Retrieval-Augmented Generation
also called RAG
Retrieving relevant documents at query time and putting them in the model's context, so answers are grounded in your data rather than in training data.
Architecturally it is a search problem with a generation step attached, and the failures are overwhelmingly search failures rather than model failures. If retrieval returns the wrong passages, no model will produce a right answer from them.
The pipeline decisions that determine quality: chunking strategy, since a chunk that splits a table from its header is useless; hybrid retrieval, because dense vectors miss exact identifiers and product codes that keyword search finds immediately; reranking, which usually improves precision more than any embedding change; and citation, so a user can verify a claim.
The reason to choose it over fine-tuning for factual grounding is operational: updating a document is a re-index, whereas updating a fine-tune is a retraining run, and RAG can enforce per-user document permissions at query time, which weights baked into a model cannot.