WeKnora Quick Start — From Zero to RAG-Powered Q&A in 15 Minutes

🧪 WeKnora Quick Start: From Zero to RAG-Powered Q&A in 15 Minutes

I'll walk you through setting up WeKnora on your own machine — Docker Compose, first document upload, first Q&A. No prior RAG experience needed. Grab a coffee, this takes about 15 minutes.

Prerequisites

• Docker & Docker Compose installed
• At least 4GB RAM (I'd recommend 8GB for production)
• A few PDF or Word documents to test with
• An OpenAI API key (or any of the 20+ supported providers)

Step 1: Clone & Configure

git clone https://github.com/Tencent/WeKnora.git
cd WeKnora
cp .env.example .env

Now edit .env. These are the minimum fields you need to set:

# LLM provider (I used OpenAI, but DeepSeek/Qwen work great too)
LLM_API_KEY=sk-your-key-here
LLM_MODEL=gpt-4o-mini

# Database
POSTGRES_PASSWORD=your-strong-password

💡 Why I chose gpt-4o-mini: WeKnora uses LLMs for both embedding and answer generation. The mini model is fast enough for Q&A and costs way less. For complex multi-step agent reasoning, you might want the full gpt-4o.

Step 2: Start WeKnora

docker compose up -d

This spins up: the main app, PostgreSQL + pgvector, Redis, and MinIO (for document storage). The first startup takes about 2-3 minutes — the frontend needs to build its assets.

Wait for: docker compose logs -f app | grep "listening" — you'll see the server is ready on http://localhost:8080.

Step 3: Upload Your First Documents

Open http://localhost:8080 in your browser.

1. Create a workspace — give it a name like "Test KB"
2. Upload documents — drag-and-drop PDFs, Word docs, or images. I tested with a 50-page API spec PDF
3. Wait for indexing — you'll see a progress bar. A 50-page PDF took about 10 seconds on my machine
4. Ask a question — type something like "What endpoints does the API expose?"

If you see a well-formatted answer with citation numbers — congrats, it works! 🎉

Step 4: Try Agent Mode

Click the "Agent" toggle in the top-right of the chat interface. Now ask something that requires multi-step reasoning:

"Find all mentions of rate limiting in the uploaded documents, check if they follow REST API best practices, and summarise what needs updating."

The agent will: retrieve relevant chunks, analyse them, potentially search the web for best practices, and compose a structured response. I tried this and got back a 3-point action plan with document references. Honestly, I was expecting a half-baked summary, not something I could forward to a colleague.

Step 5: Connect IM Channels (Optional but Cool)

WeKnora supports Slack, Telegram, WeCom, Feishu, DingTalk, and Mattermost. To connect Slack:

# In .env, add:
SLACK_BOT_TOKEN=xoxb-your-token
SLACK_SIGNING_SECRET=your-secret

docker compose restart app

Now your team can ask questions inside Slack and get answers directly. No more "let me check and get back to you."

What I Wish I Knew Before Starting

• The docker compose.dev.yml is NOT for production — use the main docker-compose.yml
• If you want to use Ollama for local LLMs, set LLM_PROVIDER=ollama and point it to your Ollama host
• WeKnora stores uploaded files in MinIO by default — configure S3-compatible storage if you need redundancy
• The wiki mode needs ~500MB of RAM extra for the graph generation agent

Final Verdict

WeKnora is surprisingly mature for a v0.6.x project. The three-mode design (RAG → Agent → Wiki) covers a genuine spectrum of needs, and the enterprise features (RBAC, audit, multi-tenant, 20+ LLM providers) aren't just checkboxes — they work well. The Docker setup is straightforward, the UI is modern (with a ⌘K command palette!), and the IM integrations make it disappear into your workflow.

Is it overkill for a small team with 50 documents? Probably. But if you're dealing with thousands of documents across multiple sources and you want a self-hosted knowledge platform that thinks as well as it retrieves — this is the one.