RWiki: A Lightweight Self-Hosted RAG Knowledge Base Q&A System

To solve the need for a lightweight internal knowledge base Q&A system, I built RWiki.

After looking around, I found that existing solutions were either cloud services (data not in your hands) or open-source projects with overwhelming dependencies. What the team needed was simple: throw in some documents, ask questions in natural language, and get answers with sources — not a complex microservices architecture.

So I built RWiki. The entire project was developed using T-Tools AI, taking less than a week and consuming over 400 million tokens.

What It Is

RWiki is a RAG (Retrieval-Augmented Generation) based knowledge base Q&A system. You upload documents, and it handles splitting, vectorization, and storage. Users can then ask questions in natural language — the system retrieves relevant content from the documents and generates answers with source citations via an LLM.

The core design goal is simple: keep it as simple as possible.

Single binary, no external dependencies
SQLite for all data storage, including vector indexes (via sqlite-vec)
Supports multiple LLM providers: OpenAI, OpenRouter, Zhipu AI, and more
Comes with an embeddable chat widget for any website

Notable Features

Document Processing

Supports three formats: Markdown (with frontmatter), Excel (.xlsx), and OpenAPI specs. Markdown works well for technical docs and knowledge bases, Excel for structured data, and OpenAPI for API documentation assistants. Uploaded files are automatically split, embedded, indexed, and made searchable once published.

Conversation Experience

Chat uses streaming (SSE) for a natural typing effect. Multi-turn conversations are supported with automatic query rewriting — follow-up questions are understood in their full semantic context. Conversation history uses a sliding window with automatic compression, so token usage won't blow up over long sessions.

Answers include source citations in the format [Source: Title / Section], with links to original documents when available. This is crucial for knowledge base scenarios — users need to be able to trace information back to its origin.

Embeddable Chat Widget

Beyond the built-in admin frontend, RWiki provides rwiki-chat.js — a chat widget you can embed in any webpage with just two lines of HTML. It uses Shadow DOM for style isolation, so it won't conflict with the host page's styles.

Multi-LLM Provider

Simply specify base_url and model in the config file — it's compatible with any OpenAI API format provider. You can use DeepSeek, Claude, GPT-4o, or Zhipu's GLM series. The same applies to embedding models — currently tested with OpenAI and Zhipu's Embedding-3.

Quick Start

The easiest way is Docker:

docker run -it --rm \
  -p 8080:8080 \
  -e LLM_API_KEY=your-api-key \
  -v rwiki-data:/app/data \
  ghcr.io/timzaak/rwiki:latest

Then visit http://localhost:8080 to see the interface.

To run locally, download the binary for your platform from GitHub Releases, prepare a config file, and run it directly.

Tech Stack

The backend is written in Rust (Axum + SQLite + rig-core), with a React + TypeScript frontend. Rust wasn't chosen to show off — it was chosen because:

Compiles to a single binary, easy to deploy
Low memory footprint, runs on a 1C1G machine
Mature SQLite ecosystem in Rust, sqlite-vec handles vector search well enough

Honestly, for most small-to-medium knowledge base scenarios (hundreds to thousands of documents), SQLite + sqlite-vec performs perfectly fine. No need to reach for Qdrant or Milvus right away.

Current Status

This is the first official release (v0.1.1), with core features ready to use:

Document upload, parsing, vectorization, and publishing
RAG retrieval + LLM streaming Q&A
Multi-turn conversation + source citations
Embeddable chat widget
OpenTelemetry observability support

Upcoming plans: more document format support, finer-grained access control, and a better frontend experience.

Wrapping Up

RWiki solves a small problem: "I have a bunch of documents and want to ask questions in natural language." But this need is quite common — internal team docs, product FAQs, API documentation assistants, personal note collections...

Want to try it out? Visit rwiki.fornetcode.com to test document upload and Q&A online.

If this fits your needs, give it a try. Feel free to open an issue if you run into problems, and stars are always welcome.

Project: https://github.com/timzaak/rwiki Live Demo: https://rwiki.fornetcode.com