R2R
R2R (Reason to Retrieve) is an open-source production-ready RAG platform. The project is available at github.com/SciPhi-AI/R2R, has approximately 7,850 GitHub stars, and was last updated in May 2026. It provides a complete backend for building RAG applications: document ingestion, hybrid search (vector + full-text), knowledge graph construction, user management, and a REST API — all as a single self-hostable service backed by PostgreSQL and pgvector.
R2R combines the components that most RAG implementations assemble separately. The ingestion pipeline accepts PDFs, DOCX, HTML, and plain text; converts them to chunks; and stores both vectors and raw text in a single PostgreSQL database. Retrieval uses hybrid search by default (dense vector search via pgvector plus sparse BM25 keyword search), with results merged by reciprocal rank fusion. Knowledge graph construction is available for entity-relationship extraction across the document corpus. The REST API is OpenAPI-documented; a Python SDK and a JavaScript SDK wrap it. User authentication, per-document access control, and conversation history are built in.
Placed in Platforms / Assess / inner.
Most RAG systems are assembled from separate services: a vector database, an embedding pipeline, a retrieval service, and a separate user management layer. R2R's value is collapsing that stack into a single Docker Compose deployment backed by PostgreSQL. For teams who have already confirmed value in RAG (Assess-stage blips like DuckDB Vector Search, Jina AI Reader, and RAG Chunking Strategies are on this radar) and are ready to operationalise it, R2R eliminates the integration cost of wiring those components.
7,850 stars with active development places it at a meaningful point of community validation — not a prototype, but not yet as established as Supabase or PostgREST (which serve a comparable "PostgreSQL as a platform" niche). The hybrid search default is a concrete advantage over embedding-only pipelines: BM25 retrieval handles named entities, code identifiers, and technical terms where vector similarity underperforms — a problem this radar's Search-based RAG blip documents separately.
Inner position reflects the Docker Compose deployment path, the single-database architecture removing operational overhead, and direct complementarity with the RAG stack already assessed here.
Trial gate: R2R deployed serving a real document corpus with hybrid search verified against at least one named-entity query type where keyword retrieval outperforms pure vector search.
