WeKnora Deep Research: Three Modes in One, with a Self-Maintaining Knowledge Loop
> Turn scattered documents into a library that reorganizes its own shelves — and writes its own encyclopedia.
>
> Subject: Tencent/WeKnora (Tencent's open-source LLM knowledge platform — RAG Q&A, ReAct Agent, AI Wiki in one)
> Research date: 2026-08-16 | Method: shallow clone (3,067 files) + static reading of README/CHANGELOG/ROADMAP/docs + external sources. Claims are marked as code-verified, doc-verified, or inferred.
Key points
- Three core modes (doc-verified): 1. RAG Quick Q&A — retrieval-augmented generation with BM25 + dense + knowledge-graph (GraphRAG) hybrid retrieval; answers are auto-attributed to sources. 2. ReAct Agent — multi-step autonomous reasoning orchestrating knowledge retrieval, web search, MCP tools, and a code sandbox, with optional human-approval gates. 3. Wiki mode (GA since v0.5.0) — the agent distills raw documents into a self-maintaining, cross-linked Markdown knowledge base with interactive knowledge graph, revision history, and one-click rollback.
- The three modes form a closed loop: the Agent can query the RAG store *and* edit the Wiki (17 Wiki tools); the Wiki in turn feeds retrieval.
internal/agent/engine.goimplements a standard ReAct loop (think → analyze → act → observe), stateless across turns, with token-budgeted context management, LLM summarization memory consolidation, skill progressive disclosure, and an event bus.internal/sandboxis production-grade: Tencent Cube, E2B, Docker, and local backends; Redis session binding for multi-instance; orphan sandbox reaper; SSRF guard.- Tool surface for the agent includes knowledge search, graph queries, web fetch/search, 17 Wiki edit tools, memory search, shell execution, DuckDB data analysis, and MCP invocation.
- 10 retrieval backends: postgres (pgvector), Elasticsearch v7/v8, OpenSearch, Milvus, Qdrant, Weaviate, Tencent VectorDB, Doris, Neo4j (knowledge graph), SQLite (Lite) — rare breadth in open-source RAG.
- Object storage: local, MinIO, S3, COS, TOS, OSS, KS3, OBS.
- 20+ LLM providers via OpenAI-compatible protocol (OpenAI, Azure, Anthropic, DeepSeek, Qwen, Zhipu, Hunyuan, Gemini, Ollama, etc.); BGE/GTE embeddings and rerankers.
- Async governance: asynq-based tiered worker pools (core/post-process/enrichment/maintenance) with per-model concurrency governors; Wiki generation gets a dedicated pool.
- Security: AES-256-GCM, unified SSRF-safe HTTP transport, gRPC/Redis TLS, 4-level RBAC with audit logs, OIDC.
autoprofiles the document (Markdown headings, page breaks, section markers) and picks the strongest strategy;headingsplits on#/##/###with breadcrumb-prefixed embeddings;heuristicuses page-break/numbering cues;legacyrecursive splitting as fallback.- Parent-child chunking: small child chunks (384 chars) for vector matching, large parent chunks (4,096 chars) fed to the LLM.
- Grounded in the Vecta Feb-2026 benchmark (recursive ~512-token/15% overlap ≈ 69% end-to-end accuracy).
- A read-only chunking preview in the KB editor makes chunking debuggable as a first-class citizen.
- Fast release cadence: v0.7.0 → v0.7.1 → v0.7.2 within three weeks (July–August 2026); 79+ DB migrations, ~50 doc pages, ~360 API endpoints, ~150 env vars.
- ⚠️ Bus factor ≈ 1: top contributor
lyingbughas 1,803 commits vs 128 for #2. - ⚠️ Doc drift:
mcp-server/PROJECT_SUMMARY.mdstill says 21 tools / v1.0.0 while CHANGELOG says 29 tools / v1.1.x; ROADMAP has many unchecked items. - ⚠️ License misreporting: several third-party articles wrongly claim Apache-2.0; it is MIT — a plus for commercial use.
- ⚠️ Stale reviews: some mid-2026 tests on v0.6.2 claiming no Excel/PPT support are outdated for v0.7.2 (10+ formats now supported), though complex-table fidelity still depends on the underlying parser.
- ⚠️ Heavy dependencies: Go 1.26 toolchain, cgo + prebuilt Rust libs (musl/gnu) affecting offline/ARM builds, Neo4j required separately for the knowledge graph.
- ℹ️ Memory trade-off: v0.7.1 removed Neo4j session/episodic memory in favor of LLM-summary consolidation, keeping Neo4j only for the knowledge graph.
- Shallow clone (depth 1): commit history reconstructed via GitHub API + CHANGELOG; no full end-to-end benchmarks were run; comparisons rely on public positioning of competitors; some third-party evaluations used older v0.6.x versions (corrected where possible).
- Repo: https://github.com/Tencent/WeKnora | Site: https://weknora.weixin.qq.com
- Announcement: https://fuwu.weixin.qq.com/community/develop/article/doc/0002e077cf8da8c1bfb36bff663c13
CHANGELOG.md,docs/ROADMAP.md,docs/CHUNKING.md,docs/KnowledgeGraph.md,docs/worker-pool-governance.md,mcp-server/- PyPI:
tencent-weknora-mcp - GitHub API ground truth (2026-08-16): 19,911 stars / 2,863 forks / 546 open issues / Go / MIT / created 2025-07-22
What it is
| Item | Detail | | --- | --- | | Positioning | Open-source LLM knowledge engine/framework | | Origin | Core doc-understanding/retrieval framework of Tencent's WeChat Dialogue Open Platform; open-sourced ~2025-08 (repo created 2025-07-22) | | Current version | v0.7.2 (2026-08-07) | | License | MIT (often misreported as Apache-2.0; LICENSE confirms MIT) | | Stack | Go 1.26 backend; Vue/TypeScript frontend; Rust document parser (anydoc) via cgo; Python MCP server; WeChat mini-program | | Scale | 3,067 files; GitHub ~19,911 stars / 2,863 forks / 546 open issues (2026-08-16) | | Deployment | Docker Compose (profiles), Helm/K8s, Lite edition, offline; Web UI, REST API, CLI, Chrome extension, mini-program |
Architecture (code-verified from internal/)
Clean layered architecture: access layer (router/handler, 9 IM channels incl. Slack, Telegram, WeCom, DingTalk, Lark) → application layer (repository/service, DI container) → capability layer (agent engine, MCP, sandbox, model registry for chat/embedding/rerank/VLM/ASR) → infrastructure (chunker, docparser, web fetch/search, connectors for Feishu/Notion/Yuque/RSS, 10 retriever backends) → cross-cutting (Langfuse tracing, async queues, SSRF guards, rate limiting).
Technology breadth
Adaptive three-tier chunking (doc-verified)
Differentiators
1. Retrieval backend breadth (10 backends).
2. Multi-backend sandbox with orphan reaping.
3. Bidirectional MCP: consumes external MCP tools (with in-chat OAuth) and *provides* 29 tools via the PyPI package tencent-weknora-mcp for Claude/Cursor/n8n etc.
4. Context/memory engineering: request-scoped res:// alias compression, LLM-summary memory consolidation.
5. Platform-level governance: RBAC, scoped API keys, OIDC, shared spaces.
Comparison
| Dimension | WeKnora | Dify | RAGFlow | FastGPT | | --- | --- | --- | --- | --- | | License | MIT | Apache-2.0 | SSPL | Apache-2.0 | | Doc parsing | anydoc/MinerU/OCR, modular | Third-party prep | DeepDoc (best-in-class layout/table) | Basic | | Retrieval backends | 10 | Limited | Built-in | Limited | | Agent | ReAct + self-maintaining Wiki + sandbox | Workflow orchestration | Weak | Graph orchestration | | App layer | Build your own | Turnkey | Complete | Turnkey | | Knowledge graph | Neo4j GraphRAG | External | Yes | No |
Positioning (inferred): WeKnora is a knowledge-engine kernel, not a turnkey app platform. RAGFlow still leads in deep document layout parsing; Dify is an app-orchestration suite where WeKnora can serve as the knowledge-recall kernel.
Maturity, risks, and caveats
Study limitations
Conclusion
WeKnora's real novelty is not any single technique but closing the retrieval → reasoning → consolidation loop: the Agent queries the knowledge base, writes the Wiki, and the Wiki feeds back into retrieval. It is a deeply modular, MIT-licensed knowledge-platform kernel — but the application layer, documentation, and maintenance currently carry real adoption costs (single-core maintenance, early ROADMAP, heavy toolchain). Choose it if you want full control over ingestion and reasoning pipelines; choose Dify/RAGFlow if you need a turnkey product.