Tencent/WeKnora

A self-hostable, Go-based LLM knowledge platform that turns documents into RAG-based Q&A, an autonomous ReAct agent, and a self-maintaining wiki.

What it is

WeKnora is an open-source knowledge framework for document understanding, semantic retrieval, and autonomous reasoning. It ingests documents from multiple sources and exposes them through three modes: RAG-based quick Q&A, a ReAct agent that orchestrates retrieval, MCP tools, and web search for multi-step tasks, and a Wiki mode that distills raw documents into interlinked markdown with a knowledge graph. It targets teams that need a self-hosted, multi-tenant knowledge assistant with control over their data.

Key features

• Three interaction modes: RAG Q&A, a ReAct agent for multi-step reasoning, and agent-generated Wiki pages with a knowledge graph.
• Multi-source ingestion from Feishu, Notion, Yuque, and RSS, and support for 10+ document formats including PDF, Word, Excel, images, EPUB, and MHTML.
• Swappable backends: 20+ LLM providers, multiple embedding providers, and vector stores including pgvector, Elasticsearch, OpenSearch, Milvus, Weaviate, Qdrant, Apache Doris, and Tencent VectorDB.
• Multiple delivery surfaces: Web UI, REST API, the weknora CLI, a Chrome extension, an embed widget, and a WeChat Mini Program, plus IM channels such as WeCom, Feishu, Slack, Telegram, DingTalk, and Mattermost.
• Enterprise controls: workspace RBAC with a four-tier role matrix, per-resource ownership, per-workspace audit log, scoped API keys, and AES-256-GCM encryption for credentials.
• Retrieval strategies including BM25 sparse, dense, GraphRAG, parent-child chunking, and HNSW-accelerated pgvector, with Langfuse observability and a runtime task-queue dashboard.

Tech stack

• Go 1.26.0 (module github.com/Tencent/WeKnora).
• Gin for the web layer and GORM over PostgreSQL/SQLite; pgvector for vector storage.
• asynq with Redis (go-redis) for the async task queue; OpenTelemetry for tracing.
• Provider and vendor SDKs including go-openai, the Lark/Feishu SDK, Milvus, Qdrant, Weaviate, Elasticsearch v7 and v8, OpenSearch, the Neo4j driver, and AWS/Alibaba/Tencent/Volcengine/Huawei/Kingsoft storage SDKs.
• mcp-go for MCP support and gojieba for tokenization; a Wails dependency is also present.
• Deployment via Docker Compose (profiles for neo4j, minio, langfuse, and full) and Kubernetes/Helm. Version 0.7.0.

When to reach for it

• You need a self-hosted RAG plus agent knowledge base with full data sovereignty.
• You want to swap LLMs, vector databases, and object storage without rebuilding your stack.
• You need multi-tenant workspaces with RBAC and audit logging.
• You want to serve knowledge Q&A through IM channels or embed it into external websites.

When not to reach for it

• You want a lightweight, single-file library — this is a full multi-service platform requiring Docker or Kubernetes, Redis, and a vector database.
• You do not need document ingestion or RAG — a plain LLM API wrapper is simpler.
• You lack the infrastructure to operate the required services.
• Your team reads only English — some deeper guides (for example the development guide and RBAC documentation) are Chinese-only.

Maturity signal

Backed by Tencent, created in July 2025, WeKnora reached roughly 18,700 stars within a year, was last pushed in July 2026, and is not archived, with a fast release cadence that moved from v0.2 to v0.7 while adding substantial features at each step. It remains pre-1.0 (v0.7.0), so ongoing change should be expected, and the 434 open issues are consistent with heavy adoption. The README labels the license MIT, but GitHub's detected SPDX is NOASSERTION, so the LICENSE file is worth verifying before reuse.

Alternatives

• RAGFlow — use it when you want a RAG engine centered on deep document parsing.
• Dify — use it when you want a broader visual LLM-app builder rather than a knowledge-base-first platform.
• LangChain or LlamaIndex — use them when you want to assemble a custom RAG pipeline as a library instead of deploying a full platform.

Notes

• The README states the license is MIT while GitHub reports NOASSERTION, a mismatch worth confirming.
• Langfuse is the sole tracing backend; the changelog notes Jaeger was removed in v0.6.2.
• WeKnora is described as the core framework behind Tencent's WeChat Dialog Open Platform.
• A Wails dependency in go.mod suggests a desktop build path alongside the web UI.

Tags

golang, rag, llm, knowledge-base, vector-search, semantic-search, agent, self-hosted, multi-tenant, chatbot, reranking