WeKnora: Open-Source Knowledge Base RAG Platform
/WeKnora
WeKnora is worth deploying if you need a self-hosted RAG system that goes beyond plain retrieval. The combination of ReAct agents, live chunk editing with rollback, and auto-generated wikis covers ground that simpler RAG SaaSes skip entirely. It loses points on setup friction and a dense feature set that takes time to absorb, but once running it handles enterprise multi-workspace workloads without obvious gaps. Worth it once you actually need the agent+wiki trifecta on your own infrastructure.
What is WeKnora
WeKnora is an open-source knowledge base RAG platform that turns raw documents into queryable knowledge assets. It combines RAG-based Q&A, a ReAct agent that orchestrates retrieval and web search across multi-step tasks, and a Wiki Mode where agents auto-generate structured interlinked Markdown pages with an interactive knowledge graph. Built in Go with a fully modular pipeline—swap LLMs, vector stores, and storage backends freely. Supports 20+ LLM providers, auto-syncs from Feishu/Notion/Yuque, and runs fully self-hosted.
Core capabilities: RAG, agents, and auto-generated wikis
- ✓RAG-based Quick Q&A with hybrid BM25 + dense retrieval, HNSW-accelerated pgvector at 1024 dimensions, configurable citation output, and RAG pipeline progress visible in chat
- ✓ReAct Agent for multi-step autonomous reasoning—orchestrates knowledge retrieval, MCP tools, and web search in sequence; scoped runtime via @Skill/@MCP mentions per turn
- ✓Wiki Mode: agents auto-generate structured interlinked Markdown pages from raw documents, with an interactive knowledge graph browser, line-level diffs, and one-click rollback
- ✓Chunk editing with per-version snapshots, diff, and rollback; automatic reindexing after edits; custom document metadata and generated Q&A per chunk
- ✓Multi-source ingestion with auto-sync from Feishu wiki/Feishu Drive, Notion, Yuque, and RSS feeds; supports PDF, Word, Excel, images, EPUB, and more
- ✓Enterprise controls: multi-workspace RBAC (Owner/Admin/Contributor/Viewer), scoped API keys with per-KB restrictions, runtime task-queue dashboard with worker-pool governance, Langfuse observability
What teams use WeKnora for
- •Internal knowledge bases that need both fast RAG lookups and a browsable wiki non-technical teams can navigate and edit directly
- •Multi-tenant SaaS or enterprise environments requiring workspace isolation, RBAC, audit logs, and self-hosted data residency
- •Teams running autonomous agents that must orchestrate document retrieval with MCP tools and live web search across complex queries
- •Organizations already invested in Feishu, Notion, or Yuque who want a unified RAG layer across those sources without migration
- •Engineering teams that need scoped API keys and observability into token usage and agent reasoning traces for cost governance
Deploying WeKnora locally or in your cloud
Deploy via Docker Compose for a single-node setup, or Helm for Kubernetes clusters. The Web UI handles initial onboarding, while `weknora` CLI manages profiles, authentication, and MCP serving. Private cloud and offline deployments are supported with no external dependencies beyond your chosen LLM and vector store. AWS S3, Alibaba OSS, Huawei OBS, and MinIO are all supported as storage backends, and multiple storage instances can run per workspace with per-KB binding.
Building a knowledge base and querying it
Upload documents through the Web UI, folder drag-and-drop, or CLI; source materials are parsed, chunked with preview, and indexed to your chosen vector store. Switch between Quick Q&A (pure RAG), Agent Mode (ReAct reasoning), or Wiki Mode (auto-generate browsable pages). Ask questions via chat, IM channels (WeCom, Feishu, Slack, Telegram), embed widget, or REST API. The chunk editor lets you fix retrieval quality manually, and revision history tracks every change. Scoped API keys let you integrate programmatically without exposing admin credentials.
What works well in WeKnora
- ✓Three distinct interaction modes—Quick Q&A, ReAct Agent, and Wiki Mode—handle different cognitive workloads in one deployment
- ✓Live chunk editing with diff and rollback closes the loop between retrieval failures and knowledge base quality
- ✓Auto-generated wikis with a visual knowledge graph give non-technical users a readable, navigable interface into otherwise opaque RAG output
- ✓Multi-workspace RBAC and scoped API keys are genuinely enterprise-grade, not bolted on
- ✓Langfuse tracing ships out of the box for token tracking, agent reasoning visualization, and pipeline debugging
- ✓Highly modular: swap OpenAI for Ollama, pgvector for Qdrant, or S3 for MinIO without platform migration
Known limitations and constraints
- △No official hosted cloud offering documented—self-hosting is the only path, which means ops overhead
- △The feature density is high; the learning curve is real for teams expecting a plug-and-play RAG tool
- △Chunk editing and Wiki Mode revision history add maintenance surface area that smaller teams may not need
- △Multi-instance storage, per-workspace audit logs, and RBAC complexity assume an admin comfortable with platform governance
- △Not clearly documented: precise SLA or uptime guarantees for the worker-pool task queue under high concurrent load
- △OSS LLM support via Ollama is functional but the README does not specify minimum Ollama version or GPU memory requirements
How WeKnora compares to similar platforms
Who should and should not deploy WeKnora
Deploy WeKnora if you run a mid-to-large team that needs a self-hosted knowledge platform with real retrieval quality controls, autonomous agent reasoning, and a browsable wiki your non-technical staff can maintain. It is built for organizations with compliance or data-sovereignty requirements that can't use hosted SaaS. Skip it if you run a small team wanting a quick RAG demo—simpler tools exist with far less setup. Also skip it if you lack someone comfortable administering a Go backend, managing worker pools, and tuning RBAC policies, because the operational surface is wider than a typical SaaS tool.
Frequently asked questions
WeKnora integrates with OpenAI, Azure OpenAI, Anthropic Claude, DeepSeek, Qwen (Alibaba Cloud), Zhipu, Hunyuan, Doubao, Gemini, MiniMax, NVIDIA, Novita AI, SiliconFlow, OpenRouter, Requesty, and Ollama. Embeddings come from Ollama, BGE, GTE, Zhipu, and OpenAI-compatible APIs.
Yes — WeKnora supports auto-sync from Feishu wiki, Feishu Drive, Lark (Feishu International), Notion, and Yuque. RSS feed ingestion is also available. Incremental and full sync modes are documented.
WeKnora includes workspace RBAC with Owner/Admin/Contributor/Viewer tiers, per-KB resource ownership, per-workspace audit logs, scoped API keys with capability-level grants, and a runtime task-queue dashboard with per-model concurrency governors. The architecture is designed for multi-tenant operation, though the team should have ops capacity for self-hosting.
Beyond pgvector, WeKnora supports Elasticsearch, OpenSearch, Milvus, Weaviate, Qdrant, Apache Doris, and Tencent VectorDB. Retrieval strategies include BM25 sparse search, dense retrieval, GraphRAG, parent-child chunking, and HNSW-accelerated indexing.
WeKnora stores each upload's directory structure as a folder tree. Retrieval chunks are editable directly in the UI with per-version snapshots, line-level diff, and one-click rollback. After any edit, the index reindexes automatically. Custom document metadata and generated Q&A per chunk are also supported.
Yes — WeKnora ships an official MCP server (PyPI package `tencent-weknora-mcp`) with 29 tools across stdio, SSE, and HTTP transports. The MCP server supports OAuth2 for remote services, mid-conversation OAuth, and scoped agent runtime via @MCP mentions. A DeepSeek Harness plugin (`@wxg-prc-cpg/dsh-weknora`) is also available.
The problem it solves
Most RAG platforms stop at "ingest documents, retrieve chunks, answer questions." They leave you with a retrieval layer, not a knowledge platform—no way to refine what gets indexed, no agent reasoning that can call external tools, and no structured output your team can browse and maintain over time. Teams either duct-tape multiple tools together or live with a black-box that nobody can audit or edit.
Related repositories
Is WeKnora worth your time?
ChatGPT, Claude and Perplexity can all read this page. Ask one of them what it makes of WeKnora.