RAG Evaluation Workbench
Measure retrieval quality before generation reaches production. Benchmark lexical, semantic, and hybrid retrieval with labelled relevance sets and reproducible IR evidence.
Evaluate the retrieval layer independently
Run BM25, MiniLM semantic embeddings, or hybrid retrieval against a controlled corpus. Inspect ranked evidence, compare standard IR metrics, and export a traceable evidence artifact with SHA-256 provenance.
Benchmark workspace
Configure corpus, labels, retrieval mode and scoring parametersEvaluation data
Edit the corpus and relevance labels, or run the included reference benchmark.
Retriever configuration
Semantic mode downloads a compact Hugging Face ONNX model into the browser cache on first use.
Evaluation summary
Aggregate metrics from the latest benchmark runPer-query metrics
Inspect query-level retrieval quality and ranking behavior.
Ranked retrieval evidence
Review the retrieved chunks, scores, source document and relevance label behind each metric.
Metric guide
How to interpret the benchmark metrics used by this workbench.
| Metric | Question it answers |
|---|---|
| Precision@K | Of the top-K retrieved documents, how many are relevant? |
| Recall@K | Of all labelled relevant documents, how many appeared in the top K? |
| MRR | How early does the first relevant document appear? |
| nDCG@K | How good is the ranking, with stronger reward for relevant results near the top? |
| Hit Rate | Did retrieval find at least one relevant document in the top K? |
| P50 / P95 latency | What retrieval latency did the benchmark observe? |
Architecture
The benchmark evaluates document-level retrieval quality even when retrieval operates on chunks.