A DV engineer asks how long to wait after releasing reset. A general language model may produce a familiar answer from another interface, but the actual requirement belongs to a particular design and specification revision. Retrieval-augmented generation, or RAG, adds a search step before answer writing. The system locates relevant source passages and supplies them as evidence. It does not retrain the model on each question, and the presence of retrieved text does not guarantee the final answer follows that text.
Think of the workflow as a reviewable chain. First choose approved documents and identify their applicable versions. Next prepare searchable passages with their source metadata. At question time, determine the design, revision, and configuration, retrieve candidate passages, inspect whether they contain enough evidence, and only then write an answer. A final check compares each engineering claim with its cited passage. If the evidence is missing, the correct outcome is a specific statement of what could not be established.
Fictional DemoFIFO specification:
revision 1.0, section 3.1: wait two rising edges after reset release.
revision 2.0, section 3.1: wait three rising edges after reset release.
Question: For DemoFIFO revision 2.0, when may traffic start?
Required scope: DemoFIFO, revision 2.0, reset section.
Hand-authored synthetic answer:
Wait three rising clock edges after reset release before accepting traffic.
Evidence: DemoFIFO 2.0, section 3.1, chunk reset-v2.
No simulator or model result is represented here.Both passages contain the same keywords, so relevance alone cannot choose the applicable rule. Apply the revision filter before choosing evidence. A newer file timestamp is not an adequate substitute: an older specification may still govern a maintenance branch. If the question omits the revision and several revisions are available, ask for the intended revision or show the ambiguity without combining incompatible requirements into one instruction.
No. Exclude it as inapplicable evidence for the requested rule. Search the approved 2.0 collection; if that source is absent, state that the requested revision is unavailable. A relevance score cannot override a revision requirement.
Keep a short evidence record with the request: question scope, collection snapshot, retrieved chunk IDs, and final citations. This helps distinguish a retrieval failure from an answer-writing failure. If the required paragraph never entered the context, improve preparation or retrieval. If it entered but the answer changed three edges into two cycles, improve answer constraints and verification. Those are different defects and need different remedies.
RAG cannot repair an incorrect specification or prove an RTL implementation complies with it. Treat retrieved instructions as source content, never as permission to run commands or change a design. Use approved read-only collections; ingestion into an external system needs data-owner and destination approval. The foundational RAG paper by Lewis and colleagues, submitted May 2020, describes retrieval combined with generation; these engineering examples are teaching scenarios, not reported results from that paper.