Capability · Framework — rag

R2R

R2R from SciPhi is a batteries-included RAG system you deploy as a service. It handles document ingestion and chunking, hybrid search (vector + keyword), knowledge-graph construction and querying, agentic RAG, observability, and user-level document permissions. It ships a REST API, Python and TypeScript SDKs, and a dashboard — positioned as the RAG equivalent of Supabase for apps.

Framework facts

Category: rag
Language: Python / TypeScript
License: MIT
Repository: https://github.com/SciPhi-AI/R2R

Install

pip install r2r
# run server:
# docker compose -f r2r-compose.yaml up -d

Quickstart

from r2r import R2RClient

client = R2RClient('http://localhost:7272')
client.documents.create(file_path='paper.pdf')
result = client.retrieval.rag(
    query='What are the main findings?'
)
print(result.results.generated_answer)

Alternatives

Haystack — pipeline framework, not a server
LlamaIndex — framework with managed LlamaCloud
Verba — Weaviate's opinionated RAG app
txtai — embeddings-first toolkit

Frequently asked questions

Is R2R a framework or a service?

R2R is a service — a deployable server with a REST API. You don't compose pipelines in Python like with Haystack or LlamaIndex; you call retrieve/rag/ingest endpoints. That makes it fast to stand up an internal RAG platform.

What does 'Reason to Retrieve' mean?

R2R supports agentic retrieval — the LLM can decide what to search for, iterate, and use knowledge graphs — rather than a fixed embed-then-retrieve pipeline. This is the 'reason to retrieve' behaviour it's named after.

Sources

R2R — docs — accessed 2026-04-20
R2R — GitHub — accessed 2026-04-20