Run local document RAG with citations over MCP using Haiku.RAG
Index local or self-hosted documents, search them with hybrid and multimodal retrieval, and answer agent questions through an MCP server with citations.
npx skills add agentskillexchange/skills --skill run-local-document-rag-with-citations-over-mcp-using-haiku-rag
pip install haiku.rag or uv pip install haiku.rag, index documents with commands such as haiku-rag add-src paper.pdf, then expose the knowledge base to an MCP client with haiku-rag mcp --stdio. Use haiku-rag --read-only mcp --stdio when the agent should only search and ask questions.Use Haiku.RAG when an agent needs a local-first retrieval layer over private documents rather than a hosted knowledge product. The operator installs the Python package, indexes files or URLs into an embedded LanceDB store, starts the MCP server, and lets an MCP-compatible assistant search documents, ask cited questions, analyze document collections, and optionally ingest new sources through bounded tools. This is strongest for evidence-backed PDF, document, and image-grounded research workflows where citations, local storage, read-only mode, and repeatable retrieval behavior matter. The scope boundary is document retrieval and cited QA over a configured knowledge base; it is not a generic RAG framework listing, vector database listing, or product card.
What this skill actually does
Inputs and prerequisites: Python 3.12+, haiku.rag or haiku.rag-slim, an embedding provider such as Ollama/OpenAI/VoyageAI/Cohere/LM Studio/vLLM, and an MCP-compatible client.
Setup notes: Install with pip install haiku.rag or uv pip install haiku.rag , index documents with commands such as haiku-rag add-src paper.pdf , then expose the knowledge base to an MCP client with haiku-rag mcp –stdio . Use haiku-rag –read-only mcp –stdio when the agent should only search and ask questions.
Source and verification boundary: use https://ggozad.github.io/haiku.rag/ as the canonical reference before running the workflow; keep commands, API calls, CLI usage, and generated outputs reviewable against that upstream source.
Framework fit: publish this as a MCP workflow only when the operator can invoke the documented toolchain directly, rather than treating the upstream project as a generic product listing.