in #2453
• fix(rerank): batch LKEAP requests within API limits by @BigFishDreamWater in #2501
• fix(chat): keep attachment upload status reactive by @jacentsao in #2498
• fix(rerank): normalize NVIDIA logit scores by @BigFishDreamWater in #2496
• fix(rerank): preserve code and math passages by @BigFishDreamWater in #2495
• fix: count skipped stages in parsing progress by @ttommybot in #2500
• fix(docparser): preserve MinerU automatic PDF parsing by @oscarlius in #2493
• feat(mcp-server): add list_shared_knowledge_bases tool by @FincherYuan in #2497
• fix: exclude archived wiki pages from lint cursor by @quan020406 in #2503
• fix(knowledge): clear stale error_message on reprocessing and finalize by @oscarlius in #2488
• fix(knowledge): clear stale error_message on the remaining reprocess paths by @lyingbug in #2506
• fix: unify organization settings modal scrolling by @ttommybot in #2508
• fix: show answer generation status after retrieval by @ttommybot in #2514
• fix(embedding): prevent ants worker panic and mutex deadlock on empty/mismatched results by @qfzc in #2520
• feat(knowledge): folder tree sidebar, folder path storage, and document re-filing by @lyingbug in #2528
• fix(frontend): restore runtime queue i18n and guard against prune by @lyingbug in #2529
• fix(chat): preserve retrieval results during embedding failures by @KKtwo in #2525
• perf(container): batch legacy storage alias lookup at startup by @lyingbug in #2530
• feat(storage): support AWS S3 NoAk default credentials by @qilifan in #2008
• fix(summary): preserve rows across table chunks by @BigFishDreamWater in #2538
• feat(api): resource_urls=public for directly loadable file URLs by @lyingbug in #2531
• fix(chunking): align preview with parent-child splitting by @BigFishDreamWater in #2539
• fix(wiki): ensure prompt language is always resolved by @lyingbug in #2543
• fix(chunking): sync preview swagger and normalize splitter config by @lyingbug in #2541
• fix(faq): prevent duplicate entries when a create is retried by @lyingbug in #2544
• Improve Wiki reader layout and revision history UX by @lyingbug in #2547
• fix(chunking): normalize line endings before splitting by @BigFishDreamWater in #2546
• feat(mcp-server): add create_knowledge_from_text tool by @01luyicheng in #2532
• fix: retry document summaries before fallback by @ttommybot in #2505
• fix: replace all exact wiki text matches by @quan020406 in #2516
• feat(im): add markdown replies for QQBot #2461 by @YS-zdck in #2483
• chore(deps): bump the cli-deps group in /cli with 3 updates by @dependabot[bot] in #2481
• feat(feishu-driver): 新增飞书云盘数据源 by @Greenplumwine in #2466
• fix(wiki): exclude archived pages from stats queries by @pengpengyi92 in #2549
• fix(im/wecom): resolve longconn deadlock between closeConn and heartbeatLoop by @yangcongcong-coding in #2551
• fix(chunks): 恢复未编辑分块的按位置合并 by @mdrkrg in #2555
• fix(im/wecom): scope connection close to its own generation by @lyingbug in #2557
• fix(chunks): 受信任分块按精确位置裁剪重叠 by @lyingbug in #2556
• fix(docs): correct @router path annotations for agent-chat and knowledge-chat by @qfzc in #2561
• fix(chunks): 流水线重写后的检索结果不用位置匹配 by @mdrkrg in #2558
• docs: add comprehensive WeKnora documentation by @lyingbug in #2350
• docs: 部署步骤增加 docker compose pull，防止复用缓存的旧镜像 by @WHUTcjh-2024 in #2567
• docs: 补全 docker compose pull 部署与升级说明 by @lyingbug in #2570
• Revert "docs: 补全 docker compose pull 部署与升级说明" by @lyingbug in #2571
• Revert "docs: add comprehensive WeKnora documentation" by @lyingbug in #2572
• update website docs by @lyingbug in #2573
• docs: complete docker compose pull guidance across deployment docs by @lyingbug in #2574
• docs(website): add official site link and inject VERSION at build time by @lyingbug in #2576
• fix(auth): do not sanitize registration password before hashing by @jiahao6635 in #2568
• docs(release): prepare v0.7.2 release notes and version bump by @lyingbug in #2585

New Contributors

• @BigFishDreamWater made their first contribution in #2329
• @koishi514-Z made their first contribution in #2318
• @mdrkrg made their first contribution in #2319
• @FincherYuan made their first contribution in #2375
• @annopick made their first contribution in #2371
• @WHUTcjh-2024 made their first contribution in #2314
• @zldkr made their first contribution in #2419
• @ZhuangShiHao1207 made their first contribution in #2422
• @asmh1989 made their first contribution in #2421
• @lazyayuan made their first contribution in #2359

Note truncated.

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 docs/api/README.md.
• 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.