Kotaemon — Open-Source RAG Document Chat

What it is

Kotaemon is an open-source RAG (retrieval-augmented generation) tool that lets you chat with your documents. Upload PDFs, DOCX files, or web pages, and ask questions in natural language. Kotaemon retrieves relevant passages and generates answers with citations pointing back to the source documents. It supports multiple LLM providers and can be self-hosted.

It targets researchers, analysts, and knowledge workers who need to extract information from large document collections without reading everything manually.

How it saves time or tokens

Kotaemon handles the full RAG pipeline internally: document parsing, chunking, embedding, vector storage, retrieval, and answer generation with citations. Instead of building this stack from individual components, you run a single application. The citation feature is particularly valuable -- you can verify every answer against the source document.

How to use

1. Install and run:

pip install kotaemon
python -m kotaemon

Or with Docker:

docker run -p 7860:7860 ghcr.io/cinnamon/kotaemon:latest

1. Open http://localhost:7860.
2. Configure your LLM provider (OpenAI, Anthropic, Ollama) in Settings.
3. Upload documents and start asking questions.

Example

User: What are the main risks identified in the annual report?

Kotaemon: The report identifies three main risks:
1. Currency fluctuation exposure in Asian markets [page 12]
2. Supply chain disruption from single-source dependencies [page 15]
3. Regulatory changes in data privacy requirements [page 23]

[Click citations to view source passages]

Each answer includes clickable citations that link to the exact source passages.

Related on TokRepo

• AI tools for RAG -- RAG tools and frameworks
• AI tools for documents -- document processing tools

Common pitfalls

• Document parsing quality varies by file type. PDFs with complex layouts (multi-column, tables, scanned images) may not parse correctly. Pre-process problematic PDFs with an OCR tool for better results.
• Embedding model choice affects retrieval quality. The default embedding model works for general text. For specialized domains (legal, medical), consider a domain-specific embedding model.
• Large document collections increase storage and retrieval latency. For hundreds of documents, ensure adequate disk space and consider using a more performant vector store backend.