Cinnamon/kotaemon

An open-source, self-hosted RAG web UI for chatting with your documents, aimed at both end users doing document QA and developers building their own retrieval pipelines.

What it is

Kotaemon is a retrieval-augmented-generation application and framework for asking questions over your own documents. It serves two audiences: end users who want a clean question-answering interface over their files, and developers who want to build and customize a RAG pipeline with a working UI already attached. The UI is built with Gradio, and the codebase is split into a reusable library (kotaemon) and the application layer (ktem). It supports both hosted LLM providers and local models.

Key features

Self-hosted document QA web UI with multi-user login and private/public collections for organizing and sharing files and chats.
Support for both API providers (OpenAI, Azure OpenAI, Cohere, Groq) and local LLMs via ollama and llama-cpp-python.
A hybrid RAG pipeline combining full-text and vector retrieval with re-ranking.
Multi-modal QA over documents with figures and tables, using selectable document parsers.
Advanced citations shown in an in-browser PDF viewer with highlights and relevance scores, plus warnings when retrieved passages are low-relevance.
Complex reasoning via question decomposition and agent-based methods (ReAct, ReWOO), and an extensible indexing layer with GraphRAG provided as an example.

Tech stack

Python >=3.10; Apache-2.0 licensed.
UI built on Gradio; project structured as a uv workspace with two packages under libs/ (kotaemon and ktem), using setuptools with setuptools-git-versioning.
Configurable document stores (Elasticsearch, LanceDB, SimpleFileDocumentStore) and vector stores (ChromaDB, LanceDB, InMemory, Milvus, Qdrant), set via flowsettings.py.
Optional multimodal parsing/OCR through Azure Document Intelligence, Adobe PDF Extract, Docling, PaddleOCR, and Unstructured for non-PDF file types.
GraphRAG indexing options: official MS GraphRAG, NanoGraphRAG, and LightRAG.
Distributed as Docker images on GHCR (lite, full, and ollama variants) for linux/amd64 and linux/arm64, plus a pip install -e path.

When to reach for it

You want a self-hosted UI for document QA that non-developers can use out of the box.
You need multi-user access with organized private/public document collections.
Verifiable answers matter and you want inline citations with a PDF viewer and relevance scores.
You are a developer who wants a customizable RAG pipeline with a ready-made UI to iterate against.
You want local, private RAG using Ollama or GGUF models.

When not to reach for it

You need a headless RAG library only; the Gradio UI is central to the project rather than optional.
You are targeting very high-throughput or large-scale production serving, where a Gradio-based app may not be the right layer.
You want a fully managed cloud service rather than something you deploy and operate yourself.
Your use case goes beyond document question-answering into workflows the pipeline is not designed for.

Maturity signal

Created in March 2024, kotaemon is one of the more established projects in this set, with roughly 25.6k stars and a last push about a week before this snapshot, so it remains actively maintained rather than dormant. It is Apache-2.0, not archived, and backed by named maintainers at Cinnamon with a support contact. Its open-issue count (around 236) is moderate and consistent with a widely used application that draws steady user reports. Overall this reads as a mature, actively developed tool.

Alternatives

Use AnythingLLM or Open WebUI instead when you want a broader chat application and model-management experience rather than a citation-focused document-QA UI.
Use RAGFlow instead when deep document layout parsing and ingestion are your priority.
Use LlamaIndex or Haystack instead when you want a library to build a RAG system from scratch without a bundled UI.
Use Onyx (formerly Danswer) instead when you need enterprise search across many connected data sources rather than QA over uploaded files.

Notes

The project is deliberately layered for three audiences — end users, developers who import kotaemon, and contributors — and ships both a library and an app to serve them.
GraphRAG support is pluggable across three implementations, but the README notes official MS GraphRAG indexing works only with OpenAI or Ollama and recommends NanoGraphRAG for most users.
Enabling the in-browser PDF viewer requires manually downloading and extracting a specific pdf.js distribution into the assets directory.

Wiki: Cinnamon/kotaemon