00 · IN THREE MINUTES
The answer in three steps
- 1RAG searches an external collection and inserts selected passages into the model’s context.
- 2The answer can be grounded in those passages and linked to inspectable sources.
- 3Bad retrieval, conflicting documents or unsupported synthesis can still produce a wrong answer.
01 · THE MODEL GAINS AN OPEN BOOK
The model gains an open book
Instead of relying only on parameters learned during training, a RAG system converts a query into a search representation and retrieves passages from an index. Those passages travel with the question into the generator.
02 · RETRIEVAL IS A SEPARATE MODEL
Retrieval is a separate model
Chunking, embeddings, filters and ranking decide what evidence reaches the language model. If the decisive passage is absent or ranked too low, generation cannot reliably recover it from the index.
The model gains an open book
Instead of relying only on parameters learned during training, a RAG system converts a query into a search representation and retrieves passages from an index. Those passages travel with the question into the generator.
queryRetrieval is a separate model
Chunking, embeddings, filters and ranking decide what evidence reaches the language model. If the decisive passage is absent or ranked too low, generation cannot reliably recover it from the index.
retrieveContext can still be misread
The generator may ignore a passage, merge incompatible claims or cite a source beside a sentence it does not support. Documents may also be stale, poisoned or authoritative-looking but wrong.
rerankEvaluation needs two layers
Retrieval quality asks whether relevant evidence was found. Answer quality asks whether the response follows that evidence; a useful test reports both instead of hiding them behind one score.
answer with evidence03 · CONTEXT CAN STILL BE MISREAD
Context can still be misread
The generator may ignore a passage, merge incompatible claims or cite a source beside a sentence it does not support. Documents may also be stale, poisoned or authoritative-looking but wrong.
04 · EVALUATION NEEDS TWO LAYERS
Evaluation needs two layers
Retrieval quality asks whether relevant evidence was found. Answer quality asks whether the response follows that evidence; a useful test reports both instead of hiding them behind one score.
05 · RAG IS AN ARCHITECTURE, NOT A GUARANTEE
RAG is an architecture, not a guarantee
Good systems expose citations, preserve document dates, restrict sources when needed and allow refusal when evidence conflicts. Verification remains necessary wherever a wrong synthesis carries real cost.
06 · SOURCES AND EVIDENCE
Sources and evidence
Claims are linked to foundational papers, standards or the primary study behind the update.
- 01Retrieval-Augmented Generation for Knowledge-Intensive NLP TasksPRIMARY STUDY ↗
Supports a defined mechanism, measurement or evidence boundary in this article.
- 02Retrieval-Augmented Generation for Large Language Models: A SurveySCHOLARLY REVIEW ↗
Supports a defined mechanism, measurement or evidence boundary in this article.