causantic 0.9.4 → 0.10.0

package/dist/eval/experiments/index-differentiation/alignment-analysis.d.ts

@@ -0,0 +1,53 @@
+/**
+ * Phase 2b: Entry-to-chunk alignment analysis.
+ *
+ * Measures how well each index entry's embedding preserves the semantic
+ * direction of its source chunk. Complements the discrimination test
+ * (Phase 2) which uses the chunk→entry direction.
+ *
+ * For each entry in a cluster, compute:
+ * - Self-alignment: cosine similarity between entry embedding and its own chunk embedding
+ * - Sibling alignment: mean cosine similarity between entry embedding and other chunks' embeddings
+ * - Alignment gap: self - mean sibling (positive = entry preserves chunk-specific signal)
+ *
+ * If the gap is small, the LLM summarization shifts the entry into a
+ * cluster-generic direction, losing chunk-specific retrievability.
+ */
+import type { ClusterForAnalysis } from './similarity-analysis.js';
+/** Per-entry alignment measurement. */
+export interface EntryAlignment {
+    entryId: string;
+    /** Cosine similarity between this entry's embedding and its own chunk embedding. */
+    selfAlignment: number;
+    /** Mean cosine similarity between this entry's embedding and all sibling chunk embeddings. */
+    meanSiblingAlignment: number;
+    /** self - meanSibling (positive = preserves chunk-specific signal). */
+    alignmentGap: number;
+    /** Max cosine similarity to any sibling chunk (worst-case confusion). */
+    maxSiblingAlignment: number;
+}
+/** Per-cluster alignment summary. */
+export interface ClusterAlignmentResult {
+    clusterId: string;
+    clusterName: string | null;
+    entryCount: number;
+    /** Mean self-alignment across all entries. */
+    meanSelfAlignment: number;
+    /** Mean sibling alignment across all entries. */
+    meanSiblingAlignment: number;
+    /** Mean alignment gap (self - sibling). */
+    meanAlignmentGap: number;
+    /** Fraction of entries where self > max sibling (entry points more to own chunk than any sibling). */
+    uniquelyAlignedFraction: number;
+    /** Per-entry details. */
+    perEntry: EntryAlignment[];
+}
+/**
+ * Run alignment analysis for a single cluster.
+ */
+export declare function analyseClusterAlignment(cluster: ClusterForAnalysis): ClusterAlignmentResult;
+/**
+ * Run alignment analysis across all eligible clusters.
+ */
+export declare function runAlignmentAnalysis(clusters: ClusterForAnalysis[]): ClusterAlignmentResult[];
+//# sourceMappingURL=alignment-analysis.d.ts.map

package/dist/eval/experiments/index-differentiation/alignment-analysis.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"alignment-analysis.d.ts","sourceRoot":"","sources":["../../../../src/eval/experiments/index-differentiation/alignment-analysis.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;GAcG;AAGH,OAAO,KAAK,EAAE,kBAAkB,EAAE,MAAM,0BAA0B,CAAC;AAEnE,uCAAuC;AACvC,MAAM,WAAW,cAAc;IAC7B,OAAO,EAAE,MAAM,CAAC;IAChB,oFAAoF;IACpF,aAAa,EAAE,MAAM,CAAC;IACtB,8FAA8F;IAC9F,oBAAoB,EAAE,MAAM,CAAC;IAC7B,uEAAuE;IACvE,YAAY,EAAE,MAAM,CAAC;IACrB,yEAAyE;IACzE,mBAAmB,EAAE,MAAM,CAAC;CAC7B;AAED,qCAAqC;AACrC,MAAM,WAAW,sBAAsB;IACrC,SAAS,EAAE,MAAM,CAAC;IAClB,WAAW,EAAE,MAAM,GAAG,IAAI,CAAC;IAC3B,UAAU,EAAE,MAAM,CAAC;IACnB,8CAA8C;IAC9C,iBAAiB,EAAE,MAAM,CAAC;IAC1B,iDAAiD;IACjD,oBAAoB,EAAE,MAAM,CAAC;IAC7B,2CAA2C;IAC3C,gBAAgB,EAAE,MAAM,CAAC;IACzB,sGAAsG;IACtG,uBAAuB,EAAE,MAAM,CAAC;IAChC,yBAAyB;IACzB,QAAQ,EAAE,cAAc,EAAE,CAAC;CAC5B;AAED;;GAEG;AACH,wBAAgB,uBAAuB,CACrC,OAAO,EAAE,kBAAkB,GAC1B,sBAAsB,CAwExB;AAED;;GAEG;AACH,wBAAgB,oBAAoB,CAClC,QAAQ,EAAE,kBAAkB,EAAE,GAC7B,sBAAsB,EAAE,CAE1B"}

package/dist/eval/experiments/index-differentiation/alignment-analysis.js

@@ -0,0 +1,91 @@
+/**
+ * Phase 2b: Entry-to-chunk alignment analysis.
+ *
+ * Measures how well each index entry's embedding preserves the semantic
+ * direction of its source chunk. Complements the discrimination test
+ * (Phase 2) which uses the chunk→entry direction.
+ *
+ * For each entry in a cluster, compute:
+ * - Self-alignment: cosine similarity between entry embedding and its own chunk embedding
+ * - Sibling alignment: mean cosine similarity between entry embedding and other chunks' embeddings
+ * - Alignment gap: self - mean sibling (positive = entry preserves chunk-specific signal)
+ *
+ * If the gap is small, the LLM summarization shifts the entry into a
+ * cluster-generic direction, losing chunk-specific retrievability.
+ */
+import { cosineSimilarity } from '../../../utils/angular-distance.js';
+/**
+ * Run alignment analysis for a single cluster.
+ */
+export function analyseClusterAlignment(cluster) {
+    const entries = cluster.entries;
+    const perEntry = [];
+    for (let i = 0; i < entries.length; i++) {
+        const entry = entries[i];
+        const entryEmb = entry.entryEmbedding;
+        const ownChunkEmb = entry.chunkEmbeddings[0];
+        if (!ownChunkEmb || ownChunkEmb.length === 0) {
+            perEntry.push({
+                entryId: entry.entryId,
+                selfAlignment: 0,
+                meanSiblingAlignment: 0,
+                alignmentGap: 0,
+                maxSiblingAlignment: 0,
+            });
+            continue;
+        }
+        const selfSim = cosineSimilarity(entryEmb, ownChunkEmb);
+        // Compute similarity to all sibling chunks
+        const siblingChunkEmbs = [];
+        for (let j = 0; j < entries.length; j++) {
+            if (j === i)
+                continue;
+            for (const chunkEmb of entries[j].chunkEmbeddings) {
+                if (chunkEmb.length > 0)
+                    siblingChunkEmbs.push(chunkEmb);
+            }
+        }
+        let meanSibSim = 0;
+        let maxSibSim = 0;
+        if (siblingChunkEmbs.length > 0) {
+            const sibSims = siblingChunkEmbs.map((se) => cosineSimilarity(entryEmb, se));
+            meanSibSim = sibSims.reduce((a, b) => a + b, 0) / sibSims.length;
+            maxSibSim = Math.max(...sibSims);
+        }
+        perEntry.push({
+            entryId: entry.entryId,
+            selfAlignment: selfSim,
+            meanSiblingAlignment: meanSibSim,
+            alignmentGap: selfSim - meanSibSim,
+            maxSiblingAlignment: maxSibSim,
+        });
+    }
+    const valid = perEntry.filter((e) => e.selfAlignment > 0);
+    const meanSelf = valid.length > 0
+        ? valid.reduce((s, e) => s + e.selfAlignment, 0) / valid.length
+        : 0;
+    const meanSib = valid.length > 0
+        ? valid.reduce((s, e) => s + e.meanSiblingAlignment, 0) / valid.length
+        : 0;
+    const meanGap = valid.length > 0
+        ? valid.reduce((s, e) => s + e.alignmentGap, 0) / valid.length
+        : 0;
+    const uniquelyAligned = valid.filter((e) => e.selfAlignment > e.maxSiblingAlignment).length;
+    return {
+        clusterId: cluster.clusterId,
+        clusterName: cluster.clusterName,
+        entryCount: entries.length,
+        meanSelfAlignment: meanSelf,
+        meanSiblingAlignment: meanSib,
+        meanAlignmentGap: meanGap,
+        uniquelyAlignedFraction: valid.length > 0 ? uniquelyAligned / valid.length : 0,
+        perEntry,
+    };
+}
+/**
+ * Run alignment analysis across all eligible clusters.
+ */
+export function runAlignmentAnalysis(clusters) {
+    return clusters.map(analyseClusterAlignment);
+}
+//# sourceMappingURL=alignment-analysis.js.map

package/dist/eval/experiments/index-differentiation/alignment-analysis.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"alignment-analysis.js","sourceRoot":"","sources":["../../../../src/eval/experiments/index-differentiation/alignment-analysis.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;GAcG;AAEH,OAAO,EAAE,gBAAgB,EAAE,MAAM,oCAAoC,CAAC;AAiCtE;;GAEG;AACH,MAAM,UAAU,uBAAuB,CACrC,OAA2B;IAE3B,MAAM,OAAO,GAAG,OAAO,CAAC,OAAO,CAAC;IAChC,MAAM,QAAQ,GAAqB,EAAE,CAAC;IAEtC,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,OAAO,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;QACxC,MAAM,KAAK,GAAG,OAAO,CAAC,CAAC,CAAC,CAAC;QACzB,MAAM,QAAQ,GAAG,KAAK,CAAC,cAAc,CAAC;QACtC,MAAM,WAAW,GAAG,KAAK,CAAC,eAAe,CAAC,CAAC,CAAC,CAAC;QAE7C,IAAI,CAAC,WAAW,IAAI,WAAW,CAAC,MAAM,KAAK,CAAC,EAAE,CAAC;YAC7C,QAAQ,CAAC,IAAI,CAAC;gBACZ,OAAO,EAAE,KAAK,CAAC,OAAO;gBACtB,aAAa,EAAE,CAAC;gBAChB,oBAAoB,EAAE,CAAC;gBACvB,YAAY,EAAE,CAAC;gBACf,mBAAmB,EAAE,CAAC;aACvB,CAAC,CAAC;YACH,SAAS;QACX,CAAC;QAED,MAAM,OAAO,GAAG,gBAAgB,CAAC,QAAQ,EAAE,WAAW,CAAC,CAAC;QAExD,2CAA2C;QAC3C,MAAM,gBAAgB,GAAe,EAAE,CAAC;QACxC,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,OAAO,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;YACxC,IAAI,CAAC,KAAK,CAAC;gBAAE,SAAS;YACtB,KAAK,MAAM,QAAQ,IAAI,OAAO,CAAC,CAAC,CAAC,CAAC,eAAe,EAAE,CAAC;gBAClD,IAAI,QAAQ,CAAC,MAAM,GAAG,CAAC;oBAAE,gBAAgB,CAAC,IAAI,CAAC,QAAQ,CAAC,CAAC;YAC3D,CAAC;QACH,CAAC;QAED,IAAI,UAAU,GAAG,CAAC,CAAC;QACnB,IAAI,SAAS,GAAG,CAAC,CAAC;QAClB,IAAI,gBAAgB,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;YAChC,MAAM,OAAO,GAAG,gBAAgB,CAAC,GAAG,CAAC,CAAC,EAAE,EAAE,EAAE,CAAC,gBAAgB,CAAC,QAAQ,EAAE,EAAE,CAAC,CAAC,CAAC;YAC7E,UAAU,GAAG,OAAO,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC,GAAG,CAAC,EAAE,CAAC,CAAC,GAAG,OAAO,CAAC,MAAM,CAAC;YACjE,SAAS,GAAG,IAAI,CAAC,GAAG,CAAC,GAAG,OAAO,CAAC,CAAC;QACnC,CAAC;QAED,QAAQ,CAAC,IAAI,CAAC;YACZ,OAAO,EAAE,KAAK,CAAC,OAAO;YACtB,aAAa,EAAE,OAAO;YACtB,oBAAoB,EAAE,UAAU;YAChC,YAAY,EAAE,OAAO,GAAG,UAAU;YAClC,mBAAmB,EAAE,SAAS;SAC/B,CAAC,CAAC;IACL,CAAC;IAED,MAAM,KAAK,GAAG,QAAQ,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,aAAa,GAAG,CAAC,CAAC,CAAC;IAC1D,MAAM,QAAQ,GAAG,KAAK,CAAC,MAAM,GAAG,CAAC;QAC/B,CAAC,CAAC,KAAK,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC,GAAG,CAAC,CAAC,aAAa,EAAE,CAAC,CAAC,GAAG,KAAK,CAAC,MAAM;QAC/D,CAAC,CAAC,CAAC,CAAC;IACN,MAAM,OAAO,GAAG,KAAK,CAAC,MAAM,GAAG,CAAC;QAC9B,CAAC,CAAC,KAAK,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC,GAAG,CAAC,CAAC,oBAAoB,EAAE,CAAC,CAAC,GAAG,KAAK,CAAC,MAAM;QACtE,CAAC,CAAC,CAAC,CAAC;IACN,MAAM,OAAO,GAAG,KAAK,CAAC,MAAM,GAAG,CAAC;QAC9B,CAAC,CAAC,KAAK,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC,GAAG,CAAC,CAAC,YAAY,EAAE,CAAC,CAAC,GAAG,KAAK,CAAC,MAAM;QAC9D,CAAC,CAAC,CAAC,CAAC;IACN,MAAM,eAAe,GAAG,KAAK,CAAC,MAAM,CAClC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,aAAa,GAAG,CAAC,CAAC,mBAAmB,CAC/C,CAAC,MAAM,CAAC;IAET,OAAO;QACL,SAAS,EAAE,OAAO,CAAC,SAAS;QAC5B,WAAW,EAAE,OAAO,CAAC,WAAW;QAChC,UAAU,EAAE,OAAO,CAAC,MAAM;QAC1B,iBAAiB,EAAE,QAAQ;QAC3B,oBAAoB,EAAE,OAAO;QAC7B,gBAAgB,EAAE,OAAO;QACzB,uBAAuB,EAAE,KAAK,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC,CAAC,eAAe,GAAG,KAAK,CAAC,MAAM,CAAC,CAAC,CAAC,CAAC;QAC9E,QAAQ;KACT,CAAC;AACJ,CAAC;AAED;;GAEG;AACH,MAAM,UAAU,oBAAoB,CAClC,QAA8B;IAE9B,OAAO,QAAQ,CAAC,GAAG,CAAC,uBAAuB,CAAC,CAAC;AAC/C,CAAC"}

package/dist/eval/experiments/index-differentiation/discrimination-test.d.ts

@@ -0,0 +1,24 @@
+/**
+ * Phase 2: Discrimination test.
+ *
+ * For each index entry within a cluster, test whether its embedding
+ * is more similar to itself than to its cluster siblings.
+ *
+ * Metric: Mean Reciprocal Rank (MRR) — if we rank cluster entries by
+ * similarity to an entry's embedding, what rank does the entry itself get?
+ * Perfect discrimination = MRR 1.0 (always rank 1).
+ *
+ * This tells us: can the search system distinguish between entries in the
+ * same topic cluster, or does the compression lose chunk-specific detail?
+ */
+import type { ClusterForAnalysis } from './similarity-analysis.js';
+import type { DiscriminationResult } from './types.js';
+/**
+ * Run discrimination test for a single cluster.
+ */
+export declare function testClusterDiscrimination(cluster: ClusterForAnalysis): DiscriminationResult;
+/**
+ * Run discrimination test across all eligible clusters.
+ */
+export declare function runDiscriminationTest(clusters: ClusterForAnalysis[]): DiscriminationResult[];
+//# sourceMappingURL=discrimination-test.d.ts.map

package/dist/eval/experiments/index-differentiation/discrimination-test.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"discrimination-test.d.ts","sourceRoot":"","sources":["../../../../src/eval/experiments/index-differentiation/discrimination-test.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;GAYG;AAGH,OAAO,KAAK,EAAE,kBAAkB,EAAE,MAAM,0BAA0B,CAAC;AACnE,OAAO,KAAK,EAAE,oBAAoB,EAAE,MAAM,YAAY,CAAC;AAEvD;;GAEG;AACH,wBAAgB,yBAAyB,CACvC,OAAO,EAAE,kBAAkB,GAC1B,oBAAoB,CAgEtB;AAED;;GAEG;AACH,wBAAgB,qBAAqB,CACnC,QAAQ,EAAE,kBAAkB,EAAE,GAC7B,oBAAoB,EAAE,CAExB"}

package/dist/eval/experiments/index-differentiation/discrimination-test.js

@@ -0,0 +1,79 @@
+/**
+ * Phase 2: Discrimination test.
+ *
+ * For each index entry within a cluster, test whether its embedding
+ * is more similar to itself than to its cluster siblings.
+ *
+ * Metric: Mean Reciprocal Rank (MRR) — if we rank cluster entries by
+ * similarity to an entry's embedding, what rank does the entry itself get?
+ * Perfect discrimination = MRR 1.0 (always rank 1).
+ *
+ * This tells us: can the search system distinguish between entries in the
+ * same topic cluster, or does the compression lose chunk-specific detail?
+ */
+import { cosineSimilarity } from '../../../utils/angular-distance.js';
+/**
+ * Run discrimination test for a single cluster.
+ */
+export function testClusterDiscrimination(cluster) {
+    const entries = cluster.entries;
+    const perEntry = [];
+    for (let i = 0; i < entries.length; i++) {
+        const target = entries[i];
+        // Compute similarity of this entry's embedding to all entries' embeddings
+        const similarities = entries.map((e, j) => ({
+            index: j,
+            entryId: e.entryId,
+            similarity: cosineSimilarity(target.entryEmbedding, e.entryEmbedding),
+        }));
+        // Sort descending by similarity
+        similarities.sort((a, b) => b.similarity - a.similarity);
+        // Find rank of the target entry (self-similarity should be 1.0, rank 1)
+        // But we want to test the embedding's ability to discriminate,
+        // so we use the *chunk* embedding as the query instead of the entry embedding.
+        // This simulates: "given the actual chunk content, can we find the right index entry?"
+        const queryEmbedding = target.chunkEmbeddings[0];
+        if (!queryEmbedding || queryEmbedding.length === 0) {
+            perEntry.push({
+                entryId: target.entryId,
+                rankAmongSiblings: entries.length,
+                correctSimilarity: 0,
+                bestSiblingSimilarity: 0,
+            });
+            continue;
+        }
+        const chunkToEntrySims = entries.map((e, j) => ({
+            index: j,
+            entryId: e.entryId,
+            similarity: cosineSimilarity(queryEmbedding, e.entryEmbedding),
+        }));
+        chunkToEntrySims.sort((a, b) => b.similarity - a.similarity);
+        const rank = chunkToEntrySims.findIndex((s) => s.index === i) + 1;
+        const correctSim = chunkToEntrySims.find((s) => s.index === i).similarity;
+        const bestSiblingSim = chunkToEntrySims.find((s) => s.index !== i)?.similarity ?? 0;
+        perEntry.push({
+            entryId: target.entryId,
+            rankAmongSiblings: rank,
+            correctSimilarity: correctSim,
+            bestSiblingSimilarity: bestSiblingSim,
+        });
+    }
+    const reciprocalRanks = perEntry.map((e) => 1 / e.rankAmongSiblings);
+    const mrr = reciprocalRanks.reduce((a, b) => a + b, 0) / reciprocalRanks.length;
+    const hitRate = perEntry.filter((e) => e.rankAmongSiblings === 1).length / perEntry.length;
+    return {
+        clusterId: cluster.clusterId,
+        clusterName: cluster.clusterName,
+        entryCount: entries.length,
+        meanReciprocalRank: mrr,
+        hitRate,
+        perEntry,
+    };
+}
+/**
+ * Run discrimination test across all eligible clusters.
+ */
+export function runDiscriminationTest(clusters) {
+    return clusters.map(testClusterDiscrimination);
+}
+//# sourceMappingURL=discrimination-test.js.map

package/dist/eval/experiments/index-differentiation/discrimination-test.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"discrimination-test.js","sourceRoot":"","sources":["../../../../src/eval/experiments/index-differentiation/discrimination-test.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;GAYG;AAEH,OAAO,EAAE,gBAAgB,EAAE,MAAM,oCAAoC,CAAC;AAItE;;GAEG;AACH,MAAM,UAAU,yBAAyB,CACvC,OAA2B;IAE3B,MAAM,OAAO,GAAG,OAAO,CAAC,OAAO,CAAC;IAChC,MAAM,QAAQ,GAAqC,EAAE,CAAC;IAEtD,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,OAAO,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;QACxC,MAAM,MAAM,GAAG,OAAO,CAAC,CAAC,CAAC,CAAC;QAE1B,0EAA0E;QAC1E,MAAM,YAAY,GAAG,OAAO,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC;YAC1C,KAAK,EAAE,CAAC;YACR,OAAO,EAAE,CAAC,CAAC,OAAO;YAClB,UAAU,EAAE,gBAAgB,CAAC,MAAM,CAAC,cAAc,EAAE,CAAC,CAAC,cAAc,CAAC;SACtE,CAAC,CAAC,CAAC;QAEJ,gCAAgC;QAChC,YAAY,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,UAAU,GAAG,CAAC,CAAC,UAAU,CAAC,CAAC;QAEzD,wEAAwE;QACxE,+DAA+D;QAC/D,+EAA+E;QAC/E,uFAAuF;QACvF,MAAM,cAAc,GAAG,MAAM,CAAC,eAAe,CAAC,CAAC,CAAC,CAAC;QACjD,IAAI,CAAC,cAAc,IAAI,cAAc,CAAC,MAAM,KAAK,CAAC,EAAE,CAAC;YACnD,QAAQ,CAAC,IAAI,CAAC;gBACZ,OAAO,EAAE,MAAM,CAAC,OAAO;gBACvB,iBAAiB,EAAE,OAAO,CAAC,MAAM;gBACjC,iBAAiB,EAAE,CAAC;gBACpB,qBAAqB,EAAE,CAAC;aACzB,CAAC,CAAC;YACH,SAAS;QACX,CAAC;QAED,MAAM,gBAAgB,GAAG,OAAO,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC;YAC9C,KAAK,EAAE,CAAC;YACR,OAAO,EAAE,CAAC,CAAC,OAAO;YAClB,UAAU,EAAE,gBAAgB,CAAC,cAAc,EAAE,CAAC,CAAC,cAAc,CAAC;SAC/D,CAAC,CAAC,CAAC;QAEJ,gBAAgB,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,UAAU,GAAG,CAAC,CAAC,UAAU,CAAC,CAAC;QAE7D,MAAM,IAAI,GAAG,gBAAgB,CAAC,SAAS,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,KAAK,KAAK,CAAC,CAAC,GAAG,CAAC,CAAC;QAClE,MAAM,UAAU,GAAG,gBAAgB,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,KAAK,KAAK,CAAC,CAAE,CAAC,UAAU,CAAC;QAC3E,MAAM,cAAc,GAAG,gBAAgB,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,KAAK,KAAK,CAAC,CAAC,EAAE,UAAU,IAAI,CAAC,CAAC;QAEpF,QAAQ,CAAC,IAAI,CAAC;YACZ,OAAO,EAAE,MAAM,CAAC,OAAO;YACvB,iBAAiB,EAAE,IAAI;YACvB,iBAAiB,EAAE,UAAU;YAC7B,qBAAqB,EAAE,cAAc;SACtC,CAAC,CAAC;IACL,CAAC;IAED,MAAM,eAAe,GAAG,QAAQ,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,GAAG,CAAC,CAAC,iBAAiB,CAAC,CAAC;IACrE,MAAM,GAAG,GAAG,eAAe,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC,GAAG,CAAC,EAAE,CAAC,CAAC,GAAG,eAAe,CAAC,MAAM,CAAC;IAChF,MAAM,OAAO,GAAG,QAAQ,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,iBAAiB,KAAK,CAAC,CAAC,CAAC,MAAM,GAAG,QAAQ,CAAC,MAAM,CAAC;IAE3F,OAAO;QACL,SAAS,EAAE,OAAO,CAAC,SAAS;QAC5B,WAAW,EAAE,OAAO,CAAC,WAAW;QAChC,UAAU,EAAE,OAAO,CAAC,MAAM;QAC1B,kBAAkB,EAAE,GAAG;QACvB,OAAO;QACP,QAAQ;KACT,CAAC;AACJ,CAAC;AAED;;GAEG;AACH,MAAM,UAAU,qBAAqB,CACnC,QAA8B;IAE9B,OAAO,QAAQ,CAAC,GAAG,CAAC,yBAAyB,CAAC,CAAC;AACjD,CAAC"}

package/dist/eval/experiments/index-differentiation/index.d.ts

@@ -0,0 +1,11 @@
+/**
+ * Index entry differentiation experiment — barrel export.
+ */
+export type { ClusterSimilarityResult, DiscriminationResult, RefinementResult, DifferentiationReport, } from './types.js';
+export { runSimilarityAnalysis, analyseCluster } from './similarity-analysis.js';
+export type { ClusterForAnalysis } from './similarity-analysis.js';
+export { runDiscriminationTest, testClusterDiscrimination } from './discrimination-test.js';
+export { runAlignmentAnalysis, analyseClusterAlignment } from './alignment-analysis.js';
+export type { EntryAlignment, ClusterAlignmentResult } from './alignment-analysis.js';
+export { runRefinementTest, testClusterRefinement } from './refinement-test.js';
+//# sourceMappingURL=index.d.ts.map

package/dist/eval/experiments/index-differentiation/index.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../../../../src/eval/experiments/index-differentiation/index.ts"],"names":[],"mappings":"AAAA;;GAEG;AAEH,YAAY,EACV,uBAAuB,EACvB,oBAAoB,EACpB,gBAAgB,EAChB,qBAAqB,GACtB,MAAM,YAAY,CAAC;AAEpB,OAAO,EAAE,qBAAqB,EAAE,cAAc,EAAE,MAAM,0BAA0B,CAAC;AACjF,YAAY,EAAE,kBAAkB,EAAE,MAAM,0BAA0B,CAAC;AACnE,OAAO,EAAE,qBAAqB,EAAE,yBAAyB,EAAE,MAAM,0BAA0B,CAAC;AAC5F,OAAO,EAAE,oBAAoB,EAAE,uBAAuB,EAAE,MAAM,yBAAyB,CAAC;AACxF,YAAY,EAAE,cAAc,EAAE,sBAAsB,EAAE,MAAM,yBAAyB,CAAC;AACtF,OAAO,EAAE,iBAAiB,EAAE,qBAAqB,EAAE,MAAM,sBAAsB,CAAC"}

package/dist/eval/experiments/index-differentiation/index.js

@@ -0,0 +1,8 @@
+/**
+ * Index entry differentiation experiment — barrel export.
+ */
+export { runSimilarityAnalysis, analyseCluster } from './similarity-analysis.js';
+export { runDiscriminationTest, testClusterDiscrimination } from './discrimination-test.js';
+export { runAlignmentAnalysis, analyseClusterAlignment } from './alignment-analysis.js';
+export { runRefinementTest, testClusterRefinement } from './refinement-test.js';
+//# sourceMappingURL=index.js.map

package/dist/eval/experiments/index-differentiation/index.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"index.js","sourceRoot":"","sources":["../../../../src/eval/experiments/index-differentiation/index.ts"],"names":[],"mappings":"AAAA;;GAEG;AASH,OAAO,EAAE,qBAAqB,EAAE,cAAc,EAAE,MAAM,0BAA0B,CAAC;AAEjF,OAAO,EAAE,qBAAqB,EAAE,yBAAyB,EAAE,MAAM,0BAA0B,CAAC;AAC5F,OAAO,EAAE,oBAAoB,EAAE,uBAAuB,EAAE,MAAM,yBAAyB,CAAC;AAExF,OAAO,EAAE,iBAAiB,EAAE,qBAAqB,EAAE,MAAM,sBAAsB,CAAC"}

package/dist/eval/experiments/index-differentiation/refinement-test.d.ts

@@ -0,0 +1,32 @@
+/**
+ * Phase 3: Cluster-aware refinement simulation.
+ *
+ * For clusters with worst discrimination scores, regenerate index entries
+ * with a prompt that includes the other entries in the cluster as context.
+ * The LLM is asked to emphasize what makes THIS chunk uniquely different.
+ *
+ * Compare discrimination metrics before and after refinement.
+ */
+import Anthropic from '@anthropic-ai/sdk';
+import { Embedder } from '../../../models/embedder.js';
+import type { IndexEntry } from '../../../storage/types.js';
+import type { ClusterForAnalysis } from './similarity-analysis.js';
+import type { RefinementResult } from './types.js';
+/**
+ * Run refinement experiment on a single cluster.
+ *
+ * For each entry in the cluster:
+ * 1. Collect sibling descriptions (all other entries)
+ * 2. Re-generate this entry's description with cluster-aware prompt
+ * 3. Embed the refined description
+ * 4. Compare discrimination metrics
+ */
+export declare function testClusterRefinement(cluster: ClusterForAnalysis, entries: IndexEntry[], chunkContents: Map<string, string>, embedder: Embedder, client: Anthropic, model: string): Promise<RefinementResult>;
+/**
+ * Run refinement test on clusters with worst discrimination.
+ *
+ * @param maxClusters Maximum clusters to refine (LLM cost control)
+ */
+export declare function runRefinementTest(clusters: ClusterForAnalysis[], entries: IndexEntry[], chunkContents: Map<string, string>, discriminationScores: Map<string, number>, // clusterId → MRR
+maxClusters?: number): Promise<RefinementResult[] | null>;
+//# sourceMappingURL=refinement-test.d.ts.map

package/dist/eval/experiments/index-differentiation/refinement-test.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"refinement-test.d.ts","sourceRoot":"","sources":["../../../../src/eval/experiments/index-differentiation/refinement-test.ts"],"names":[],"mappings":"AAAA;;;;;;;;GAQG;AAEH,OAAO,SAAS,MAAM,mBAAmB,CAAC;AAE1C,OAAO,EAAE,QAAQ,EAAE,MAAM,6BAA6B,CAAC;AAGvD,OAAO,KAAK,EAAE,UAAU,EAAE,MAAM,2BAA2B,CAAC;AAC5D,OAAO,KAAK,EAAE,kBAAkB,EAAE,MAAM,0BAA0B,CAAC;AAInE,OAAO,KAAK,EAAE,gBAAgB,EAAE,MAAM,YAAY,CAAC;AAkEnD;;;;;;;;GAQG;AACH,wBAAsB,qBAAqB,CACzC,OAAO,EAAE,kBAAkB,EAC3B,OAAO,EAAE,UAAU,EAAE,EACrB,aAAa,EAAE,GAAG,CAAC,MAAM,EAAE,MAAM,CAAC,EAClC,QAAQ,EAAE,QAAQ,EAClB,MAAM,EAAE,SAAS,EACjB,KAAK,EAAE,MAAM,GACZ,OAAO,CAAC,gBAAgB,CAAC,CAyG3B;AAED;;;;GAIG;AACH,wBAAsB,iBAAiB,CACrC,QAAQ,EAAE,kBAAkB,EAAE,EAC9B,OAAO,EAAE,UAAU,EAAE,EACrB,aAAa,EAAE,GAAG,CAAC,MAAM,EAAE,MAAM,CAAC,EAClC,oBAAoB,EAAE,GAAG,CAAC,MAAM,EAAE,MAAM,CAAC,EAAE,kBAAkB;AAC7D,WAAW,GAAE,MAAU,GACtB,OAAO,CAAC,gBAAgB,EAAE,GAAG,IAAI,CAAC,CA0CpC"}

package/dist/eval/experiments/index-differentiation/refinement-test.js

@@ -0,0 +1,203 @@
+/**
+ * Phase 3: Cluster-aware refinement simulation.
+ *
+ * For clusters with worst discrimination scores, regenerate index entries
+ * with a prompt that includes the other entries in the cluster as context.
+ * The LLM is asked to emphasize what makes THIS chunk uniquely different.
+ *
+ * Compare discrimination metrics before and after refinement.
+ */
+import Anthropic from '@anthropic-ai/sdk';
+import { createSecretStore } from '../../../utils/secret-store.js';
+import { Embedder } from '../../../models/embedder.js';
+import { getModel } from '../../../models/model-registry.js';
+import { getConfig } from '../../../config/memory-config.js';
+import { testClusterDiscrimination } from './discrimination-test.js';
+import { analyseCluster } from './similarity-analysis.js';
+import { cosineSimilarity } from '../../../utils/angular-distance.js';
+/** Max entries to refine per cluster (limits API cost for large clusters). */
+const MAX_ENTRIES_PER_CLUSTER = 50;
+/**
+ * Build a refinement prompt that includes cluster context.
+ *
+ * The key difference from standard generation: the LLM sees the OTHER
+ * entries in the cluster and is instructed to emphasize what makes
+ * this chunk uniquely different from them.
+ */
+function buildRefinementPrompt(chunkContent, siblingDescriptions, targetTokens) {
+    const maxContentChars = 500 * 4; // ~500 tokens
+    const truncatedContent = chunkContent.length > maxContentChars
+        ? chunkContent.slice(0, maxContentChars) + '\n...[truncated]'
+        : chunkContent;
+    const siblingList = siblingDescriptions
+        .map((d, i) => `  ${i + 1}. ${d}`)
+        .join('\n');
+    return `You are refining a search index description for a memory system. The goal is to make this description MAXIMALLY DISTINGUISHABLE from similar entries in the same topic cluster.
+
+## Other entries in this cluster:
+${siblingList}
+
+## This chunk's content:
+${truncatedContent}
+
+Write a concise description (~${targetTokens} tokens) that:
+- Captures what is UNIQUE about this chunk compared to the cluster siblings above
+- Focuses on specific decisions, file names, error messages, or outcomes that differentiate it
+- Avoids generic terms that would match all entries in this cluster
+- Does NOT include dates, project names, or agent IDs (metadata stored separately)
+
+If someone searched for the specific topic of this chunk, your description should match THIS chunk and not the siblings.
+
+Description:`;
+}
+/**
+ * Get Anthropic client, returning null if unavailable.
+ */
+async function getClient() {
+    if (!process.env.ANTHROPIC_API_KEY) {
+        try {
+            const store = createSecretStore();
+            const storedKey = await store.get('anthropic-api-key');
+            if (storedKey) {
+                process.env.ANTHROPIC_API_KEY = storedKey;
+            }
+        }
+        catch {
+            // Keychain not available
+        }
+    }
+    if (!process.env.ANTHROPIC_API_KEY)
+        return null;
+    return new Anthropic();
+}
+/**
+ * Run refinement experiment on a single cluster.
+ *
+ * For each entry in the cluster:
+ * 1. Collect sibling descriptions (all other entries)
+ * 2. Re-generate this entry's description with cluster-aware prompt
+ * 3. Embed the refined description
+ * 4. Compare discrimination metrics
+ */
+export async function testClusterRefinement(cluster, entries, chunkContents, embedder, client, model) {
+    const config = getConfig();
+    const targetTokens = config.semanticIndex.targetDescriptionTokens;
+    // Build entry map for description lookup
+    const entryMap = new Map(entries.map((e) => [e.id, e]));
+    // Baseline discrimination
+    const baselineResult = testClusterDiscrimination(cluster);
+    const baselineSimilarity = analyseCluster(cluster);
+    // Generate refined descriptions (sample if cluster is large)
+    const refinedEntries = [];
+    const entriesToRefine = cluster.entries.length > MAX_ENTRIES_PER_CLUSTER
+        ? cluster.entries.slice(0, MAX_ENTRIES_PER_CLUSTER)
+        : cluster.entries;
+    for (const entry of entriesToRefine) {
+        const indexEntry = entryMap.get(entry.entryId);
+        if (!indexEntry)
+            continue;
+        const chunkId = indexEntry.chunkIds[0];
+        const content = chunkContents.get(chunkId);
+        if (!content)
+            continue;
+        // Collect sibling descriptions (cap at 20 nearest to keep prompt manageable)
+        const siblingDescs = cluster.entries
+            .filter((e) => e.entryId !== entry.entryId)
+            .map((e) => {
+            const desc = entryMap.get(e.entryId)?.description ?? '';
+            const sim = cosineSimilarity(entry.entryEmbedding, e.entryEmbedding);
+            return { desc, sim };
+        })
+            .filter((s) => s.desc.length > 0)
+            .sort((a, b) => b.sim - a.sim)
+            .slice(0, 20)
+            .map((s) => s.desc);
+        if (siblingDescs.length === 0)
+            continue;
+        const prompt = buildRefinementPrompt(content, siblingDescs, targetTokens);
+        try {
+            const response = await client.messages.create({
+                model,
+                max_tokens: 300,
+                messages: [{ role: 'user', content: prompt }],
+            });
+            const text = response.content[0].type === 'text' ? response.content[0].text.trim() : '';
+            if (!text)
+                continue;
+            const embedResult = await embedder.embed(text, false);
+            refinedEntries.push({
+                entryId: entry.entryId,
+                original: indexEntry.description,
+                refined: text,
+                embedding: embedResult.embedding,
+            });
+        }
+        catch (error) {
+            console.warn(`  Refinement failed for ${entry.entryId}: ${error.message}`);
+        }
+    }
+    // Build refined cluster for discrimination test
+    const refinedCluster = {
+        ...cluster,
+        entries: cluster.entries.map((e) => {
+            const refined = refinedEntries.find((r) => r.entryId === e.entryId);
+            return {
+                ...e,
+                entryEmbedding: refined?.embedding ?? e.entryEmbedding,
+            };
+        }),
+    };
+    const refinedDiscrimination = testClusterDiscrimination(refinedCluster);
+    const refinedSimilarity = analyseCluster(refinedCluster);
+    return {
+        clusterId: cluster.clusterId,
+        clusterName: cluster.clusterName,
+        entryCount: cluster.entries.length,
+        baselineMRR: baselineResult.meanReciprocalRank,
+        refinedMRR: refinedDiscrimination.meanReciprocalRank,
+        baselineHitRate: baselineResult.hitRate,
+        refinedHitRate: refinedDiscrimination.hitRate,
+        compressionRatioDelta: refinedSimilarity.compressionRatio - baselineSimilarity.compressionRatio,
+        sampleRefinements: refinedEntries.slice(0, 3).map((r) => ({
+            entryId: r.entryId,
+            original: r.original,
+            refined: r.refined,
+        })),
+    };
+}
+/**
+ * Run refinement test on clusters with worst discrimination.
+ *
+ * @param maxClusters Maximum clusters to refine (LLM cost control)
+ */
+export async function runRefinementTest(clusters, entries, chunkContents, discriminationScores, // clusterId → MRR
+maxClusters = 5) {
+    const client = await getClient();
+    if (!client) {
+        console.log('  No API key available, skipping refinement test');
+        return null;
+    }
+    const config = getConfig();
+    const embedder = new Embedder();
+    await embedder.load(getModel(config.embeddingModel));
+    try {
+        // Select clusters with worst discrimination
+        const ranked = [...discriminationScores.entries()]
+            .sort(([, a], [, b]) => a - b)
+            .slice(0, maxClusters);
+        const results = [];
+        for (const [clusterId] of ranked) {
+            const cluster = clusters.find((c) => c.clusterId === clusterId);
+            if (!cluster)
+                continue;
+            console.log(`  Refining cluster ${clusterId} (${cluster.clusterName ?? 'unnamed'}, MRR=${discriminationScores.get(clusterId)?.toFixed(3)})`);
+            const result = await testClusterRefinement(cluster, entries, chunkContents, embedder, client, config.clusterRefreshModel);
+            results.push(result);
+        }
+        return results;
+    }
+    finally {
+        await embedder.dispose();
+    }
+}
+//# sourceMappingURL=refinement-test.js.map

package/dist/eval/experiments/index-differentiation/refinement-test.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"refinement-test.js","sourceRoot":"","sources":["../../../../src/eval/experiments/index-differentiation/refinement-test.ts"],"names":[],"mappings":"AAAA;;;;;;;;GAQG;AAEH,OAAO,SAAS,MAAM,mBAAmB,CAAC;AAC1C,OAAO,EAAE,iBAAiB,EAAE,MAAM,gCAAgC,CAAC;AACnE,OAAO,EAAE,QAAQ,EAAE,MAAM,6BAA6B,CAAC;AACvD,OAAO,EAAE,QAAQ,EAAE,MAAM,mCAAmC,CAAC;AAC7D,OAAO,EAAE,SAAS,EAAE,MAAM,kCAAkC,CAAC;AAG7D,OAAO,EAAE,yBAAyB,EAAE,MAAM,0BAA0B,CAAC;AACrE,OAAO,EAAE,cAAc,EAAE,MAAM,0BAA0B,CAAC;AAC1D,OAAO,EAAE,gBAAgB,EAAE,MAAM,oCAAoC,CAAC;AAGtE,8EAA8E;AAC9E,MAAM,uBAAuB,GAAG,EAAE,CAAC;AAEnC;;;;;;GAMG;AACH,SAAS,qBAAqB,CAC5B,YAAoB,EACpB,mBAA6B,EAC7B,YAAoB;IAEpB,MAAM,eAAe,GAAG,GAAG,GAAG,CAAC,CAAC,CAAC,cAAc;IAC/C,MAAM,gBAAgB,GACpB,YAAY,CAAC,MAAM,GAAG,eAAe;QACnC,CAAC,CAAC,YAAY,CAAC,KAAK,CAAC,CAAC,EAAE,eAAe,CAAC,GAAG,kBAAkB;QAC7D,CAAC,CAAC,YAAY,CAAC;IAEnB,MAAM,WAAW,GAAG,mBAAmB;SACpC,GAAG,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,KAAK,CAAC,GAAG,CAAC,KAAK,CAAC,EAAE,CAAC;SACjC,IAAI,CAAC,IAAI,CAAC,CAAC;IAEd,OAAO;;;EAGP,WAAW;;;EAGX,gBAAgB;;gCAEc,YAAY;;;;;;;;aAQ/B,CAAC;AACd,CAAC;AAED;;GAEG;AACH,KAAK,UAAU,SAAS;IACtB,IAAI,CAAC,OAAO,CAAC,GAAG,CAAC,iBAAiB,EAAE,CAAC;QACnC,IAAI,CAAC;YACH,MAAM,KAAK,GAAG,iBAAiB,EAAE,CAAC;YAClC,MAAM,SAAS,GAAG,MAAM,KAAK,CAAC,GAAG,CAAC,mBAAmB,CAAC,CAAC;YACvD,IAAI,SAAS,EAAE,CAAC;gBACd,OAAO,CAAC,GAAG,CAAC,iBAAiB,GAAG,SAAS,CAAC;YAC5C,CAAC;QACH,CAAC;QAAC,MAAM,CAAC;YACP,yBAAyB;QAC3B,CAAC;IACH,CAAC;IAED,IAAI,CAAC,OAAO,CAAC,GAAG,CAAC,iBAAiB;QAAE,OAAO,IAAI,CAAC;IAChD,OAAO,IAAI,SAAS,EAAE,CAAC;AACzB,CAAC;AAED;;;;;;;;GAQG;AACH,MAAM,CAAC,KAAK,UAAU,qBAAqB,CACzC,OAA2B,EAC3B,OAAqB,EACrB,aAAkC,EAClC,QAAkB,EAClB,MAAiB,EACjB,KAAa;IAEb,MAAM,MAAM,GAAG,SAAS,EAAE,CAAC;IAC3B,MAAM,YAAY,GAAG,MAAM,CAAC,aAAa,CAAC,uBAAuB,CAAC;IAElE,yCAAyC;IACzC,MAAM,QAAQ,GAAG,IAAI,GAAG,CAAC,OAAO,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,CAAC,CAAC;IAExD,0BAA0B;IAC1B,MAAM,cAAc,GAAG,yBAAyB,CAAC,OAAO,CAAC,CAAC;IAC1D,MAAM,kBAAkB,GAAG,cAAc,CAAC,OAAO,CAAC,CAAC;IAEnD,6DAA6D;IAC7D,MAAM,cAAc,GAKf,EAAE,CAAC;IAER,MAAM,eAAe,GAAG,OAAO,CAAC,OAAO,CAAC,MAAM,GAAG,uBAAuB;QACtE,CAAC,CAAC,OAAO,CAAC,OAAO,CAAC,KAAK,CAAC,CAAC,EAAE,uBAAuB,CAAC;QACnD,CAAC,CAAC,OAAO,CAAC,OAAO,CAAC;IAEpB,KAAK,MAAM,KAAK,IAAI,eAAe,EAAE,CAAC;QACpC,MAAM,UAAU,GAAG,QAAQ,CAAC,GAAG,CAAC,KAAK,CAAC,OAAO,CAAC,CAAC;QAC/C,IAAI,CAAC,UAAU;YAAE,SAAS;QAE1B,MAAM,OAAO,GAAG,UAAU,CAAC,QAAQ,CAAC,CAAC,CAAC,CAAC;QACvC,MAAM,OAAO,GAAG,aAAa,CAAC,GAAG,CAAC,OAAO,CAAC,CAAC;QAC3C,IAAI,CAAC,OAAO;YAAE,SAAS;QAEvB,6EAA6E;QAC7E,MAAM,YAAY,GAAG,OAAO,CAAC,OAAO;aACjC,MAAM,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,OAAO,KAAK,KAAK,CAAC,OAAO,CAAC;aAC1C,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE;YACT,MAAM,IAAI,GAAG,QAAQ,CAAC,GAAG,CAAC,CAAC,CAAC,OAAO,CAAC,EAAE,WAAW,IAAI,EAAE,CAAC;YACxD,MAAM,GAAG,GAAG,gBAAgB,CAAC,KAAK,CAAC,cAAc,EAAE,CAAC,CAAC,cAAc,CAAC,CAAC;YACrE,OAAO,EAAE,IAAI,EAAE,GAAG,EAAE,CAAC;QACvB,CAAC,CAAC;aACD,MAAM,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,CAAC,MAAM,GAAG,CAAC,CAAC;aAChC,IAAI,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,GAAG,GAAG,CAAC,CAAC,GAAG,CAAC;aAC7B,KAAK,CAAC,CAAC,EAAE,EAAE,CAAC;aACZ,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC;QAEtB,IAAI,YAAY,CAAC,MAAM,KAAK,CAAC;YAAE,SAAS;QAExC,MAAM,MAAM,GAAG,qBAAqB,CAAC,OAAO,EAAE,YAAY,EAAE,YAAY,CAAC,CAAC;QAE1E,IAAI,CAAC;YACH,MAAM,QAAQ,GAAG,MAAM,MAAM,CAAC,QAAQ,CAAC,MAAM,CAAC;gBAC5C,KAAK;gBACL,UAAU,EAAE,GAAG;gBACf,QAAQ,EAAE,CAAC,EAAE,IAAI,EAAE,MAAM,EAAE,OAAO,EAAE,MAAM,EAAE,CAAC;aAC9C,CAAC,CAAC;YAEH,MAAM,IAAI,GACR,QAAQ,CAAC,OAAO,CAAC,CAAC,CAAC,CAAC,IAAI,KAAK,MAAM,CAAC,CAAC,CAAC,QAAQ,CAAC,OAAO,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,IAAI,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;YAC7E,IAAI,CAAC,IAAI;gBAAE,SAAS;YAEpB,MAAM,WAAW,GAAG,MAAM,QAAQ,CAAC,KAAK,CAAC,IAAI,EAAE,KAAK,CAAC,CAAC;YAEtD,cAAc,CAAC,IAAI,CAAC;gBAClB,OAAO,EAAE,KAAK,CAAC,OAAO;gBACtB,QAAQ,EAAE,UAAU,CAAC,WAAW;gBAChC,OAAO,EAAE,IAAI;gBACb,SAAS,EAAE,WAAW,CAAC,SAAS;aACjC,CAAC,CAAC;QACL,CAAC;QAAC,OAAO,KAAK,EAAE,CAAC;YACf,OAAO,CAAC,IAAI,CACV,2BAA2B,KAAK,CAAC,OAAO,KAAM,KAAe,CAAC,OAAO,EAAE,CACxE,CAAC;QACJ,CAAC;IACH,CAAC;IAED,gDAAgD;IAChD,MAAM,cAAc,GAAuB;QACzC,GAAG,OAAO;QACV,OAAO,EAAE,OAAO,CAAC,OAAO,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE;YACjC,MAAM,OAAO,GAAG,cAAc,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,OAAO,KAAK,CAAC,CAAC,OAAO,CAAC,CAAC;YACpE,OAAO;gBACL,GAAG,CAAC;gBACJ,cAAc,EAAE,OAAO,EAAE,SAAS,IAAI,CAAC,CAAC,cAAc;aACvD,CAAC;QACJ,CAAC,CAAC;KACH,CAAC;IAEF,MAAM,qBAAqB,GAAG,yBAAyB,CAAC,cAAc,CAAC,CAAC;IACxE,MAAM,iBAAiB,GAAG,cAAc,CAAC,cAAc,CAAC,CAAC;IAEzD,OAAO;QACL,SAAS,EAAE,OAAO,CAAC,SAAS;QAC5B,WAAW,EAAE,OAAO,CAAC,WAAW;QAChC,UAAU,EAAE,OAAO,CAAC,OAAO,CAAC,MAAM;QAClC,WAAW,EAAE,cAAc,CAAC,kBAAkB;QAC9C,UAAU,EAAE,qBAAqB,CAAC,kBAAkB;QACpD,eAAe,EAAE,cAAc,CAAC,OAAO;QACvC,cAAc,EAAE,qBAAqB,CAAC,OAAO;QAC7C,qBAAqB,EACnB,iBAAiB,CAAC,gBAAgB,GAAG,kBAAkB,CAAC,gBAAgB;QAC1E,iBAAiB,EAAE,cAAc,CAAC,KAAK,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC;YACxD,OAAO,EAAE,CAAC,CAAC,OAAO;YAClB,QAAQ,EAAE,CAAC,CAAC,QAAQ;YACpB,OAAO,EAAE,CAAC,CAAC,OAAO;SACnB,CAAC,CAAC;KACJ,CAAC;AACJ,CAAC;AAED;;;;GAIG;AACH,MAAM,CAAC,KAAK,UAAU,iBAAiB,CACrC,QAA8B,EAC9B,OAAqB,EACrB,aAAkC,EAClC,oBAAyC,EAAE,kBAAkB;AAC7D,cAAsB,CAAC;IAEvB,MAAM,MAAM,GAAG,MAAM,SAAS,EAAE,CAAC;IACjC,IAAI,CAAC,MAAM,EAAE,CAAC;QACZ,OAAO,CAAC,GAAG,CAAC,kDAAkD,CAAC,CAAC;QAChE,OAAO,IAAI,CAAC;IACd,CAAC;IAED,MAAM,MAAM,GAAG,SAAS,EAAE,CAAC;IAC3B,MAAM,QAAQ,GAAG,IAAI,QAAQ,EAAE,CAAC;IAChC,MAAM,QAAQ,CAAC,IAAI,CAAC,QAAQ,CAAC,MAAM,CAAC,cAAc,CAAC,CAAC,CAAC;IAErD,IAAI,CAAC;QACH,4CAA4C;QAC5C,MAAM,MAAM,GAAG,CAAC,GAAG,oBAAoB,CAAC,OAAO,EAAE,CAAC;aAC/C,IAAI,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,EAAE,CAAC,EAAE,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,GAAG,CAAC,CAAC;aAC7B,KAAK,CAAC,CAAC,EAAE,WAAW,CAAC,CAAC;QAEzB,MAAM,OAAO,GAAuB,EAAE,CAAC;QACvC,KAAK,MAAM,CAAC,SAAS,CAAC,IAAI,MAAM,EAAE,CAAC;YACjC,MAAM,OAAO,GAAG,QAAQ,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,SAAS,KAAK,SAAS,CAAC,CAAC;YAChE,IAAI,CAAC,OAAO;gBAAE,SAAS;YAEvB,OAAO,CAAC,GAAG,CACT,sBAAsB,SAAS,KAAK,OAAO,CAAC,WAAW,IAAI,SAAS,SAAS,oBAAoB,CAAC,GAAG,CAAC,SAAS,CAAC,EAAE,OAAO,CAAC,CAAC,CAAC,GAAG,CAChI,CAAC;YAEF,MAAM,MAAM,GAAG,MAAM,qBAAqB,CACxC,OAAO,EACP,OAAO,EACP,aAAa,EACb,QAAQ,EACR,MAAM,EACN,MAAM,CAAC,mBAAmB,CAC3B,CAAC;YAEF,OAAO,CAAC,IAAI,CAAC,MAAM,CAAC,CAAC;QACvB,CAAC;QAED,OAAO,OAAO,CAAC;IACjB,CAAC;YAAS,CAAC;QACT,MAAM,QAAQ,CAAC,OAAO,EAAE,CAAC;IAC3B,CAAC;AACH,CAAC"}

package/dist/eval/experiments/index-differentiation/run-experiment.d.ts

@@ -0,0 +1,20 @@
+/**
+ * Index Entry Differentiation Experiment
+ *
+ * Tests whether semantic index entries within the same topic cluster are
+ * too similar to each other, reducing the search system's ability to
+ * discriminate between chunks covering the same topic.
+ *
+ * Three phases:
+ *   1. Similarity analysis — do entries converge more than raw chunks?
+ *   2. Discrimination test — can we find the right entry among siblings?
+ *   3. Refinement simulation — does cluster-aware regeneration help?
+ *
+ * Usage:
+ *   npx tsx src/eval/experiments/index-differentiation/run-experiment.ts [--refine] [--max-clusters N]
+ *
+ * Requires: populated index entries and chunk cluster assignments.
+ * Run `npx causantic maintenance run backfill-index` first if needed.
+ */
+export {};
+//# sourceMappingURL=run-experiment.d.ts.map

package/dist/eval/experiments/index-differentiation/run-experiment.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"run-experiment.d.ts","sourceRoot":"","sources":["../../../../src/eval/experiments/index-differentiation/run-experiment.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;GAiBG"}