test: configure raw diversity eval budgets

Helweg · Helweg · commit a57e4af8c656 · 2026-04-11T22:01:38.000+02:00
diff --git a/benchmarks/baselines/eval-baseline-summary.json b/benchmarks/baselines/eval-baseline-summary.json
@@ -19,6 +19,8 @@
     "hitAt10": 1,
     "mrrAt10": 0.875,
     "ndcgAt10": 0.9127302324517832,
+    "distinctTop3Ratio": 1,
+    "rawDistinctTop3Ratio": 1,
     "latencyMs": {
       "p50": 26.173166000000037,
       "p95": 52.931082999999944,
diff --git a/benchmarks/budgets/default.json b/benchmarks/budgets/default.json
@@ -5,9 +5,11 @@
   "thresholds": {
     "hitAt5MaxDrop": 0.03,
     "mrrAt10MaxDrop": 0.03,
+    "rawDistinctTop3RatioMaxDrop": 0.1,
     "p95LatencyMaxMultiplier": 1.35,
     "p95LatencyMaxAbsoluteMs": 4000,
     "minHitAt5": 0.4,
-    "minMrrAt10": 0.25
+    "minMrrAt10": 0.25,
+    "minRawDistinctTop3Ratio": 0.5
   }
 }
diff --git a/benchmarks/budgets/github-models.json b/benchmarks/budgets/github-models.json
@@ -4,6 +4,7 @@
   "thresholds": {
     "minHitAt5": 0.5,
     "minMrrAt10": 0.45,
+    "minRawDistinctTop3Ratio": 0.5,
     "p95LatencyMaxAbsoluteMs": 500
   }
 }

Original file line number	Diff line number	Diff line change
`@@ -5,9 +5,11 @@`
`5`	`5`	`"thresholds": {`
`6`	`6`	`"hitAt5MaxDrop": 0.03,`
`7`	`7`	`"mrrAt10MaxDrop": 0.03,`
	`8`	`+ "rawDistinctTop3RatioMaxDrop": 0.1,`
`8`	`9`	`"p95LatencyMaxMultiplier": 1.35,`
`9`	`10`	`"p95LatencyMaxAbsoluteMs": 4000,`
`10`	`11`	`"minHitAt5": 0.4,`
`11`		`- "minMrrAt10": 0.25`
	`12`	`+ "minMrrAt10": 0.25,`
	`13`	`+ "minRawDistinctTop3Ratio": 0.5`
`12`	`14`	`}`
`13`	`15`	`}`