Retrieval-augmented generation
Combining two memories rather than enlarging one. A pre-trained sequence-to-sequence model supplies the parametric memory; a dense vector index, reached through a neural retriever, supplies the non-parametric one.
Official source: Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks →
The founding work fine-tunes both together for generation, which is the distinction worth keeping: retrieval feeds generation here, instead of extracting an answer span the way earlier retrieval systems did.
Sources 1 source on record · Automated review to redo
Sources
What this page is based on — every source is verified, and links out whenever the document is still reachable.