Retrieval relevance
Measure hit quality, ranking, required-document recall, and irrelevant-context pressure.
RAG Evaluation
Know why retrieval succeeds—and how it fails.
Evaluate the complete retrieval chain—not just the final answer. Prooflane separates retriever failures from generation failures and turns the evidence into release policy.
Seven dimensionsGolden baselinesFault injection
support-rag · release suite
$ prooflane rag run --suite release retrieval quality 94 tool selection 100 grounded behavior 91 contract validity 100 baseline drift −12 answer chain 88 fault resilience 79 BLOCK · poisoning case newly failed evidence retained locally
Seven dimensions
Trace the answer back through the chain.
Every score maps to evidence at the retrieval, tool, contract, and response layers.
Tool selection
Verify the agent chooses the intended retriever with valid arguments and avoids redundant calls.
Grounded response
Check whether the answer uses retrieved evidence, cites support, and abstains when evidence is absent.
Output validity
Validate retrieval and answer schemas so structural breakage cannot masquerade as quality drift.
Golden comparison
Compare documents, ordering, metadata, and result shape against a versioned baseline.
Judging and resilience
Evaluate end-to-end faithfulness plus no-hit, timeout, partial, duplicate, stale, and poisoned cases.
RAG Evaluation walkthrough
Reproduce a retrieval regression in one run.
Watch the query-to-retrieval-to-answer chain execute against reusable cases and explicit judging modes.
REAL PRODUCT · RAG Evaluation
Measure the complete chain
Run the baseline suite, introduce one synthetic poisoned document, inspect the failed dimension, and enforce the release gate.
See the CLI workflow
prooflane rag run --suite releaseprooflane rag show <run-id> --case poisoned-policyprooflane gate --rag releaseREAL FIXTURE RUN
See the weak dimension, the run history, and the unsafe retrieval path.
These views come from Prooflane fixture runs. Third-party model, target, and tool identifiers are permanently redacted in the published assets.
Exported report anatomy
Retrieval94
Grounding91
Drift−12
Poisoningfailed
Sample evidence is generated from synthetic documents and authorized fixtures, not customer content. The report distinguishes relevance, grounded behavior, contracts, drift, resilience, and poisoning outcomes.
Workflow
Version the chain from corpus to answer.
Define cases
Version queries, expected evidence, assertions, and allowed abstentions.
Retrieve
Capture documents, ranking, metadata, timings, and tool behavior locally.
Stress
Inject no-hit, timeout, stale, duplicate, partial, and poisoning cases.
Gate
Compare with baseline and block meaningful quality or safety regressions.
prooflane rag run --suite release --json rag-run.json prooflane gate --rag release --baseline .prooflane/rag.json
Privacy boundary
Documents remain close to the retriever.
Local by default
- Corpus credentials and connection strings
- Retrieved documents and raw model outputs
- Queries, customer content, and full traces
Sanitized by policy
- Dimension scores and normalized failures
- Content hashes and version metadata
- Consented evidence for independent judging
Part of the spine
Connect retrieval evidence to system decisions.
Assure the chain
Ship retrieval changes with evidence.
See what changed, which layer failed, and whether the release should proceed.
30-day full-access trial. No credit card required.