Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks
ReadPairs a dense retriever with a generator so answers can draw on documents outside the weights. The reference design behind retrieval-augmented systems.
My notes
The original framing. Worth re-reading for how little the retriever was tuned — they only fine-tune the query encoder and keep the document index frozen.
RAG-Sequence vs RAG-Token
- RAG-Sequence: one document conditions the whole answer.
- RAG-Token: each generated token can lean on a different document.
- The index is swappable — update the documents without retraining. That property is still the main reason teams pick RAG over fine-tuning.
What Production Taught Me That the Paper Does Not
- Chunking and hybrid search (BM25 + dense) matter more than the generator.
- Evaluate retrieval separately from generation, or you cannot tell which half failed.