RLMs

Retrieval-Augmented Language Models (RLMs) fuse massive external knowledge bases with neural generators to eliminate hallucinations and provide verifiable citations.

RLMs solve the static knowledge bottleneck by decoupling memory from parameters. Instead of relying solely on weights frozen during training, these systems use a retriever (like FAISS or Scann) to pull relevant document snippets from datasets like Wikipedia or internal corporate wikis. This architecture, popularized by frameworks like RAG and Google's REALM, ensures models provide factual, up-to-date responses while significantly reducing the parameter count needed for high performance. By grounding outputs in specific retrieved text, operators gain a clear audit trail for every claim the model makes.

https://github.com/google-research/google-research/tree/master/rlm