Google retrieval-augmented generation guide
Official guidance on grounding generated responses in retrieved information and evaluating the retrieval pipeline.
← Back to Compute College
Compute College
Evaluate retrieval and grounded answers
Measure whether retrieval finds useful evidence and whether the final answer stays grounded in that evidence.
Plain-English definition
Retrieval evaluation asks whether the system found the right evidence; grounded-answer evaluation asks whether the model used that evidence accurately. They are related but separate: a correct passage can be retrieved and ignored, while a fluent answer can sound grounded without being supported by the retrieved text.
Memory trick: Find the evidence, then prove the answer used it.
Why it matters
A single end-to-end answer score cannot show whether failure came from chunking, ranking, missing documents, prompt instructions, or generation. Separating retrieval and answer checks makes fixes more targeted and prevents adding more context when the real problem is poor ranking or unsupported synthesis.
- Retrieval relevance checks whether useful passages appear in the candidate set.
- Evidence coverage checks whether claims in the answer are supported.
- Abstention and “not found” behavior matter when the corpus lacks the answer.
Simple example
For a policy question, the evaluator records whether the authoritative policy section appears in the top five retrieved passages, whether each answer claim has supporting text, and whether the assistant declines when the policy is absent or outdated.
- Top-k retrieval and final-answer support are scored separately.
- Authority and freshness are part of evidence quality.
- A correct abstention can be better than an unsupported answer.
Example figures are illustrative calculations, not current quoted market prices.
Current example
Primary reference
This source explains the method or risk boundary. It does not replace testing the actual model, data, provider, and workload conditions in production.
Source discipline: examples on this page are illustrative; measure the real workload before making a release or capacity decision.
Common mistake
More documents do not automatically make an answer more grounded. Extra context can bury the right passage, increase cost, and introduce contradictory or untrusted evidence.
Practical takeaway
What you can do with this
Create a retrieval test set with a known authoritative passage, distractors, stale documents, and unanswered questions. Score retrieval and final answer support separately.
- Record document authority and freshness in the test data.
- Check every material claim against evidence.
- Add unanswered cases to test honest abstention.
Decision check: when an answer fails, can you identify whether retrieval, evidence selection, generation, or abstention caused the failure?
Compute College learning path
AI Engineering
Step 24 of 48: Evaluate retrieval and grounded answers