WeKnora: Tencent’s Open Source RAG and Agent Knowledge Platform

1. What WeKnora Is

The GitHub description says it all: “Open-source LLM knowledge platform: turn raw documents into a queryable RAG, an autonomous reasoning agent, and a self-maintaining Wiki.” It is a full product rather than a library. Per the README overview, you use RAG to look things up, the agent for multi-step tasks and the wiki to organize knowledge — all on the same knowledge bases, behind one web UI, with workspaces, roles, API keys and channels around them.

If you read my OpenRAG post, here is the difference in one paragraph: OpenRAG packages Langflow, Docling and OpenSearch behind a single uvx command and gives you visual flows to edit. WeKnora is a heavier, more product-shaped option for the same job — a Go backend and Vue UI with multi-tenant workspaces and RBAC, an agent with skills and a sandbox, an auto-generated wiki, and connectors for WeCom, Feishu, Slack and more. Same “chat with your documents” starting point, very different weight class.

2. The Architecture in 4 Steps

The README’s architecture note says it is a modular pipeline from document parsing and vectorization through retrieval to LLM inference, with every component replaceable. The v0.8.3 architecture page (in Chinese) fills in the detail:

3. Quickstart: From Clone to Login

Prerequisites per the README: Docker, Docker Compose and Git. These are the README’s own commands, verified against the README on 2026-10-14 (I did not execute them):

git clone https://github.com/Tencent/WeKnora.git
cd WeKnora
cp .env.example .env    # Edit .env as needed, see comments in the file
docker compose pull     # Pull the latest images
docker compose up -d    # Start core services

Then open http://localhost and follow the onboarding guide. The README’s service table lists the Web UI on http://localhost, the backend API on http://localhost:8080 and, only with the langfuse profile, Langfuse tracing on http://localhost:3000. The Chinese quickstart page adds a health check and the first-login rule: there is no default account — the first person to register gets a personal workspace and becomes its Owner.

curl http://localhost:8080/health

You still need models. The quickstart docs say you need a chat model and an embedding model, local Ollama or any OpenAI-compatible service, configured per knowledge base in an initialization wizard (use its test button before saving); changing the embedding model later means rebuilding the index. For a local Ollama, the README’s tip is to start it first, and .env.example points OLLAMA_BASE_URL at host.docker.internal:11434 because localhost inside a container is the container itself:

ollama serve > /dev/null 2>&1 &

Want graph retrieval or object storage? Profiles are combinable; the README lists full (everything), neo4j, minio and langfuse:

docker compose --profile neo4j --profile minio pull
docker compose --profile neo4j --profile minio up -d
docker compose down     # Stop services

Upgrading has one trap the README calls out: docker compose up -d alone can reuse cached images and leave the UI out of sync with the release, so pull first. It also says to read the v0.8.3 upgrade notes before upgrading, and the v0.8.2 notes if you are coming from an older version.

# Set WEKNORA_VERSION in .env to the target release (e.g. 0.8.3), or keep latest
docker compose pull     # Pull images matching WEKNORA_VERSION
docker compose up -d    # Recreate containers with new images

Prefer scripting to clicking? The README describes an “agent-first” weknora CLI (JSON output by default, --format text for humans). These are its example commands; I did not verify how to install the CLI, so start from cli/README.md in the repo:

weknora profile add prod --host https://kb.example.com --use
weknora auth login
weknora kb list
weknora link --kb my-knowledge-base
weknora doc upload notes.md
weknora chat "summarise the design doc"

4. Before You Open a Port: Security Defaults

Right under the deployment options, the README carries a warning block. I am paraphrasing it closely: WeKnora ships with login authentication, but for production the maintainers strongly recommend that you:

That advice matters because of what the compose file and .env.example actually do out of the box:

5. Where It Shines: Use Cases

WeKnora earns its weight when knowledge has to be shared, permissioned and kept fresh across a team:

The pattern across all six: one deployment, many users, knowledge that keeps changing — not one developer who needs quick semantic search.

6. When NOT to Use It

WeKnora is impressive, but it is not the right default for every RAG job. Honest boundaries:

Conclusion

WeKnora is a rare thing among trending RAG repos: a product with permissions, channels, evaluation and tracing, not only a pipeline demo. About 34K stars in roughly 15 months says teams want exactly that. The price is weight — a multi-service stack, a 0.x release train and docs that live mostly in Chinese.

My suggestion: read the README’s security note first, run the core stack on a private network with registration locked down, index one real corpus and compare its answers with your current setup — measure recall instead of guessing (my RAG evaluation post shows how). If all you need is quick semantic search, OpenRAG or pgvector gets you there with fewer moving parts.