01 · Prepare the evidence
Upload text PDFs, Markdown, TXT, CSV, JSON or JSONL. Each file can be up to 4 MB; an audit accepts up to 50 selected files and 25 MB in total. Scanned PDFs need OCR before upload. CSV rows and JSON records become separate source documents. Choose exactly which versions belong to a run; later uploads do not change that selection.
02 · Define the questions
Supply evaluation.csv with query and relevant_id columns and an optional segment. Each row defines one question and one relevant source. A full source ID is safest; unique filenames also work. For a CSV, append #row=17 to target that row. Missing or ambiguous references stop the audit instead of producing misleading scores. Without a query set, the engine generates synthetic questions from corpus passages.
03 · Reconstruct the baseline
The optional current-RAG form selects a supported chunking strategy, embedding model and retrieval mode. Without it, the baseline is approximately 800-token chunks, the first configured embedding model and dense retrieval. This is an offline reconstruction, not a measurement of your live application. Unsupported baselines are rejected instead of silently replaced.
04 · Compare the candidates
Free compares BM25, dense and hybrid retrieval on one chunking strategy with up to 12 questions. Full audits test four chunking strategies and the available embedding candidates, with up to 75 questions. If a reranking provider is configured, a reranked version of the strongest dense or hybrid candidate is evaluated as well.
05 · Read the decision correctly
The report shows hit-based Recall@1/3/5, reciprocal rank and nDCG@5 for the single-relevant-source evaluation contract. Candidate selection penalizes segment regressions. NO-GO means no material Recall@5 improvement justified changing the baseline. Conditional recommendations need review. Too few documents or chunks produce insufficient evidence. The benchmark does not measure end-to-end answer accuracy or production latency.
06 · Validate the change
Download the full JSON report, review the implementation card, and make the change in your own environment. A paid audit includes one rerun within 14 days after its first completed analysis. Use new, held-out evaluation questions for that rerun: selecting and reporting on the same queries can overestimate improvement. Nothing is deployed to your production system automatically.