eval: answer-targeting metric + retrieval-eval process gate + quote-truncation flag (Stage 1: measurement & guardrails) #243
There are no files selected for viewing
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,166 @@ | ||
| // Answer-quality / targeting eval (P3). Runs the 30-case `answerQualityEvalCases` fixture through | ||
| // the live answer path and reports the five answer-quality metrics PLUS a structural per-intent | ||
| // `targeting` metric (dose→figure, red-result→withhold action, monitoring→schedule, etc.). | ||
| // | ||
| // This is a SEPARATE, opt-in script (npm run eval:answer-quality) — it does not touch eval:quality | ||
| // or add spend to any existing run. All metrics are reported informationally; the script exits 0 | ||
| // unless --fail-on-threshold is passed with an explicit floor, so it can be calibrated safely. | ||
| import { loadEnvConfig } from "@next/env"; | ||
| import { | ||
| answerQualityEvalCases, | ||
| answerQualityMetricLabels, | ||
| answerTargetingMetricLabel, | ||
| scoreAnswerQualityEvalCase, | ||
| scoreAnswerTargeting, | ||
| type AnswerQualityEvalCase, | ||
| type AnswerQualityMetric, | ||
| } from "@/lib/rag-eval-cases"; | ||
| import type { RagAnswer } from "@/lib/types"; | ||
| import { findOwnerIdByEmail, loadAdminClient, withProviderBackoff } from "./eval-utils"; | ||
|
|
||
| loadEnvConfig(process.cwd()); | ||
|
|
||
| type Args = { | ||
| ownerEmail?: string; | ||
| ownerId?: string; | ||
| limit?: number; | ||
| intent?: string; | ||
| json: boolean; | ||
| targetingFloor?: number; | ||
| }; | ||
|
|
||
| function parseArgs(argv: string[]): Args { | ||
| const args: Args = { | ||
| ownerEmail: process.env.RAG_EVAL_OWNER_EMAIL, | ||
| ownerId: process.env.RAG_EVAL_OWNER_ID ?? process.env.LOCAL_NO_AUTH_OWNER_ID, | ||
| json: false, | ||
| }; | ||
| for (let index = 0; index < argv.length; index += 1) { | ||
| const token = argv[index]; | ||
| if (!token.startsWith("--")) continue; | ||
| if (token === "--json") { | ||
| args.json = true; | ||
| continue; | ||
| } | ||
| const value = argv[index + 1]; | ||
| if (token === "--owner-id") args.ownerId = value; | ||
| if (token === "--owner-email") args.ownerEmail = value; | ||
| if (token === "--intent") args.intent = value; | ||
| if (token === "--limit") args.limit = Number.parseInt(value, 10); | ||
| if (token === "--targeting-floor") args.targetingFloor = Number.parseFloat(value); | ||
|
Comment on lines
+49
to
+50
There was a problem hiding this comment.
When Useful? React with 👍 / 👎. |
||
| } | ||
| return args; | ||
| } | ||
|
|
||
| const METRICS: AnswerQualityMetric[] = ["relevance", "readability", "artifact_leaks", "intent_coverage", "fail_closed"]; | ||
|
|
||
| async function main() { | ||
| const args = parseArgs(process.argv.slice(2)); | ||
| const [{ requireOpenAIEnv, requireServerEnv }, { answerQuestionWithScope }, supabase] = await Promise.all([ | ||
| import("@/lib/env"), | ||
| import("@/lib/rag"), | ||
| loadAdminClient(), | ||
| ]); | ||
| requireServerEnv(); | ||
| requireOpenAIEnv(); | ||
|
|
||
| const ownerId = args.ownerId ?? (args.ownerEmail ? await findOwnerIdByEmail(supabase, args.ownerEmail) : undefined); | ||
| let cases: AnswerQualityEvalCase[] = answerQualityEvalCases; | ||
| if (args.intent) cases = cases.filter((testCase) => testCase.expectedIntent === args.intent); | ||
| if (args.limit) cases = cases.slice(0, args.limit); | ||
|
|
||
| const metricTotals: Record<AnswerQualityMetric, number> = { | ||
| relevance: 0, | ||
| readability: 0, | ||
| artifact_leaks: 0, | ||
| intent_coverage: 0, | ||
| fail_closed: 0, | ||
| }; | ||
| const targetingByIntent = new Map<string, { applicable: number; hit: number }>(); | ||
| let targetingApplicable = 0; | ||
| let targetingHit = 0; | ||
| const targetingMisses: Array<{ id: string; intent: string; reason: string; answer: string }> = []; | ||
|
|
||
| for (const testCase of cases) { | ||
| const answer = (await withProviderBackoff(`answer-quality:${testCase.id}`, () => | ||
| answerQuestionWithScope({ query: testCase.question, ownerId, logQuery: false, skipCache: true }), | ||
| )) as RagAnswer; | ||
|
|
||
| for (const score of scoreAnswerQualityEvalCase(testCase, answer)) { | ||
| metricTotals[score.metric] += score.score; | ||
| } | ||
|
|
||
| const targeting = scoreAnswerTargeting(testCase, answer); | ||
| const bucket = targetingByIntent.get(testCase.expectedIntent) ?? { applicable: 0, hit: 0 }; | ||
| if (targeting.applicable) { | ||
| bucket.applicable += 1; | ||
| targetingApplicable += 1; | ||
| if (targeting.score === 1) { | ||
| bucket.hit += 1; | ||
| targetingHit += 1; | ||
| } else { | ||
| targetingMisses.push({ | ||
| id: testCase.id, | ||
| intent: testCase.expectedIntent, | ||
| reason: targeting.reason, | ||
| answer: (answer.answer ?? "").replace(/\s+/g, " ").slice(0, 160), | ||
| }); | ||
| } | ||
| } | ||
| targetingByIntent.set(testCase.expectedIntent, bucket); | ||
| } | ||
|
|
||
| const caseCount = cases.length; | ||
| const metricRates = Object.fromEntries( | ||
| METRICS.map((metric) => [metric, caseCount ? Number((metricTotals[metric] / caseCount).toFixed(4)) : 0]), | ||
| ) as Record<AnswerQualityMetric, number>; | ||
| const targetingRate = targetingApplicable ? Number((targetingHit / targetingApplicable).toFixed(4)) : null; | ||
| const targetingByIntentRates = Object.fromEntries( | ||
| [...targetingByIntent.entries()].map(([intent, { applicable, hit }]) => [ | ||
| intent, | ||
| { applicable, hit, rate: applicable ? Number((hit / applicable).toFixed(4)) : null }, | ||
| ]), | ||
| ); | ||
|
|
||
| const summary = { | ||
| case_count: caseCount, | ||
| metric_labels: answerQualityMetricLabels, | ||
| metric_rates: metricRates, | ||
| targeting_label: answerTargetingMetricLabel, | ||
| targeting_applicable: targetingApplicable, | ||
| targeting_rate: targetingRate, | ||
| targeting_by_intent: targetingByIntentRates, | ||
| targeting_misses: targetingMisses, | ||
| }; | ||
|
|
||
| if (args.json) { | ||
| console.log(JSON.stringify(summary, null, 2)); | ||
| } else { | ||
| console.log(`Answer-quality eval: ${caseCount} case(s).`); | ||
| console.log(" metric_rates:"); | ||
| for (const metric of METRICS) console.log(` ${metric}=${metricRates[metric]}`); | ||
| console.log(` targeting_rate=${targetingRate} (applicable=${targetingApplicable})`); | ||
| console.log(" targeting_by_intent:"); | ||
| for (const [intent, value] of Object.entries(targetingByIntentRates)) { | ||
| console.log(` ${intent}=${value.rate} (${value.hit}/${value.applicable})`); | ||
| } | ||
| if (targetingMisses.length) { | ||
| console.log(" targeting_misses:"); | ||
| for (const miss of targetingMisses) { | ||
| console.log(` [${miss.intent}] ${miss.id}: ${miss.reason} :: "${miss.answer}"`); | ||
| } | ||
| } | ||
| } | ||
|
|
||
| if (args.targetingFloor !== undefined && targetingRate !== null && targetingRate < args.targetingFloor) { | ||
| console.error(`FAIL: targeting_rate ${targetingRate} < floor ${args.targetingFloor}`); | ||
| process.exit(1); | ||
| } | ||
| } | ||
|
|
||
| main() | ||
| .then(() => process.exit(0)) | ||
| .catch((error) => { | ||
| console.error(error); | ||
| process.exit(1); | ||
| }); | ||
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Oops, something went wrong.
Oops, something went wrong.
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
The checked-in
scripts/fixtures/rag-retrieval-golden.jsoncurrently has 24 cases, and this PR's verification summary reportseval:retrieval:qualityas 24/24, but the new PR checkbox tells reviewers the gate only needs to stay 23/23. For retrieval/ranking PRs this can make a one-case regression look acceptable; require zero failures or the current 24/24 count instead of the stale 23/23 number.Useful? React with 👍 / 👎.