Haiku Rag
Stand up an opinionated agentic RAG stack with LanceDB, Pydantic AI, and Docling exposed to your IDE agent via MCP.
Overview
Haiku RAG is a MCP server for the Build phase that serves an agentic RAG stack built on LanceDB, Pydantic AI, and Docling.
What is this MCP server?
- Agentic RAG design combining LanceDB vector storage, Pydantic AI orchestration, and Docling document parsing
- PyPI package haiku-rag (0.56.0) launched via uvx with serve and --mcp flags
- Stdio MCP transport for Claude Code, Cursor, and other MCP hosts
- Opinionated defaults so you skip wiring embeddings, chunkers, and parsers from scratch
- Fits products that need grounded answers over PDFs, docs, and internal knowledge
- PyPI package haiku-rag version 0.56.0 in server metadata
- Runtime hint uvx with serve and --mcp arguments
- Stack components named: LanceDB, Pydantic AI, Docling
Community signal: 537 GitHub stars.
What problem does it solve?
Rolling your own RAG means gluing vector DBs, parsers, and agent frameworks manually before your MVP can answer questions over real documents.
Who is it for?
Technical indie builders shipping doc-grounded agents who want LanceDB plus Docling without maintaining a custom MCP wrapper.
Skip if: No-code founders who need a hosted console only, or teams that require a non-Python-only retrieval stack.
What do I get? / Deliverables
After adding Haiku RAG MCP, your agent can call a maintained RAG serve endpoint over stdio while you focus on product logic.
- Running stdio MCP RAG server via haiku-rag serve --mcp
- Docling-driven ingestion path into LanceDB for agent queries
- Agent-callable retrieval tools integrated into your MCP config
Recommended MCP Servers
Journey fit
Retrieval infrastructure is built when you productize agent memory and document Q&A, not when you are only validating a landing page. Haiku RAG runs as serve --mcp over uvx from PyPI, which is classic build-phase agent-tooling for solo builders.
How it compares
Full agentic RAG MCP server, not a single-file prompt skill or compliance verdict tool.
Common Questions / FAQ
Who is haiku-rag for?
Developers building agent products who want LanceDB, Pydantic AI, and Docling wired as an MCP server from PyPI.
When should I use haiku-rag?
Use it during build when you need ingestion, chunking, and retrieval tools callable from Claude Code or Cursor over MCP.
How do I add haiku-rag to my agent?
Configure uvx to run haiku-rag with positional serve and the --mcp named flag per registry runtimeArguments, version 0.56.0, stdio transport.