haiku.rag

Agentic RAG that answers questions about your own documents with citations to page numbers and section headings. It runs on an embedded LanceDB database with open models through Ollama by default, so no server or API key is needed. Any provider Pydantic AI supports works in their place, and the same database can live on S3, GCS, Azure or LanceDB Cloud.

Built on LanceDB, Pydantic AI and Docling. Documentation: ggozad.github.io/haiku.rag.

Features

• Ingest PDFs, office documents, HTML, Markdown and images with Docling, in-process or on docling-serve. The stored DoclingDocument keeps headings, tables, pictures and page provenance.
• Search with hybrid vector and full-text retrieval, optional reranking (cross-encoders, Jina, Cohere, Zero Entropy, vLLM, OpenRouter), section-aware context expansion, and image search with a multimodal embedder (vLLM, OpenRouter, VoyageAI, Cohere). Across several named databases at once.
• Answer with the RAG capability: it searches, runs sandboxed Python over the documents for counting and aggregation, and cites page numbers and headings. Vision models receive the figures. Optional capabilities compact earlier evidence in long conversations and require every answer to declare its grounding.
• Check a citation by drawing its chunk on the page image, from the CLI, the chat TUI or Python.
• Integrate through the Python API, native Pydantic AI capabilities, an MCP server for Claude Code, Codex and Claude Desktop, and a reference web app.
• Operate with the haiku-ingester service (filesystem, HTTP, S3 and WebDAV sources, a SQLite or Postgres job queue with retries, a control plane and dashboard), tags and rollback, vacuum, and haiku-rag doctor health checks.

Installation

Python 3.12 or newer.

pip install haiku.rag        # Docling, the VoyageAI and Cohere embedders, every reranker, the TUI
pip install haiku.rag-slim   # the core, with extras chosen by you

The ingester, S3 access and model providers other than Ollama and OpenAI-compatible endpoints are extras. See Installation.

Quick start

The default configuration uses Ollama for embeddings and answers. The quickstart covers the models to pull and using OpenAI instead.

haiku-rag init                                  # create the database
haiku-rag add-src paper.pdf                     # index a file, URL or directory
haiku-rag search "attention mechanism"
haiku-rag ask "What datasets were used for evaluation?"
haiku-rag ask "How many documents mention transformers?"
haiku-rag ask "Does this figure match the spec?" --image figure.png
haiku-rag chat                                  # multi-turn chat in the terminal

Continuous ingestion from configured sources runs as a separate service, with the ingester extra (pip install 'haiku.rag[ingester]'):

haiku-ingester serve

Python API

from haiku.rag.client import HaikuRAG

async with HaikuRAG("knowledge.lancedb", create=True) as rag:
    await rag.create_document_from_source("paper.pdf")
    await rag.create_document_from_source("https://arxiv.org/pdf/1706.03762")