RAG Systems

Retrieval-augmented generation, or RAG, combines a model with a search system. The answer depends on the user query, query rewriting, embedding or search settings, retrieved documents, ranking, prompt assembly, and generation. A transparent RAG system exposes enough of that path to judge the answer.

Inspect a RAG Pipeline

Step through the pipeline behind a policy question-answering assistant. Then toggle retrieval quality and the grounding check to see when the system should answer — and when it should abstain.

Query

"What is the maximum reimbursement for remote work equipment?"

The system answers, and the citation is faithful to a retrieved passage.
The question is not just what the model said — it's what sources it used, whether they actually support the claim, and what should happen when they don't.

Useful RAG transparency includes the rewritten query, retrieved chunks, source metadata, rank scores, citation mapping, freshness, access permissions, and grounding checks. Users should be able to distinguish supported claims from plausible-sounding additions.

RAG systems also need abstention. If retrieval fails or sources conflict, the system should say so instead of producing a confident answer from weak evidence.

Citations Can Be Decorative

A citation is only useful if the cited passage actually supports the claim. Check whether citations are faithful, current, and permissioned for the user.

Checkpoint

A RAG assistant gives a correct-looking answer with three citations, but none of the cited passages contain the key claim. What failed?