haiku.rag

haiku.rag lets you search your own documents and ask questions with answers you can check against the original pages. It's for people building a private document library or giving an AI assistant access to their files.

• Cited answers: Answers point to page numbers and section headings, and visual grounding highlights the supporting passage on the original page image.
• Document ingestion: Import PDFs, office documents, HTML, Markdown and images, or keep a collection updated from directories, HTTP, S3 and WebDAV sources.
• Text and image search: Combine semantic and full-text search, optionally rerank results, and search images with a multimodal embedding model across multiple databases.
• Document calculations: The answering agent can run sandboxed Python over documents for questions that need counting or aggregation; an optional grounding policy requires answers to state their evidence or acknowledge its absence.
• Chat and agent access: Use browser chat, connect Claude Desktop, Claude Code or Codex through MCP, or build document-aware agents with the Python API and Pydantic AI capabilities.

The default setup keeps documents on your machine in embedded LanceDB and uses Ollama for embeddings and answers, with no API key or separate database server required. You can choose other Pydantic AI model providers or store the database on S3, GCS, Azure or LanceDB Cloud; those choices use external services rather than the local defaults. The project is open source under the MIT license and requires Python 3.12 or newer.