An AI retrieval stack that does its own homework

What it does R2R is a self-hostable retrieval system that ingests multimodal documents—text, PDF, images, audio—and exposes them through a RESTful API. It handles the usual RAG pipeline but also builds automatic knowledge graphs and layers a reasoning agent on top that can perform multi-step research across your indexed data and, when needed, the internet.

The interesting bit Most open-source RAG tools stop at retrieval; R2R treats search as the opening move for an agent that can plan, fetch, and cite. The Deep Research API essentially turns your document store into a collaborator that does its own homework.

Key highlights

• Multimodal ingestion for .txt, .pdf, .json, .png, .mp3, and more
• Hybrid search combining semantic and keyword retrieval with reciprocal rank fusion
• Automatic knowledge-graph extraction for entities and relationships
• Agentic RAG with multi-step reasoning that queries both local knowledge and the internet
• Built-in user authentication and document collections

Caveats

• The README bills R2R as “the most advanced” and “production-ready,” but provides no benchmarks, latency figures, or comparative data to support those superlatives.
• Running in “light mode” versus “full mode” implies a wide gap between demo and production deployments, though the exact resource and feature trade-offs are left unclear.

Verdict Teams that want a document-to-API pipeline with built-in auth and an agentic research layer should look here; if you just need simple vector search, it is likely overkill.