Haiku RAG is a feature-rich agentic RAG server combining LanceDB, Pydantic AI, and Docling, offering hybrid search, multimodal search, citation-based QA, and analysis capabilities. It is suitable for local-first RAG applications, but users should be aware of the risks of sandboxed code execution and its non-official maintenance status.
Haiku RAG is an opinionated agentic RAG server built on LanceDB, Pydantic AI, and Docling. It offers hybrid search (vector + full-text with Reciprocal Rank Fusion), multimodal and cross-modal search (text-as-query → figure hits, image-as-query), question answering with citations (page numbers, section headings), vision QA (vision-capable models receive figure bytes alongside chunk text), reranking (local cross-encoders, Cohere, Zero Entropy, or vLLM), analysis capability (complex analytical tasks via sandboxed Python execution), conversational RAG (chat TUI and web application), and document structure storage. It supports multiple embedding providers (Ollama, OpenAI, VoyageAI, Cohere, LM Studio, vLLM) and is local-first with embedded LanceDB, also supporting S3, GCS, Azure, and LanceDB Cloud. It provides a CLI, Python API, and MCP server to expose tools for AI assistants like Claude Desktop.
pip install haiku.rag (or use the slim package haiku.rag-slim).{
"mcpServers": {
"haiku-rag": {
"command": "haiku-rag",
"args": ["mcp", "--stdio"]
}
}
}