Back to Ecosystem Pulse
ToolProduction ReadyMCPAG-UI
haiku.rag
by ggozad
Self-hosted agentic RAG with LanceDB, Docling parsing, reranking and MCP server
Python
Updated Sep 18, 2026
Share:
What It Does
Implements an agentic RAG stack for local and self-hosted document search, combining hybrid retrieval, reranking and multimodal retrieval over embedded LanceDB. Runs a lightweight MCP server and uses Docling for document parsing and structured ingestion, with cross-encoder reranking and reciprocal-rank-fusion. Designed for offline LLMs and self-hosted pipelines to reduce hallucination and improve answer grounding, leveraging the Agentic RAG Pattern.
Key Benefits
As agents exchange retrieved knowledge and delegate subqueries, tracking where answers came from and how reliable retrieval was becomes critical for trust. Haiku.rag makes provenance and reranking explicit in self-hosted RAG stacks, enabling better signal for LLM-as-Judge Pattern and agent track record assessments. This matters because grounded retrieval + reranking is one of the most effective ways to reduce hallucinations and create reproducible evidence for downstream trust signals.
Ideal For
Teams building offline or self-hosted RAG pipelines that need tight document provenance, reranking, and integration with multi-agent MCP workflows.
How It's Used
- Build self-hosted RAG search for an offline LLM with provenance and reranking
- Integrate document parsing (Docling) and LanceDB embeddings into agent workflows
- Run multi-agent retrieval via an MCP server to enable agent-to-agent query delegation
- Reduce hallucination by applying cross-encoder reranking and reciprocal-rank-fusion
Works With
lancedbdoclingoffline-llmpydantic-ai
Topics
ag-uiagentic-ragcode-executioncode-modecross-encoderdoclingdocument-aidocument-ingestiondocument-parsinghallucination-prevention+9 more
Similar Tools
langchainllamaindex
Keywords
multi-agent trustagent-to-agent evaluationraglancedb