Test set
Representative questions, user groups, documents, edge cases, adversarial cases, and expected evidence.
Evidence:Question taxonomy, source mapping, inclusion rationaleEvaluation framework
Evaluate the knowledge system—not just the final sentence.
A useful knowledge assistant must retrieve the right evidence, respect permissions, support its claims, abstain when needed, perform reliably, and fit a human workflow.
Evaluation matrix
Six surfaces to test before release.
Retrieval
Whether relevant, permitted, sufficiently complete evidence appears in the retrieved context.
Evidence:Recall/precision variants, rank review, failure examplesGrounding
Whether important answer claims are supported by the supplied sources and citations point to the right material.
Evidence:Claim-source checks, citation correctness, unsupported-claim rateAbstention
Whether the system recognizes missing evidence, ambiguity, conflicts, and questions outside the approved scope.
Evidence:Refusal/clarification cases, false-confidence reviewOperations
Latency, cost, availability, logging, source freshness, change control, monitoring, and incident response.
Evidence:Service measures, runbook, alerts, ownershipHuman workflow
Whether users understand citations, limitations, review duties, escalation, and the decisions they still own.
Evidence:Task testing, adoption feedback, override and escalation recordsRelease decision
Do not average away a critical failure.
A high overall score can hide permission leakage, unsupported consequential claims, or failure for a key user group. Release gates should include non-negotiable conditions as well as aggregate measures.
Aggregate measuresUseful for comparison and trends
Critical gatesRequired for unacceptable failure modes
Human decisionNamed owner accepts residual risk
Minimum evaluation record
What another reviewer should be able to inspect.
PurposeUsers, questions, decisions, and exclusions
System versionModels, prompts, retrieval, sources, settings
Test constructionSampling, labels, reviewers, disagreements
ResultsMeasures, gates, slices, and failures
DecisionRelease, restrictions, mitigations, owner
MonitoringSignals, feedback, changes, incidents
This framework is an educational working method, not a certification, legal opinion, security assessment, or guarantee that a system is safe or compliant.