RAGBrain GitHub PyPI Demo

RAG with
a brain.

An open-source bi-temporal RAG platform. It remembers what it knew, and when it knew it, so an answer about the past is defensible rather than merely plausible.

Try the replay Source pip install ragbrain

The invariant, live

What did the system believe, and when?

Acme Corp moved its headquarters. The 2019 filing said Boston; the 2022 filing said Denver. Drag the slider to replay what the ledger knew at any past moment. Later corrections never leak backward.

The system knew, as of Jun 2018
2018↑ 2022 correction2024
Acme Corp is headquartered in Boston
Acme Corp is headquartered in Denver

Two time axes

Event time [valid_at, invalid_at) When the fact was true in the world. Answers “where was it in 2020?”
System time [created_at, expired_at) When the system learned it. Answers “what did we believe last March?” and powers replay.

What that buys you

Corrections supersede

A correcting fact expires the old one and stamps a back-link at write time. Nothing is silently overwritten, so the timeline stays intact.

Cited, checked answers

Generation is constrained to served facts and verified claim by claim afterwards. Unsupported claims are flagged, never quietly shipped.

Self-hostable

FalkorDB or Neo4j, local or hosted embedders, your infrastructure. The core imports only the standard library.