Colette

Colette is open source, self hosted software for retrieval augmented generation and LLM serving, designed for searching and interacting with technical documents that must not be sent to external APIs or online model services. It ingests Word files, PowerPoint decks, PDFs and images and lets users ask questions about them locally. Its central feature is a vision based RAG pipeline that treats every document page as an image and analyses it with vision language models. Because most technical documents are written for human eyes, this approach keeps figures, schemas, diagrams, highlights and page layout in play instead of discarding them during text extraction. Colette also offers a conventional text RAG mode with unstructured text extraction, embeddings and standard LLMs, lets users switch between several embedding and inference models, and integrates image generation through diffusers. It ships as a dockerized setup, which suits engineering, industrial and research teams that run their own infrastructure and handle confidential documentation. Colette itself is free to download and run. There is no published pricing for hosted plans. Its developer, Jolibrain, a French firm specialising in AI for critical industrial applications, offers demos and customised versions on request, with pricing for that work not published. The website is bilingual in English and French, and documentation and source code are hosted on GitHub.