Evaluate retrieval before evaluating answers
A final answer can fail because the relevant source was never retrieved, because the model ignored it, or because the evaluation label is weak. RetrievalGate isolates the first question: did the retrieval layer surface the intended source near the top of the result list? This does not replace answer-quality, latency or safety evaluation in your application.
Bring a corpus and, when possible, real query evidence
Upload the documents your system searches. For stronger evidence, add evaluation.csv with query and relevant_id, plus an optional segment. A segment can represent identifier lookups, policy questions or languages. Without a CSV, the Free Snapshot creates a limited synthetic benchmark and labels its evidence accordingly.
Read more than one global score
The audit reports Recall at 1, 3 and 5, MRR and nDCG at 5. A global improvement is not enough when an important query type gets worse. Full Audit shows segment and query-level regressions so a team can see whether a candidate improves the average by sacrificing a commercially important tail.
Test a bounded set of changes
The Free Snapshot compares the current bounded baseline with a limited retrieval alternative. Full Audit tests supported chunking strategies, available embedding candidates, BM25, dense and hybrid retrieval, and an optional reranker when configured. The corpus and query evidence remain fixed while those variables change.
Turn the evaluation into a release decision
The report recommends a candidate only when its evidence supports a change. It can also return NO-GO or insufficient evidence. The useful output is not a leaderboard: it is an implementation card, known regressions and the next validation step for the team that owns production.