Evaluate with evidence

Measure retrieval recall@k against evidence spans, precision or nDCG, answer groundedness, citation correctness, duplicate rate, average retrieved tokens, and latency. Build queries that require facts near boundaries and compare chunking variants while holding the embedder and generator fixed. An evaluation is useful only if its result changes a choice: ship, hold, route, tune, collect data, or retire. Define that choice and its hard gates before selecting metrics.

For an employee handbook assistant, create cases from real task distributions plus intentionally difficult boundaries. Keep a locked set for final comparison and a development set for iteration. Include slices by input type, language, risk, and consequence. Random sampling estimates common behavior; targeted challenge sets expose rare severe failures. You need both.

Measure three layers separately:

1. Component evidence asks whether retrieval, memory selection, ranking, encoding, or coordination worked.
2. End-to-end quality asks whether the user’s task was completed correctly and safely.
3. Operational outcome asks about latency, cost, availability, escalation, and downstream value.

Chunking creates retrievable units; embedding maps units to vectors; retrieval selects candidates; reranking reorders them; generation writes the answer. Increasing top-k cannot reliably repair a chunk that never contained a coherent fact. A component improvement is valuable only when it preserves gates and helps the end-to-end decision.

Suppose 84 of 100 cases pass. The observed pass rate is 84%, but another sample would differ. Report a confidence interval or bootstrap distribution, not false precision. For rare severe errors, count and inspect every event; an average quality score must not wash out a security or privacy breach.

Use deterministic scoring for exact properties such as schema validity or known calculations. Use human rubrics for nuanced correctness and harm. Model judges can scale review, but calibrate them against blinded human labels, measure agreement by slice, and periodically recheck after model or prompt updates.

Hold input cases, prompts, tools, timeouts, and scoring constant between candidates. Pair results case by case because the pattern of wins matters more than two independent averages. Record failures and adjudication notes. Reject contaminated cases that appeared in training only when the protocol says how contamination is detected.

A leave-policy answer spans the final sentence of “Eligibility” and first list under “Exceptions.” With 400-token non-overlapping windows, recall@5 is 71%. Adding 60-token overlap raises it to 86% but duplicates 28% of results. Heading-aware 250-token clauses with parent expansion reach 91% recall and only 7% duplicates. That trace demonstrates practical significance: a setting can raise one metric while violating a gate or harming a critical slice. The report should make that conflict visible.

Write a release rule such as: “Ship to 10% only if severe errors are zero on the challenge set, primary task success improves at least three points, every protected slice stays within two points, and p95 latency remains below the agreed budget.” After release, monitor the same constructs with production-appropriate proxies and delayed labels. Offline evaluation and online monitoring form a loop, not competing rituals.