Retrieval & language models · 2020 · Patrick Lewis et al.

Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks

Combine parametric generation with explicit non-parametric memory so evidence can be retrieved, updated, and inspected separately from model weights.

The central move

Combine parametric generation with explicit non-parametric memory so evidence can be retrieved, updated, and inspected separately from model weights.

Why it had to exist

A model’s parameters are an opaque, lossy, and difficult-to-update store of factual knowledge. Knowledge-intensive answers need an external evidence path.

Where it leads

Search plus generation → grounded assistants → context engineering, provenance, and evidence-aware evaluation.

Study the guided reading in Bits →