RAG Evaluation
Measure whether your Retrieval-Augmented Generation system actually retrieves the necessary knowledge, uses it faithfully, and answers user queries without hallucinated interpolations.
What We Evaluate
We test the complete RAG lifecycle: chunking strategy validity, embedding model relevance, vector search precision, reranker performance, context window packing, and generator synthesis faithfulness under domain documents.
Why Generic Metrics Fail in Production
High-level synthetic benchmarks rarely surface the catastrophic edge cases that cause production RAG failure: conflicting effective dates in policy documents, subtle retrieval cutoff on multi-hop questions, and generator tendency to hallucinate plausible facts when retrieved evidence is thin.
| Metric | Target Phase | Description |
|---|---|---|
| Context Precision | RETRIEVAL | Proportion of top-k retrieved chunks that contain relevant factual evidence. |
| Context Recall | RETRIEVAL | Proportion of required ground-truth evidence successfully retrieved into the context window. |
| Citation Faithfulness | GENERATION | Mathematical groundedness: claims in the generated response directly supported by retrieved context. |
| Answer Relevance | GENERATION | Directness and completeness with which the response answers the user's specific prompt. |
| Retrieval Latency (p95) | PIPELINE | Time-to-first-chunk and full context assembly timing across variable query complexities. |
Crucial amendment or clause is ranked just below the top-k threshold, depriving the generator of evidence.
High-scoring distractor chunks with matching semantic keywords mislead the generator into incorrect conclusions.
The model generates a factual claim and cites a retrieved chunk that actually contains no supporting evidence.
| Test Partition | Tasks | Context Precision | Faithfulness | Observed Finding |
|---|---|---|---|---|
| Single-Hop Direct Lookups | 1,200 | 95.4% | 98.1% | Nominal retrieval performance |
| Multi-Document Syntheses | 800 | 82.1% | 86.4% | Chunk boundary cuts cross-references |
| Temporal Conflict Questions | 200 | 74.2% | 81.5% | Omitted appendix amendment clauses |
Audit Your RAG Architecture
Connect with our engineering team to construct a task-grounded RAG evaluation suite under mutual NDA.