@remnic/bench 9.3.577 → 9.3.579

package/dist/index.js

@@ -22580,70 +22580,50 @@ function versioningConfig(overrides) {
 // src/benchmarks/remnic/retrieval-personalization/runner.ts
 import { randomUUID as randomUUID15 } from "crypto";
 
-// src/benchmarks/remnic/retrieval-shared.ts
-function prefixAggregates(prefix, aggregates) {
-  const prefixed = {};
-  for (const [metric, aggregate] of Object.entries(aggregates)) {
-    prefixed[`${prefix}.${metric}`] = aggregate;
-  }
-  return prefixed;
-}
-function pairIdFromTaskId(taskId) {
-  return taskId.replace(/^(clean|dirty):/, "");
-}
-function buildTieredAggregates(tasks) {
-  const cleanTasks = tasks.filter((task) => task.details?.tier === "clean");
-  const dirtyTasks = tasks.filter((task) => task.details?.tier === "dirty");
-  const dirtyTasksByPairId = new Map(
-    dirtyTasks.map((task) => [pairIdFromTaskId(task.taskId), task])
-  );
-  const pairedDeltas = cleanTasks.flatMap((task) => {
-    const dirtyTask = dirtyTasksByPairId.get(pairIdFromTaskId(task.taskId));
-    if (!dirtyTask) return [];
-    const metricNames = /* @__PURE__ */ new Set([
-      ...Object.keys(task.scores),
-      ...Object.keys(dirtyTask.scores)
-    ]);
-    const deltaScores = {};
-    for (const metricName of metricNames) {
-      deltaScores[metricName] = (task.scores[metricName] ?? 0) - (dirtyTask.scores[metricName] ?? 0);
-    }
-    return [deltaScores];
-  });
-  return {
-    ...prefixAggregates("clean", aggregateTaskScores(cleanTasks.map((task) => task.scores))),
-    ...prefixAggregates("dirty", aggregateTaskScores(dirtyTasks.map((task) => task.scores))),
-    ...prefixAggregates("dirty_penalty", aggregateTaskScores(pairedDeltas))
-  };
-}
-
 // src/benchmarks/remnic/retrieval-page-ids.ts
-function extractRankedPageIds(recallText, pages) {
-  const pageIdByLowercase = new Map(
-    pages.map((page) => [page.id.toLowerCase(), page.id])
-  );
-  return collectPageIdMarkers(recallText).sort((left, right) => {
+function extractRankedPageIds(recallText, pages, options = {}) {
+  const pageIdByLowercase = new Map(pages.map((page) => [page.id.toLowerCase(), page.id]));
+  const rankedPageIds = collectPageIdMarkers(recallText).sort((left, right) => {
     if (left.index !== right.index) {
       return left.index - right.index;
     }
     return left.id.localeCompare(right.id);
   }).map((match) => resolvePageId(match.id, pageIdByLowercase));
+  const seenKnown = /* @__PURE__ */ new Map();
+  const out = [];
+  for (const pageId of rankedPageIds) {
+    if (!pageId.known) {
+      out.push(pageId.id);
+      continue;
+    }
+    const seenCount = seenKnown.get(pageId.id) ?? 0;
+    seenKnown.set(pageId.id, seenCount + 1);
+    if (seenCount === 0) {
+      out.push(pageId.id);
+    } else if (options.preserveDuplicateRankSlots) {
+      out.push(formatDuplicatePageIdSlot(pageId.id, seenCount + 1));
+    }
+  }
+  return out;
 }
 function resolvePageId(markerId, pageIdByLowercase) {
   const normalizedMarkerId = stripLeadingFormattingDelimiters(markerId).toLowerCase();
   const exact = pageIdByLowercase.get(normalizedMarkerId);
   if (exact !== void 0) {
-    return exact;
+    return { id: exact, known: true };
   }
   let candidate = normalizedMarkerId;
   while (candidate.length > 0 && isTrailingFormattingDelimiter(candidate.charCodeAt(candidate.length - 1), { includePeriod: true })) {
     candidate = candidate.slice(0, -1);
     const known = pageIdByLowercase.get(candidate);
     if (known !== void 0) {
-      return known;
+      return { id: known, known: true };
     }
   }
-  return stripTrailingFormattingDelimiters(normalizedMarkerId);
+  return { id: stripTrailingFormattingDelimiters(normalizedMarkerId), known: false };
+}
+function formatDuplicatePageIdSlot(pageId, ordinal) {
+  return `__duplicate_page_id_slot__:${ordinal}:${pageId}`;
 }
 function collectPageIdMarkers(recallText) {
   const out = [];
@@ -22734,6 +22714,43 @@ function buildSchemaTierMessages(pages) {
   }));
 }
 
+// src/benchmarks/remnic/retrieval-shared.ts
+function prefixAggregates(prefix, aggregates) {
+  const prefixed = {};
+  for (const [metric, aggregate] of Object.entries(aggregates)) {
+    prefixed[`${prefix}.${metric}`] = aggregate;
+  }
+  return prefixed;
+}
+function pairIdFromTaskId(taskId) {
+  return taskId.replace(/^(clean|dirty):/, "");
+}
+function buildTieredAggregates(tasks) {
+  const cleanTasks = tasks.filter((task) => task.details?.tier === "clean");
+  const dirtyTasks = tasks.filter((task) => task.details?.tier === "dirty");
+  const dirtyTasksByPairId = new Map(
+    dirtyTasks.map((task) => [pairIdFromTaskId(task.taskId), task])
+  );
+  const pairedDeltas = cleanTasks.flatMap((task) => {
+    const dirtyTask = dirtyTasksByPairId.get(pairIdFromTaskId(task.taskId));
+    if (!dirtyTask) return [];
+    const metricNames = /* @__PURE__ */ new Set([
+      ...Object.keys(task.scores),
+      ...Object.keys(dirtyTask.scores)
+    ]);
+    const deltaScores = {};
+    for (const metricName of metricNames) {
+      deltaScores[metricName] = (task.scores[metricName] ?? 0) - (dirtyTask.scores[metricName] ?? 0);
+    }
+    return [deltaScores];
+  });
+  return {
+    ...prefixAggregates("clean", aggregateTaskScores(cleanTasks.map((task) => task.scores))),
+    ...prefixAggregates("dirty", aggregateTaskScores(dirtyTasks.map((task) => task.scores))),
+    ...prefixAggregates("dirty_penalty", aggregateTaskScores(pairedDeltas))
+  };
+}
+
 // src/fixtures/schema-tiers/index.ts
 var DEFAULT_SCHEMA_TIER_SEED = 448;
 var CLEAN_PAGES = [
@@ -23215,7 +23232,9 @@ async function runRetrievalPersonalizationBenchmark(options) {
     await options.system.drain?.();
     const recallText = await options.system.recall(sessionId, sample.query, 12e3);
     const latencyMs = Math.round(performance.now() - startedAt);
-    const rankedPageIds = extractRankedPageIds(recallText, sample.pages);
+    const rankedPageIds = extractRankedPageIds(recallText, sample.pages, {
+      preserveDuplicateRankSlots: true
+    });
     const topRetrievedPageIds = rankedPageIds.slice(0, 5);
     const expectedJson = JSON.stringify(sample.expectedPageIds);
     const actualJson = JSON.stringify(topRetrievedPageIds);
@@ -23356,14 +23375,11 @@ async function runRetrievalTemporalBenchmark(options) {
     await options.system.reset(sessionId);
     await options.system.store(sessionId, buildSchemaTierMessages(sample.pages));
     await options.system.drain?.();
-    const recallText = await options.system.recall(
-      sessionId,
-      sample.query,
-      12e3,
-      { asOf: sample.window.end }
-    );
+    const recallText = await options.system.recall(sessionId, sample.query, 12e3, { asOf: sample.window.end });
     const latencyMs = Math.round(performance.now() - startedAt);
-    const retrievedPageIds = extractRankedPageIds(recallText, sample.pages);
+    const retrievedPageIds = extractRankedPageIds(recallText, sample.pages, {
+      preserveDuplicateRankSlots: true
+    });
     const topRetrievedPageIds = retrievedPageIds.slice(0, 5);
     const matchedPageIds = matchingPageIds(retrievedPageIds, sample);
     const expectedJson = JSON.stringify({

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@remnic/bench",
-  "version": "9.3.577",
+  "version": "9.3.579",
   "description": "Retrieval latency ladder benchmarks + CI regression gates for @remnic/core",
   "type": "module",
   "main": "./dist/index.js",
@@ -34,7 +34,7 @@
   "dependencies": {
     "hyparquet": "^1.25.7",
     "yaml": "^2.4.2",
-    "@remnic/core": "^9.3.577"
+    "@remnic/core": "^9.3.579"
   },
   "devDependencies": {
     "tsup": "^8.5.1",