@dreb/coding-agent 1.18.0 → 2.0.0

package/dist/core/search/poem.d.ts

@@ -1,38 +0,0 @@
-/**
- * POEM — Pareto-Optimal Embedded Modeling, multi-metric ranking.
- *
- * Ranks search candidates across multiple relevance metrics without requiring
- * hand-tuned weights. Uses the TFPR (Top-Fraction Pareto Ranker) approach:
- * vectorized dominance matrix computation with column duplication for
- * query-type-dependent metric weighting.
- *
- * Algorithm:
- *  1. Prune: per-metric top-K → union of survivors
- *  2. Build objectives matrix with column duplication for query-type weighting
- *  3. For each objective column, sort candidates and accumulate pairwise
- *     dominance counts (duplicate columns contribute via weight multiplier)
- *  4. Compute fitness: meanDominance × (numDominating + ε) / (numSubmitting + ε)
- *  5. Sort by fitness, assign ranks
- *
- * References:
- *  - POEM paper: https://iopscience.iop.org/article/10.1088/2632-2153/ab891b
- *  - TFPR: https://github.com/merckgroup/aidd_tfpr
- *  - colourdle: https://github.com/aebrer/colourdle
- */
-import type { QueryType } from "./query-classifier.js";
-import { type MetricScores } from "./types.js";
-export interface RankedCandidate {
-    id: number;
-    scores: MetricScores;
-    rank: number;
-}
-/**
- * Rank candidates using POEM / TFPR.
- *
- * @param candidates Map of candidateId → MetricScores (all values 0–1)
- * @param queryType  Query type for column duplication weighting
- * @param topK       Per-metric pruning limit (default: 1000)
- * @returns Candidates ordered best-first with assigned ranks (0 = best)
- */
-export declare function poemRank(candidates: Map<number, MetricScores>, queryType: QueryType, topK?: number): RankedCandidate[];
-//# sourceMappingURL=poem.d.ts.map

package/dist/core/search/poem.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"poem.d.ts","sourceRoot":"","sources":["../../../src/core/search/poem.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;GAoBG;AAEH,OAAO,KAAK,EAAE,SAAS,EAAE,MAAM,uBAAuB,CAAC;AACvD,OAAO,EAAiC,KAAK,YAAY,EAAE,MAAM,YAAY,CAAC;AAM9E,MAAM,WAAW,eAAe;IAC/B,EAAE,EAAE,MAAM,CAAC;IACX,MAAM,EAAE,YAAY,CAAC;IACrB,IAAI,EAAE,MAAM,CAAC;CACb;AA4KD;;;;;;;GAOG;AACH,wBAAgB,QAAQ,CAAC,UAAU,EAAE,GAAG,CAAC,MAAM,EAAE,YAAY,CAAC,EAAE,SAAS,EAAE,SAAS,EAAE,IAAI,SAAO,GAAG,eAAe,EAAE,CA+CpH","sourcesContent":["/**\n * POEM — Pareto-Optimal Embedded Modeling, multi-metric ranking.\n *\n * Ranks search candidates across multiple relevance metrics without requiring\n * hand-tuned weights. Uses the TFPR (Top-Fraction Pareto Ranker) approach:\n * vectorized dominance matrix computation with column duplication for\n * query-type-dependent metric weighting.\n *\n * Algorithm:\n *  1. Prune: per-metric top-K → union of survivors\n *  2. Build objectives matrix with column duplication for query-type weighting\n *  3. For each objective column, sort candidates and accumulate pairwise\n *     dominance counts (duplicate columns contribute via weight multiplier)\n *  4. Compute fitness: meanDominance × (numDominating + ε) / (numSubmitting + ε)\n *  5. Sort by fitness, assign ranks\n *\n * References:\n *  - POEM paper: https://iopscience.iop.org/article/10.1088/2632-2153/ab891b\n *  - TFPR: https://github.com/merckgroup/aidd_tfpr\n *  - colourdle: https://github.com/aebrer/colourdle\n */\n\nimport type { QueryType } from \"./query-classifier.js\";\nimport { METRIC_NAMES, type MetricName, type MetricScores } from \"./types.js\";\n\n// ============================================================================\n// Public types\n// ============================================================================\n\nexport interface RankedCandidate {\n\tid: number;\n\tscores: MetricScores;\n\trank: number;\n}\n\n// ============================================================================\n// Column duplication config per query type\n// ============================================================================\n\n/** How many times each metric column appears in the objectives matrix. */\ntype ColumnWeights = Record<MetricName, number>;\n\nconst COLUMN_WEIGHTS: Record<QueryType, ColumnWeights> = {\n\tidentifier: {\n\t\tbm25: 2,\n\t\tcosine: 1,\n\t\tpathMatch: 1,\n\t\tsymbolMatch: 2,\n\t\timportGraph: 1,\n\t\tgitRecency: 1,\n\t},\n\tnatural_language: {\n\t\tbm25: 1,\n\t\tcosine: 2,\n\t\tpathMatch: 1,\n\t\tsymbolMatch: 1,\n\t\timportGraph: 1,\n\t\tgitRecency: 1,\n\t},\n\tpath_like: {\n\t\tbm25: 1,\n\t\tcosine: 1,\n\t\tpathMatch: 3,\n\t\tsymbolMatch: 1,\n\t\timportGraph: 1,\n\t\tgitRecency: 1,\n\t},\n};\n\n// ============================================================================\n// Fitness smoothing constant (avoids division by zero)\n// ============================================================================\n\nconst EPSILON = 0.05;\n\n// ============================================================================\n// Pruning\n// ============================================================================\n\n/**\n * Per-metric top-K pruning → union of surviving candidate IDs.\n */\nfunction pruneTopK(candidates: Map<number, MetricScores>, topK: number): Set<number> {\n\tif (candidates.size <= topK) {\n\t\treturn new Set(candidates.keys());\n\t}\n\n\tconst union = new Set<number>();\n\n\tfor (const metric of METRIC_NAMES) {\n\t\tconst pairs: Array<[number, number]> = [];\n\t\tfor (const [id, scores] of candidates) {\n\t\t\tpairs.push([id, scores[metric] ?? 0]);\n\t\t}\n\t\tpairs.sort((a, b) => b[1] - a[1]);\n\t\tconst limit = Math.min(topK, pairs.length);\n\t\tfor (let i = 0; i < limit; i++) {\n\t\t\tunion.add(pairs[i][0]);\n\t\t}\n\t}\n\n\treturn union;\n}\n\n// ============================================================================\n// Dominance matrix computation\n// ============================================================================\n\n/**\n * Build the dominance count matrix using the TFPR approach.\n *\n * For each objective (metric), sorts candidates and accumulates pairwise\n * dominance: if candidate i ranks above candidate j on an objective,\n * dominanceCounts[i][j] increases by the column weight.\n *\n * Duplicate columns (from column duplication) are handled by multiplying\n * the contribution by the weight rather than re-sorting — same result,\n * no redundant work.\n *\n * Uses Uint16Array to keep memory compact (max possible value per cell\n * is the sum of all weights, which is ≤ 10).\n *\n * @param scores  Dense array of MetricScores, indexed 0..n-1\n * @param weights Column weights from query type\n * @param topK    Only top-K per objective contribute to pairwise dominance\n * @returns Flat dominance count matrix [n × n] and the total weight sum\n */\nfunction buildDominanceCounts(\n\tscores: MetricScores[],\n\tweights: ColumnWeights,\n\ttopK: number,\n): [counts: Uint16Array, totalWeight: number] {\n\tconst n = scores.length;\n\tconst counts = new Uint16Array(n * n);\n\tlet totalWeight = 0;\n\n\t// Reusable index array for sorting\n\tconst indices = new Array<number>(n);\n\tfor (let i = 0; i < n; i++) indices[i] = i;\n\n\tfor (const metric of METRIC_NAMES) {\n\t\tconst weight = weights[metric];\n\t\tif (weight === 0) continue;\n\t\ttotalWeight += weight;\n\n\t\t// Sort candidates by this metric (descending)\n\t\tconst sorted = indices.slice();\n\t\tsorted.sort((a, b) => (scores[b][metric] ?? 0) - (scores[a][metric] ?? 0));\n\n\t\t// Only consider top-K for pairwise dominance\n\t\tconst k = Math.min(topK, n);\n\n\t\t// For each pair in the top-K where i ranks above j:\n\t\t// i dominates j on this objective → add weight to counts[i, j]\n\t\tfor (let ri = 0; ri < k; ri++) {\n\t\t\tconst i = sorted[ri];\n\t\t\tconst iBase = i * n;\n\t\t\tfor (let rj = ri + 1; rj < k; rj++) {\n\t\t\t\tcounts[iBase + sorted[rj]] += weight;\n\t\t\t}\n\t\t}\n\t}\n\n\treturn [counts, totalWeight];\n}\n\n/**\n * Compute fitness scores from the dominance count matrix.\n *\n * Fitness = meanDominance × (numDominating + ε) / (numSubmitting + ε)\n *\n * Where:\n *  - meanDominance = average normalized dominance across all other candidates\n *  - numDominating = count of candidates this one dominates (>50% of objectives)\n *  - numSubmitting = count of candidates this one fails to dominate (<50%)\n */\nfunction computeFitness(counts: Uint16Array, n: number, totalWeight: number): Float64Array {\n\tconst fitness = new Float64Array(n);\n\tconst threshold = totalWeight * 0.5;\n\n\tfor (let i = 0; i < n; i++) {\n\t\tlet sumDom = 0;\n\t\tlet numDominating = 0;\n\t\tlet numSubmitting = 0;\n\t\tconst iBase = i * n;\n\n\t\tfor (let j = 0; j < n; j++) {\n\t\t\tif (i === j) continue;\n\t\t\tconst count = counts[iBase + j];\n\t\t\tsumDom += count;\n\t\t\tif (count > threshold) numDominating++;\n\t\t\tif (count < threshold) numSubmitting++;\n\t\t}\n\n\t\tconst meanDom = n > 1 ? sumDom / ((n - 1) * totalWeight) : 0;\n\t\tfitness[i] = (meanDom * (numDominating + EPSILON)) / (numSubmitting + EPSILON);\n\t}\n\n\treturn fitness;\n}\n\n// ============================================================================\n// Public API\n// ============================================================================\n\n/**\n * Rank candidates using POEM / TFPR.\n *\n * @param candidates Map of candidateId → MetricScores (all values 0–1)\n * @param queryType  Query type for column duplication weighting\n * @param topK       Per-metric pruning limit (default: 1000)\n * @returns Candidates ordered best-first with assigned ranks (0 = best)\n */\nexport function poemRank(candidates: Map<number, MetricScores>, queryType: QueryType, topK = 1000): RankedCandidate[] {\n\tif (candidates.size === 0) return [];\n\n\t// 1. Prune: per-metric top-K → union\n\tconst surviving = pruneTopK(candidates, topK);\n\n\t// 2. Build dense arrays\n\tconst ids: number[] = [];\n\tconst scores: MetricScores[] = [];\n\n\tfor (const id of surviving) {\n\t\tconst s = candidates.get(id)!;\n\t\tids.push(id);\n\t\tscores.push({\n\t\t\tbm25: s.bm25 ?? 0,\n\t\t\tcosine: s.cosine ?? 0,\n\t\t\tpathMatch: s.pathMatch ?? 0,\n\t\t\tsymbolMatch: s.symbolMatch ?? 0,\n\t\t\timportGraph: s.importGraph ?? 0,\n\t\t\tgitRecency: s.gitRecency ?? 0,\n\t\t});\n\t}\n\n\tconst n = ids.length;\n\n\t// Single candidate → rank 0\n\tif (n === 1) {\n\t\treturn [{ id: ids[0], scores: scores[0], rank: 0 }];\n\t}\n\n\t// 3. Compute dominance matrix with column duplication\n\tconst weights = COLUMN_WEIGHTS[queryType];\n\tconst [counts, totalWeight] = buildDominanceCounts(scores, weights, topK);\n\n\t// 4. Compute fitness scores\n\tconst fitness = computeFitness(counts, n, totalWeight);\n\n\t// 5. Sort by fitness (descending) and assign ranks\n\tconst order = new Array<number>(n);\n\tfor (let i = 0; i < n; i++) order[i] = i;\n\torder.sort((a, b) => fitness[b] - fitness[a]);\n\n\treturn order.map((idx, rank) => ({\n\t\tid: ids[idx],\n\t\tscores: scores[idx],\n\t\trank,\n\t}));\n}\n"]}

package/dist/core/search/poem.js

@@ -1,214 +0,0 @@
-/**
- * POEM — Pareto-Optimal Embedded Modeling, multi-metric ranking.
- *
- * Ranks search candidates across multiple relevance metrics without requiring
- * hand-tuned weights. Uses the TFPR (Top-Fraction Pareto Ranker) approach:
- * vectorized dominance matrix computation with column duplication for
- * query-type-dependent metric weighting.
- *
- * Algorithm:
- *  1. Prune: per-metric top-K → union of survivors
- *  2. Build objectives matrix with column duplication for query-type weighting
- *  3. For each objective column, sort candidates and accumulate pairwise
- *     dominance counts (duplicate columns contribute via weight multiplier)
- *  4. Compute fitness: meanDominance × (numDominating + ε) / (numSubmitting + ε)
- *  5. Sort by fitness, assign ranks
- *
- * References:
- *  - POEM paper: https://iopscience.iop.org/article/10.1088/2632-2153/ab891b
- *  - TFPR: https://github.com/merckgroup/aidd_tfpr
- *  - colourdle: https://github.com/aebrer/colourdle
- */
-import { METRIC_NAMES } from "./types.js";
-const COLUMN_WEIGHTS = {
-    identifier: {
-        bm25: 2,
-        cosine: 1,
-        pathMatch: 1,
-        symbolMatch: 2,
-        importGraph: 1,
-        gitRecency: 1,
-    },
-    natural_language: {
-        bm25: 1,
-        cosine: 2,
-        pathMatch: 1,
-        symbolMatch: 1,
-        importGraph: 1,
-        gitRecency: 1,
-    },
-    path_like: {
-        bm25: 1,
-        cosine: 1,
-        pathMatch: 3,
-        symbolMatch: 1,
-        importGraph: 1,
-        gitRecency: 1,
-    },
-};
-// ============================================================================
-// Fitness smoothing constant (avoids division by zero)
-// ============================================================================
-const EPSILON = 0.05;
-// ============================================================================
-// Pruning
-// ============================================================================
-/**
- * Per-metric top-K pruning → union of surviving candidate IDs.
- */
-function pruneTopK(candidates, topK) {
-    if (candidates.size <= topK) {
-        return new Set(candidates.keys());
-    }
-    const union = new Set();
-    for (const metric of METRIC_NAMES) {
-        const pairs = [];
-        for (const [id, scores] of candidates) {
-            pairs.push([id, scores[metric] ?? 0]);
-        }
-        pairs.sort((a, b) => b[1] - a[1]);
-        const limit = Math.min(topK, pairs.length);
-        for (let i = 0; i < limit; i++) {
-            union.add(pairs[i][0]);
-        }
-    }
-    return union;
-}
-// ============================================================================
-// Dominance matrix computation
-// ============================================================================
-/**
- * Build the dominance count matrix using the TFPR approach.
- *
- * For each objective (metric), sorts candidates and accumulates pairwise
- * dominance: if candidate i ranks above candidate j on an objective,
- * dominanceCounts[i][j] increases by the column weight.
- *
- * Duplicate columns (from column duplication) are handled by multiplying
- * the contribution by the weight rather than re-sorting — same result,
- * no redundant work.
- *
- * Uses Uint16Array to keep memory compact (max possible value per cell
- * is the sum of all weights, which is ≤ 10).
- *
- * @param scores  Dense array of MetricScores, indexed 0..n-1
- * @param weights Column weights from query type
- * @param topK    Only top-K per objective contribute to pairwise dominance
- * @returns Flat dominance count matrix [n × n] and the total weight sum
- */
-function buildDominanceCounts(scores, weights, topK) {
-    const n = scores.length;
-    const counts = new Uint16Array(n * n);
-    let totalWeight = 0;
-    // Reusable index array for sorting
-    const indices = new Array(n);
-    for (let i = 0; i < n; i++)
-        indices[i] = i;
-    for (const metric of METRIC_NAMES) {
-        const weight = weights[metric];
-        if (weight === 0)
-            continue;
-        totalWeight += weight;
-        // Sort candidates by this metric (descending)
-        const sorted = indices.slice();
-        sorted.sort((a, b) => (scores[b][metric] ?? 0) - (scores[a][metric] ?? 0));
-        // Only consider top-K for pairwise dominance
-        const k = Math.min(topK, n);
-        // For each pair in the top-K where i ranks above j:
-        // i dominates j on this objective → add weight to counts[i, j]
-        for (let ri = 0; ri < k; ri++) {
-            const i = sorted[ri];
-            const iBase = i * n;
-            for (let rj = ri + 1; rj < k; rj++) {
-                counts[iBase + sorted[rj]] += weight;
-            }
-        }
-    }
-    return [counts, totalWeight];
-}
-/**
- * Compute fitness scores from the dominance count matrix.
- *
- * Fitness = meanDominance × (numDominating + ε) / (numSubmitting + ε)
- *
- * Where:
- *  - meanDominance = average normalized dominance across all other candidates
- *  - numDominating = count of candidates this one dominates (>50% of objectives)
- *  - numSubmitting = count of candidates this one fails to dominate (<50%)
- */
-function computeFitness(counts, n, totalWeight) {
-    const fitness = new Float64Array(n);
-    const threshold = totalWeight * 0.5;
-    for (let i = 0; i < n; i++) {
-        let sumDom = 0;
-        let numDominating = 0;
-        let numSubmitting = 0;
-        const iBase = i * n;
-        for (let j = 0; j < n; j++) {
-            if (i === j)
-                continue;
-            const count = counts[iBase + j];
-            sumDom += count;
-            if (count > threshold)
-                numDominating++;
-            if (count < threshold)
-                numSubmitting++;
-        }
-        const meanDom = n > 1 ? sumDom / ((n - 1) * totalWeight) : 0;
-        fitness[i] = (meanDom * (numDominating + EPSILON)) / (numSubmitting + EPSILON);
-    }
-    return fitness;
-}
-// ============================================================================
-// Public API
-// ============================================================================
-/**
- * Rank candidates using POEM / TFPR.
- *
- * @param candidates Map of candidateId → MetricScores (all values 0–1)
- * @param queryType  Query type for column duplication weighting
- * @param topK       Per-metric pruning limit (default: 1000)
- * @returns Candidates ordered best-first with assigned ranks (0 = best)
- */
-export function poemRank(candidates, queryType, topK = 1000) {
-    if (candidates.size === 0)
-        return [];
-    // 1. Prune: per-metric top-K → union
-    const surviving = pruneTopK(candidates, topK);
-    // 2. Build dense arrays
-    const ids = [];
-    const scores = [];
-    for (const id of surviving) {
-        const s = candidates.get(id);
-        ids.push(id);
-        scores.push({
-            bm25: s.bm25 ?? 0,
-            cosine: s.cosine ?? 0,
-            pathMatch: s.pathMatch ?? 0,
-            symbolMatch: s.symbolMatch ?? 0,
-            importGraph: s.importGraph ?? 0,
-            gitRecency: s.gitRecency ?? 0,
-        });
-    }
-    const n = ids.length;
-    // Single candidate → rank 0
-    if (n === 1) {
-        return [{ id: ids[0], scores: scores[0], rank: 0 }];
-    }
-    // 3. Compute dominance matrix with column duplication
-    const weights = COLUMN_WEIGHTS[queryType];
-    const [counts, totalWeight] = buildDominanceCounts(scores, weights, topK);
-    // 4. Compute fitness scores
-    const fitness = computeFitness(counts, n, totalWeight);
-    // 5. Sort by fitness (descending) and assign ranks
-    const order = new Array(n);
-    for (let i = 0; i < n; i++)
-        order[i] = i;
-    order.sort((a, b) => fitness[b] - fitness[a]);
-    return order.map((idx, rank) => ({
-        id: ids[idx],
-        scores: scores[idx],
-        rank,
-    }));
-}
-//# sourceMappingURL=poem.js.map

package/dist/core/search/poem.js.map

@@ -1 +0,0 @@
-{"version":3,"file":"poem.js","sourceRoot":"","sources":["../../../src/core/search/poem.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;GAoBG;AAGH,OAAO,EAAE,YAAY,EAAsC,MAAM,YAAY,CAAC;AAmB9E,MAAM,cAAc,GAAqC;IACxD,UAAU,EAAE;QACX,IAAI,EAAE,CAAC;QACP,MAAM,EAAE,CAAC;QACT,SAAS,EAAE,CAAC;QACZ,WAAW,EAAE,CAAC;QACd,WAAW,EAAE,CAAC;QACd,UAAU,EAAE,CAAC;KACb;IACD,gBAAgB,EAAE;QACjB,IAAI,EAAE,CAAC;QACP,MAAM,EAAE,CAAC;QACT,SAAS,EAAE,CAAC;QACZ,WAAW,EAAE,CAAC;QACd,WAAW,EAAE,CAAC;QACd,UAAU,EAAE,CAAC;KACb;IACD,SAAS,EAAE;QACV,IAAI,EAAE,CAAC;QACP,MAAM,EAAE,CAAC;QACT,SAAS,EAAE,CAAC;QACZ,WAAW,EAAE,CAAC;QACd,WAAW,EAAE,CAAC;QACd,UAAU,EAAE,CAAC;KACb;CACD,CAAC;AAEF,+EAA+E;AAC/E,uDAAuD;AACvD,+EAA+E;AAE/E,MAAM,OAAO,GAAG,IAAI,CAAC;AAErB,+EAA+E;AAC/E,UAAU;AACV,+EAA+E;AAE/E;;GAEG;AACH,SAAS,SAAS,CAAC,UAAqC,EAAE,IAAY,EAAe;IACpF,IAAI,UAAU,CAAC,IAAI,IAAI,IAAI,EAAE,CAAC;QAC7B,OAAO,IAAI,GAAG,CAAC,UAAU,CAAC,IAAI,EAAE,CAAC,CAAC;IACnC,CAAC;IAED,MAAM,KAAK,GAAG,IAAI,GAAG,EAAU,CAAC;IAEhC,KAAK,MAAM,MAAM,IAAI,YAAY,EAAE,CAAC;QACnC,MAAM,KAAK,GAA4B,EAAE,CAAC;QAC1C,KAAK,MAAM,CAAC,EAAE,EAAE,MAAM,CAAC,IAAI,UAAU,EAAE,CAAC;YACvC,KAAK,CAAC,IAAI,CAAC,CAAC,EAAE,EAAE,MAAM,CAAC,MAAM,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC;QACvC,CAAC;QACD,KAAK,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;QAClC,MAAM,KAAK,GAAG,IAAI,CAAC,GAAG,CAAC,IAAI,EAAE,KAAK,CAAC,MAAM,CAAC,CAAC;QAC3C,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,KAAK,EAAE,CAAC,EAAE,EAAE,CAAC;YAChC,KAAK,CAAC,GAAG,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;QACxB,CAAC;IACF,CAAC;IAED,OAAO,KAAK,CAAC;AAAA,CACb;AAED,+EAA+E;AAC/E,+BAA+B;AAC/B,+EAA+E;AAE/E;;;;;;;;;;;;;;;;;;GAkBG;AACH,SAAS,oBAAoB,CAC5B,MAAsB,EACtB,OAAsB,EACtB,IAAY,EACiC;IAC7C,MAAM,CAAC,GAAG,MAAM,CAAC,MAAM,CAAC;IACxB,MAAM,MAAM,GAAG,IAAI,WAAW,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC;IACtC,IAAI,WAAW,GAAG,CAAC,CAAC;IAEpB,mCAAmC;IACnC,MAAM,OAAO,GAAG,IAAI,KAAK,CAAS,CAAC,CAAC,CAAC;IACrC,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,EAAE,CAAC,EAAE;QAAE,OAAO,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC;IAE3C,KAAK,MAAM,MAAM,IAAI,YAAY,EAAE,CAAC;QACnC,MAAM,MAAM,GAAG,OAAO,CAAC,MAAM,CAAC,CAAC;QAC/B,IAAI,MAAM,KAAK,CAAC;YAAE,SAAS;QAC3B,WAAW,IAAI,MAAM,CAAC;QAEtB,8CAA8C;QAC9C,MAAM,MAAM,GAAG,OAAO,CAAC,KAAK,EAAE,CAAC;QAC/B,MAAM,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,CAAC,MAAM,CAAC,IAAI,CAAC,CAAC,GAAG,CAAC,MAAM,CAAC,CAAC,CAAC,CAAC,MAAM,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC;QAE3E,6CAA6C;QAC7C,MAAM,CAAC,GAAG,IAAI,CAAC,GAAG,CAAC,IAAI,EAAE,CAAC,CAAC,CAAC;QAE5B,oDAAoD;QACpD,iEAA+D;QAC/D,KAAK,IAAI,EAAE,GAAG,CAAC,EAAE,EAAE,GAAG,CAAC,EAAE,EAAE,EAAE,EAAE,CAAC;YAC/B,MAAM,CAAC,GAAG,MAAM,CAAC,EAAE,CAAC,CAAC;YACrB,MAAM,KAAK,GAAG,CAAC,GAAG,CAAC,CAAC;YACpB,KAAK,IAAI,EAAE,GAAG,EAAE,GAAG,CAAC,EAAE,EAAE,GAAG,CAAC,EAAE,EAAE,EAAE,EAAE,CAAC;gBACpC,MAAM,CAAC,KAAK,GAAG,MAAM,CAAC,EAAE,CAAC,CAAC,IAAI,MAAM,CAAC;YACtC,CAAC;QACF,CAAC;IACF,CAAC;IAED,OAAO,CAAC,MAAM,EAAE,WAAW,CAAC,CAAC;AAAA,CAC7B;AAED;;;;;;;;;GASG;AACH,SAAS,cAAc,CAAC,MAAmB,EAAE,CAAS,EAAE,WAAmB,EAAgB;IAC1F,MAAM,OAAO,GAAG,IAAI,YAAY,CAAC,CAAC,CAAC,CAAC;IACpC,MAAM,SAAS,GAAG,WAAW,GAAG,GAAG,CAAC;IAEpC,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC;QAC5B,IAAI,MAAM,GAAG,CAAC,CAAC;QACf,IAAI,aAAa,GAAG,CAAC,CAAC;QACtB,IAAI,aAAa,GAAG,CAAC,CAAC;QACtB,MAAM,KAAK,GAAG,CAAC,GAAG,CAAC,CAAC;QAEpB,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC;YAC5B,IAAI,CAAC,KAAK,CAAC;gBAAE,SAAS;YACtB,MAAM,KAAK,GAAG,MAAM,CAAC,KAAK,GAAG,CAAC,CAAC,CAAC;YAChC,MAAM,IAAI,KAAK,CAAC;YAChB,IAAI,KAAK,GAAG,SAAS;gBAAE,aAAa,EAAE,CAAC;YACvC,IAAI,KAAK,GAAG,SAAS;gBAAE,aAAa,EAAE,CAAC;QACxC,CAAC;QAED,MAAM,OAAO,GAAG,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,GAAG,WAAW,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;QAC7D,OAAO,CAAC,CAAC,CAAC,GAAG,CAAC,OAAO,GAAG,CAAC,aAAa,GAAG,OAAO,CAAC,CAAC,GAAG,CAAC,aAAa,GAAG,OAAO,CAAC,CAAC;IAChF,CAAC;IAED,OAAO,OAAO,CAAC;AAAA,CACf;AAED,+EAA+E;AAC/E,aAAa;AACb,+EAA+E;AAE/E;;;;;;;GAOG;AACH,MAAM,UAAU,QAAQ,CAAC,UAAqC,EAAE,SAAoB,EAAE,IAAI,GAAG,IAAI,EAAqB;IACrH,IAAI,UAAU,CAAC,IAAI,KAAK,CAAC;QAAE,OAAO,EAAE,CAAC;IAErC,uCAAqC;IACrC,MAAM,SAAS,GAAG,SAAS,CAAC,UAAU,EAAE,IAAI,CAAC,CAAC;IAE9C,wBAAwB;IACxB,MAAM,GAAG,GAAa,EAAE,CAAC;IACzB,MAAM,MAAM,GAAmB,EAAE,CAAC;IAElC,KAAK,MAAM,EAAE,IAAI,SAAS,EAAE,CAAC;QAC5B,MAAM,CAAC,GAAG,UAAU,CAAC,GAAG,CAAC,EAAE,CAAE,CAAC;QAC9B,GAAG,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC;QACb,MAAM,CAAC,IAAI,CAAC;YACX,IAAI,EAAE,CAAC,CAAC,IAAI,IAAI,CAAC;YACjB,MAAM,EAAE,CAAC,CAAC,MAAM,IAAI,CAAC;YACrB,SAAS,EAAE,CAAC,CAAC,SAAS,IAAI,CAAC;YAC3B,WAAW,EAAE,CAAC,CAAC,WAAW,IAAI,CAAC;YAC/B,WAAW,EAAE,CAAC,CAAC,WAAW,IAAI,CAAC;YAC/B,UAAU,EAAE,CAAC,CAAC,UAAU,IAAI,CAAC;SAC7B,CAAC,CAAC;IACJ,CAAC;IAED,MAAM,CAAC,GAAG,GAAG,CAAC,MAAM,CAAC;IAErB,8BAA4B;IAC5B,IAAI,CAAC,KAAK,CAAC,EAAE,CAAC;QACb,OAAO,CAAC,EAAE,EAAE,EAAE,GAAG,CAAC,CAAC,CAAC,EAAE,MAAM,EAAE,MAAM,CAAC,CAAC,CAAC,EAAE,IAAI,EAAE,CAAC,EAAE,CAAC,CAAC;IACrD,CAAC;IAED,sDAAsD;IACtD,MAAM,OAAO,GAAG,cAAc,CAAC,SAAS,CAAC,CAAC;IAC1C,MAAM,CAAC,MAAM,EAAE,WAAW,CAAC,GAAG,oBAAoB,CAAC,MAAM,EAAE,OAAO,EAAE,IAAI,CAAC,CAAC;IAE1E,4BAA4B;IAC5B,MAAM,OAAO,GAAG,cAAc,CAAC,MAAM,EAAE,CAAC,EAAE,WAAW,CAAC,CAAC;IAEvD,mDAAmD;IACnD,MAAM,KAAK,GAAG,IAAI,KAAK,CAAS,CAAC,CAAC,CAAC;IACnC,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,EAAE,CAAC,EAAE;QAAE,KAAK,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC;IACzC,KAAK,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,OAAO,CAAC,CAAC,CAAC,GAAG,OAAO,CAAC,CAAC,CAAC,CAAC,CAAC;IAE9C,OAAO,KAAK,CAAC,GAAG,CAAC,CAAC,GAAG,EAAE,IAAI,EAAE,EAAE,CAAC,CAAC;QAChC,EAAE,EAAE,GAAG,CAAC,GAAG,CAAC;QACZ,MAAM,EAAE,MAAM,CAAC,GAAG,CAAC;QACnB,IAAI;KACJ,CAAC,CAAC,CAAC;AAAA,CACJ","sourcesContent":["/**\n * POEM — Pareto-Optimal Embedded Modeling, multi-metric ranking.\n *\n * Ranks search candidates across multiple relevance metrics without requiring\n * hand-tuned weights. Uses the TFPR (Top-Fraction Pareto Ranker) approach:\n * vectorized dominance matrix computation with column duplication for\n * query-type-dependent metric weighting.\n *\n * Algorithm:\n *  1. Prune: per-metric top-K → union of survivors\n *  2. Build objectives matrix with column duplication for query-type weighting\n *  3. For each objective column, sort candidates and accumulate pairwise\n *     dominance counts (duplicate columns contribute via weight multiplier)\n *  4. Compute fitness: meanDominance × (numDominating + ε) / (numSubmitting + ε)\n *  5. Sort by fitness, assign ranks\n *\n * References:\n *  - POEM paper: https://iopscience.iop.org/article/10.1088/2632-2153/ab891b\n *  - TFPR: https://github.com/merckgroup/aidd_tfpr\n *  - colourdle: https://github.com/aebrer/colourdle\n */\n\nimport type { QueryType } from \"./query-classifier.js\";\nimport { METRIC_NAMES, type MetricName, type MetricScores } from \"./types.js\";\n\n// ============================================================================\n// Public types\n// ============================================================================\n\nexport interface RankedCandidate {\n\tid: number;\n\tscores: MetricScores;\n\trank: number;\n}\n\n// ============================================================================\n// Column duplication config per query type\n// ============================================================================\n\n/** How many times each metric column appears in the objectives matrix. */\ntype ColumnWeights = Record<MetricName, number>;\n\nconst COLUMN_WEIGHTS: Record<QueryType, ColumnWeights> = {\n\tidentifier: {\n\t\tbm25: 2,\n\t\tcosine: 1,\n\t\tpathMatch: 1,\n\t\tsymbolMatch: 2,\n\t\timportGraph: 1,\n\t\tgitRecency: 1,\n\t},\n\tnatural_language: {\n\t\tbm25: 1,\n\t\tcosine: 2,\n\t\tpathMatch: 1,\n\t\tsymbolMatch: 1,\n\t\timportGraph: 1,\n\t\tgitRecency: 1,\n\t},\n\tpath_like: {\n\t\tbm25: 1,\n\t\tcosine: 1,\n\t\tpathMatch: 3,\n\t\tsymbolMatch: 1,\n\t\timportGraph: 1,\n\t\tgitRecency: 1,\n\t},\n};\n\n// ============================================================================\n// Fitness smoothing constant (avoids division by zero)\n// ============================================================================\n\nconst EPSILON = 0.05;\n\n// ============================================================================\n// Pruning\n// ============================================================================\n\n/**\n * Per-metric top-K pruning → union of surviving candidate IDs.\n */\nfunction pruneTopK(candidates: Map<number, MetricScores>, topK: number): Set<number> {\n\tif (candidates.size <= topK) {\n\t\treturn new Set(candidates.keys());\n\t}\n\n\tconst union = new Set<number>();\n\n\tfor (const metric of METRIC_NAMES) {\n\t\tconst pairs: Array<[number, number]> = [];\n\t\tfor (const [id, scores] of candidates) {\n\t\t\tpairs.push([id, scores[metric] ?? 0]);\n\t\t}\n\t\tpairs.sort((a, b) => b[1] - a[1]);\n\t\tconst limit = Math.min(topK, pairs.length);\n\t\tfor (let i = 0; i < limit; i++) {\n\t\t\tunion.add(pairs[i][0]);\n\t\t}\n\t}\n\n\treturn union;\n}\n\n// ============================================================================\n// Dominance matrix computation\n// ============================================================================\n\n/**\n * Build the dominance count matrix using the TFPR approach.\n *\n * For each objective (metric), sorts candidates and accumulates pairwise\n * dominance: if candidate i ranks above candidate j on an objective,\n * dominanceCounts[i][j] increases by the column weight.\n *\n * Duplicate columns (from column duplication) are handled by multiplying\n * the contribution by the weight rather than re-sorting — same result,\n * no redundant work.\n *\n * Uses Uint16Array to keep memory compact (max possible value per cell\n * is the sum of all weights, which is ≤ 10).\n *\n * @param scores  Dense array of MetricScores, indexed 0..n-1\n * @param weights Column weights from query type\n * @param topK    Only top-K per objective contribute to pairwise dominance\n * @returns Flat dominance count matrix [n × n] and the total weight sum\n */\nfunction buildDominanceCounts(\n\tscores: MetricScores[],\n\tweights: ColumnWeights,\n\ttopK: number,\n): [counts: Uint16Array, totalWeight: number] {\n\tconst n = scores.length;\n\tconst counts = new Uint16Array(n * n);\n\tlet totalWeight = 0;\n\n\t// Reusable index array for sorting\n\tconst indices = new Array<number>(n);\n\tfor (let i = 0; i < n; i++) indices[i] = i;\n\n\tfor (const metric of METRIC_NAMES) {\n\t\tconst weight = weights[metric];\n\t\tif (weight === 0) continue;\n\t\ttotalWeight += weight;\n\n\t\t// Sort candidates by this metric (descending)\n\t\tconst sorted = indices.slice();\n\t\tsorted.sort((a, b) => (scores[b][metric] ?? 0) - (scores[a][metric] ?? 0));\n\n\t\t// Only consider top-K for pairwise dominance\n\t\tconst k = Math.min(topK, n);\n\n\t\t// For each pair in the top-K where i ranks above j:\n\t\t// i dominates j on this objective → add weight to counts[i, j]\n\t\tfor (let ri = 0; ri < k; ri++) {\n\t\t\tconst i = sorted[ri];\n\t\t\tconst iBase = i * n;\n\t\t\tfor (let rj = ri + 1; rj < k; rj++) {\n\t\t\t\tcounts[iBase + sorted[rj]] += weight;\n\t\t\t}\n\t\t}\n\t}\n\n\treturn [counts, totalWeight];\n}\n\n/**\n * Compute fitness scores from the dominance count matrix.\n *\n * Fitness = meanDominance × (numDominating + ε) / (numSubmitting + ε)\n *\n * Where:\n *  - meanDominance = average normalized dominance across all other candidates\n *  - numDominating = count of candidates this one dominates (>50% of objectives)\n *  - numSubmitting = count of candidates this one fails to dominate (<50%)\n */\nfunction computeFitness(counts: Uint16Array, n: number, totalWeight: number): Float64Array {\n\tconst fitness = new Float64Array(n);\n\tconst threshold = totalWeight * 0.5;\n\n\tfor (let i = 0; i < n; i++) {\n\t\tlet sumDom = 0;\n\t\tlet numDominating = 0;\n\t\tlet numSubmitting = 0;\n\t\tconst iBase = i * n;\n\n\t\tfor (let j = 0; j < n; j++) {\n\t\t\tif (i === j) continue;\n\t\t\tconst count = counts[iBase + j];\n\t\t\tsumDom += count;\n\t\t\tif (count > threshold) numDominating++;\n\t\t\tif (count < threshold) numSubmitting++;\n\t\t}\n\n\t\tconst meanDom = n > 1 ? sumDom / ((n - 1) * totalWeight) : 0;\n\t\tfitness[i] = (meanDom * (numDominating + EPSILON)) / (numSubmitting + EPSILON);\n\t}\n\n\treturn fitness;\n}\n\n// ============================================================================\n// Public API\n// ============================================================================\n\n/**\n * Rank candidates using POEM / TFPR.\n *\n * @param candidates Map of candidateId → MetricScores (all values 0–1)\n * @param queryType  Query type for column duplication weighting\n * @param topK       Per-metric pruning limit (default: 1000)\n * @returns Candidates ordered best-first with assigned ranks (0 = best)\n */\nexport function poemRank(candidates: Map<number, MetricScores>, queryType: QueryType, topK = 1000): RankedCandidate[] {\n\tif (candidates.size === 0) return [];\n\n\t// 1. Prune: per-metric top-K → union\n\tconst surviving = pruneTopK(candidates, topK);\n\n\t// 2. Build dense arrays\n\tconst ids: number[] = [];\n\tconst scores: MetricScores[] = [];\n\n\tfor (const id of surviving) {\n\t\tconst s = candidates.get(id)!;\n\t\tids.push(id);\n\t\tscores.push({\n\t\t\tbm25: s.bm25 ?? 0,\n\t\t\tcosine: s.cosine ?? 0,\n\t\t\tpathMatch: s.pathMatch ?? 0,\n\t\t\tsymbolMatch: s.symbolMatch ?? 0,\n\t\t\timportGraph: s.importGraph ?? 0,\n\t\t\tgitRecency: s.gitRecency ?? 0,\n\t\t});\n\t}\n\n\tconst n = ids.length;\n\n\t// Single candidate → rank 0\n\tif (n === 1) {\n\t\treturn [{ id: ids[0], scores: scores[0], rank: 0 }];\n\t}\n\n\t// 3. Compute dominance matrix with column duplication\n\tconst weights = COLUMN_WEIGHTS[queryType];\n\tconst [counts, totalWeight] = buildDominanceCounts(scores, weights, topK);\n\n\t// 4. Compute fitness scores\n\tconst fitness = computeFitness(counts, n, totalWeight);\n\n\t// 5. Sort by fitness (descending) and assign ranks\n\tconst order = new Array<number>(n);\n\tfor (let i = 0; i < n; i++) order[i] = i;\n\torder.sort((a, b) => fitness[b] - fitness[a]);\n\n\treturn order.map((idx, rank) => ({\n\t\tid: ids[idx],\n\t\tscores: scores[idx],\n\t\trank,\n\t}));\n}\n"]}

package/dist/core/search/query-classifier.d.ts

@@ -1,17 +0,0 @@
-/**
- * Classify search queries into types for POEM column weighting.
- *
- * Query types affect how metric columns are duplicated during ranking:
- * - identifier: emphasise BM25 and symbol-match scores
- * - path_like: emphasise path-match scores
- * - natural_language: emphasise cosine similarity scores
- */
-export type QueryType = "identifier" | "natural_language" | "path_like";
-/**
- * Classify a search query to guide POEM column weighting.
- *
- * @param query Raw user query string
- * @returns The detected query type
- */
-export declare function classifyQuery(query: string): QueryType;
-//# sourceMappingURL=query-classifier.d.ts.map

package/dist/core/search/query-classifier.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"query-classifier.d.ts","sourceRoot":"","sources":["../../../src/core/search/query-classifier.ts"],"names":[],"mappings":"AAAA;;;;;;;GAOG;AAEH,MAAM,MAAM,SAAS,GAAG,YAAY,GAAG,kBAAkB,GAAG,WAAW,CAAC;AAoBxE;;;;;GAKG;AACH,wBAAgB,aAAa,CAAC,KAAK,EAAE,MAAM,GAAG,SAAS,CAyBtD","sourcesContent":["/**\n * Classify search queries into types for POEM column weighting.\n *\n * Query types affect how metric columns are duplicated during ranking:\n * - identifier: emphasise BM25 and symbol-match scores\n * - path_like: emphasise path-match scores\n * - natural_language: emphasise cosine similarity scores\n */\n\nexport type QueryType = \"identifier\" | \"natural_language\" | \"path_like\";\n\n/** Matches camelCase or PascalCase boundaries (lowercase→uppercase). */\nconst CAMEL_RE = /[a-z][A-Z]/;\n\n/** Matches snake_case — word chars around an underscore. */\nconst SNAKE_RE = /\\w+_\\w+/;\n\n/** SCREAMING_SNAKE_CASE — two or more uppercase-letter groups joined by underscores. */\nconst SCREAMING_SNAKE_RE = /^[A-Z][A-Z0-9]*(?:_[A-Z0-9]+)+$/;\n\n/** Path separators. */\nconst PATH_SEP_RE = /[/\\\\]/;\n\n/** Dotted path like `foo.bar.baz` (3+ segments). */\nconst DOTTED_PATH_RE = /^\\w+\\.\\w+\\.\\w+/;\n\n/** File extension pattern — ends with `.ext` where ext is 1-5 alphanumeric chars. */\nconst FILE_EXT_RE = /\\.\\w{1,5}$/;\n\n/**\n * Classify a search query to guide POEM column weighting.\n *\n * @param query Raw user query string\n * @returns The detected query type\n */\nexport function classifyQuery(query: string): QueryType {\n\tconst trimmed = query.trim();\n\tif (trimmed.length === 0) return \"natural_language\";\n\n\t// --- path_like ---\n\tif (PATH_SEP_RE.test(trimmed)) return \"path_like\";\n\tif (DOTTED_PATH_RE.test(trimmed)) return \"path_like\";\n\t// File extension at end of a single token (e.g. \"config.yaml\", \"auth.ts\")\n\tconst words = trimmed.split(/\\s+/);\n\tif (words.length === 1 && FILE_EXT_RE.test(trimmed) && /\\./.test(trimmed)) {\n\t\treturn \"path_like\";\n\t}\n\n\t// --- identifier ---\n\t// Single token or short (≤3 words) with code-style naming\n\tif (words.length === 1) return \"identifier\";\n\tif (words.length <= 3) {\n\t\t// If any word looks like a code identifier, classify as identifier\n\t\tif (words.some((w) => CAMEL_RE.test(w) || SNAKE_RE.test(w) || SCREAMING_SNAKE_RE.test(w))) {\n\t\t\treturn \"identifier\";\n\t\t}\n\t}\n\n\t// --- natural_language ---\n\treturn \"natural_language\";\n}\n"]}

package/dist/core/search/query-classifier.js

@@ -1,54 +0,0 @@
-/**
- * Classify search queries into types for POEM column weighting.
- *
- * Query types affect how metric columns are duplicated during ranking:
- * - identifier: emphasise BM25 and symbol-match scores
- * - path_like: emphasise path-match scores
- * - natural_language: emphasise cosine similarity scores
- */
-/** Matches camelCase or PascalCase boundaries (lowercase→uppercase). */
-const CAMEL_RE = /[a-z][A-Z]/;
-/** Matches snake_case — word chars around an underscore. */
-const SNAKE_RE = /\w+_\w+/;
-/** SCREAMING_SNAKE_CASE — two or more uppercase-letter groups joined by underscores. */
-const SCREAMING_SNAKE_RE = /^[A-Z][A-Z0-9]*(?:_[A-Z0-9]+)+$/;
-/** Path separators. */
-const PATH_SEP_RE = /[/\\]/;
-/** Dotted path like `foo.bar.baz` (3+ segments). */
-const DOTTED_PATH_RE = /^\w+\.\w+\.\w+/;
-/** File extension pattern — ends with `.ext` where ext is 1-5 alphanumeric chars. */
-const FILE_EXT_RE = /\.\w{1,5}$/;
-/**
- * Classify a search query to guide POEM column weighting.
- *
- * @param query Raw user query string
- * @returns The detected query type
- */
-export function classifyQuery(query) {
-    const trimmed = query.trim();
-    if (trimmed.length === 0)
-        return "natural_language";
-    // --- path_like ---
-    if (PATH_SEP_RE.test(trimmed))
-        return "path_like";
-    if (DOTTED_PATH_RE.test(trimmed))
-        return "path_like";
-    // File extension at end of a single token (e.g. "config.yaml", "auth.ts")
-    const words = trimmed.split(/\s+/);
-    if (words.length === 1 && FILE_EXT_RE.test(trimmed) && /\./.test(trimmed)) {
-        return "path_like";
-    }
-    // --- identifier ---
-    // Single token or short (≤3 words) with code-style naming
-    if (words.length === 1)
-        return "identifier";
-    if (words.length <= 3) {
-        // If any word looks like a code identifier, classify as identifier
-        if (words.some((w) => CAMEL_RE.test(w) || SNAKE_RE.test(w) || SCREAMING_SNAKE_RE.test(w))) {
-            return "identifier";
-        }
-    }
-    // --- natural_language ---
-    return "natural_language";
-}
-//# sourceMappingURL=query-classifier.js.map

package/dist/core/search/query-classifier.js.map

@@ -1 +0,0 @@
-{"version":3,"file":"query-classifier.js","sourceRoot":"","sources":["../../../src/core/search/query-classifier.ts"],"names":[],"mappings":"AAAA;;;;;;;GAOG;AAIH,0EAAwE;AACxE,MAAM,QAAQ,GAAG,YAAY,CAAC;AAE9B,8DAA4D;AAC5D,MAAM,QAAQ,GAAG,SAAS,CAAC;AAE3B,0FAAwF;AACxF,MAAM,kBAAkB,GAAG,iCAAiC,CAAC;AAE7D,uBAAuB;AACvB,MAAM,WAAW,GAAG,OAAO,CAAC;AAE5B,oDAAoD;AACpD,MAAM,cAAc,GAAG,gBAAgB,CAAC;AAExC,uFAAqF;AACrF,MAAM,WAAW,GAAG,YAAY,CAAC;AAEjC;;;;;GAKG;AACH,MAAM,UAAU,aAAa,CAAC,KAAa,EAAa;IACvD,MAAM,OAAO,GAAG,KAAK,CAAC,IAAI,EAAE,CAAC;IAC7B,IAAI,OAAO,CAAC,MAAM,KAAK,CAAC;QAAE,OAAO,kBAAkB,CAAC;IAEpD,oBAAoB;IACpB,IAAI,WAAW,CAAC,IAAI,CAAC,OAAO,CAAC;QAAE,OAAO,WAAW,CAAC;IAClD,IAAI,cAAc,CAAC,IAAI,CAAC,OAAO,CAAC;QAAE,OAAO,WAAW,CAAC;IACrD,0EAA0E;IAC1E,MAAM,KAAK,GAAG,OAAO,CAAC,KAAK,CAAC,KAAK,CAAC,CAAC;IACnC,IAAI,KAAK,CAAC,MAAM,KAAK,CAAC,IAAI,WAAW,CAAC,IAAI,CAAC,OAAO,CAAC,IAAI,IAAI,CAAC,IAAI,CAAC,OAAO,CAAC,EAAE,CAAC;QAC3E,OAAO,WAAW,CAAC;IACpB,CAAC;IAED,qBAAqB;IACrB,4DAA0D;IAC1D,IAAI,KAAK,CAAC,MAAM,KAAK,CAAC;QAAE,OAAO,YAAY,CAAC;IAC5C,IAAI,KAAK,CAAC,MAAM,IAAI,CAAC,EAAE,CAAC;QACvB,mEAAmE;QACnE,IAAI,KAAK,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,QAAQ,CAAC,IAAI,CAAC,CAAC,CAAC,IAAI,QAAQ,CAAC,IAAI,CAAC,CAAC,CAAC,IAAI,kBAAkB,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC,EAAE,CAAC;YAC3F,OAAO,YAAY,CAAC;QACrB,CAAC;IACF,CAAC;IAED,2BAA2B;IAC3B,OAAO,kBAAkB,CAAC;AAAA,CAC1B","sourcesContent":["/**\n * Classify search queries into types for POEM column weighting.\n *\n * Query types affect how metric columns are duplicated during ranking:\n * - identifier: emphasise BM25 and symbol-match scores\n * - path_like: emphasise path-match scores\n * - natural_language: emphasise cosine similarity scores\n */\n\nexport type QueryType = \"identifier\" | \"natural_language\" | \"path_like\";\n\n/** Matches camelCase or PascalCase boundaries (lowercase→uppercase). */\nconst CAMEL_RE = /[a-z][A-Z]/;\n\n/** Matches snake_case — word chars around an underscore. */\nconst SNAKE_RE = /\\w+_\\w+/;\n\n/** SCREAMING_SNAKE_CASE — two or more uppercase-letter groups joined by underscores. */\nconst SCREAMING_SNAKE_RE = /^[A-Z][A-Z0-9]*(?:_[A-Z0-9]+)+$/;\n\n/** Path separators. */\nconst PATH_SEP_RE = /[/\\\\]/;\n\n/** Dotted path like `foo.bar.baz` (3+ segments). */\nconst DOTTED_PATH_RE = /^\\w+\\.\\w+\\.\\w+/;\n\n/** File extension pattern — ends with `.ext` where ext is 1-5 alphanumeric chars. */\nconst FILE_EXT_RE = /\\.\\w{1,5}$/;\n\n/**\n * Classify a search query to guide POEM column weighting.\n *\n * @param query Raw user query string\n * @returns The detected query type\n */\nexport function classifyQuery(query: string): QueryType {\n\tconst trimmed = query.trim();\n\tif (trimmed.length === 0) return \"natural_language\";\n\n\t// --- path_like ---\n\tif (PATH_SEP_RE.test(trimmed)) return \"path_like\";\n\tif (DOTTED_PATH_RE.test(trimmed)) return \"path_like\";\n\t// File extension at end of a single token (e.g. \"config.yaml\", \"auth.ts\")\n\tconst words = trimmed.split(/\\s+/);\n\tif (words.length === 1 && FILE_EXT_RE.test(trimmed) && /\\./.test(trimmed)) {\n\t\treturn \"path_like\";\n\t}\n\n\t// --- identifier ---\n\t// Single token or short (≤3 words) with code-style naming\n\tif (words.length === 1) return \"identifier\";\n\tif (words.length <= 3) {\n\t\t// If any word looks like a code identifier, classify as identifier\n\t\tif (words.some((w) => CAMEL_RE.test(w) || SNAKE_RE.test(w) || SCREAMING_SNAKE_RE.test(w))) {\n\t\t\treturn \"identifier\";\n\t\t}\n\t}\n\n\t// --- natural_language ---\n\treturn \"natural_language\";\n}\n"]}

package/dist/core/search/scanner.d.ts

@@ -1,30 +0,0 @@
-/**
- * File scanner for the semantic search subsystem.
- *
- * Discovers project files for indexing by walking the directory tree,
- * respecting .gitignore rules, and classifying files by type.
- */
-import type { FileType } from "./types.js";
-/** A file discovered by the scanner, ready for indexing. */
-export interface ScannedFile {
-    /** Path relative to the project root (posix separators). */
-    filePath: string;
-    /** Detected file type. */
-    fileType: FileType;
-    /** File modification time in milliseconds since epoch. */
-    mtime: number;
-}
-/**
- * Detect the {@link FileType} for a file path based on its extension.
- * Returns `null` for unrecognized extensions or files without an extension.
- */
-export declare function detectFileType(filePath: string): FileType | null;
-/**
- * Scan a project directory and return all indexable files.
- *
- * Walks the tree rooted at {@link projectRoot}, respects `.gitignore` rules,
- * skips binary / oversized files, and optionally includes memory files from
- * a global memory directory.
- */
-export declare function scanProject(projectRoot: string, globalMemoryDir?: string): Promise<ScannedFile[]>;
-//# sourceMappingURL=scanner.d.ts.map

package/dist/core/search/scanner.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"scanner.d.ts","sourceRoot":"","sources":["../../../src/core/search/scanner.ts"],"names":[],"mappings":"AAAA;;;;;GAKG;AAOH,OAAO,KAAK,EAAE,QAAQ,EAAE,MAAM,YAAY,CAAC;AAM3C,4DAA4D;AAC5D,MAAM,WAAW,WAAW;IAC3B,4DAA4D;IAC5D,QAAQ,EAAE,MAAM,CAAC;IACjB,0BAA0B;IAC1B,QAAQ,EAAE,QAAQ,CAAC;IACnB,0DAA0D;IAC1D,KAAK,EAAE,MAAM,CAAC;CACd;AAiED;;;GAGG;AACH,wBAAgB,cAAc,CAAC,QAAQ,EAAE,MAAM,GAAG,QAAQ,GAAG,IAAI,CAIhE;AAED;;;;;;GAMG;AACH,wBAAsB,WAAW,CAAC,WAAW,EAAE,MAAM,EAAE,eAAe,CAAC,EAAE,MAAM,GAAG,OAAO,CAAC,WAAW,EAAE,CAAC,CAgCvG","sourcesContent":["/**\n * File scanner for the semantic search subsystem.\n *\n * Discovers project files for indexing by walking the directory tree,\n * respecting .gitignore rules, and classifying files by type.\n */\n\nimport { existsSync, readdirSync, readFileSync, type Stats, statSync } from \"node:fs\";\nimport { homedir } from \"node:os\";\nimport { extname, isAbsolute, join, relative, sep } from \"node:path\";\nimport ignore from \"ignore\";\nimport { getDrebToolVisibleDirs } from \"../tools/dreb-paths.js\";\nimport type { FileType } from \"./types.js\";\n\n// ============================================================================\n// Public types\n// ============================================================================\n\n/** A file discovered by the scanner, ready for indexing. */\nexport interface ScannedFile {\n\t/** Path relative to the project root (posix separators). */\n\tfilePath: string;\n\t/** Detected file type. */\n\tfileType: FileType;\n\t/** File modification time in milliseconds since epoch. */\n\tmtime: number;\n}\n\n// ============================================================================\n// Constants\n// ============================================================================\n\n/** Maximum file size to index (1 MB). */\nconst MAX_FILE_SIZE = 1024 * 1024;\n\n/** Directories unconditionally skipped during traversal. */\nconst SKIP_DIRS = new Set([\n\t\"node_modules\",\n\t\".git\",\n\t\".dreb/index\",\n\t\".hg\",\n\t\".svn\",\n\t\"__pycache__\",\n\t\".tox\",\n\t\".venv\",\n\t\"dist\",\n\t\"build\",\n\t\".next\",\n\t\".nuxt\",\n\t\"coverage\",\n\t\".cache\",\n]);\n\n/** Extension → FileType mapping. */\nconst EXTENSION_MAP: ReadonlyMap<string, FileType> = new Map<string, FileType>([\n\t// Tree-sitter languages\n\t[\".ts\", \"typescript\"],\n\t[\".tsx\", \"tsx\"],\n\t[\".js\", \"javascript\"],\n\t[\".mjs\", \"javascript\"],\n\t[\".cjs\", \"javascript\"],\n\t[\".py\", \"python\"],\n\t[\".go\", \"go\"],\n\t[\".rs\", \"rust\"],\n\t[\".java\", \"java\"],\n\t[\".c\", \"c\"],\n\t[\".h\", \"c\"],\n\t[\".cpp\", \"cpp\"],\n\t[\".hpp\", \"cpp\"],\n\t[\".cc\", \"cpp\"],\n\t[\".cxx\", \"cpp\"],\n\t[\".hh\", \"cpp\"],\n\t[\".hxx\", \"cpp\"],\n\t// Text file types\n\t[\".md\", \"markdown\"],\n\t[\".mdx\", \"markdown\"],\n\t[\".yml\", \"yaml\"],\n\t[\".yaml\", \"yaml\"],\n\t[\".json\", \"json\"],\n\t[\".toml\", \"toml\"],\n\t[\".txt\", \"plaintext\"],\n\t[\".cfg\", \"plaintext\"],\n\t[\".ini\", \"plaintext\"],\n\t[\".env\", \"plaintext\"],\n\t[\".conf\", \"plaintext\"],\n]);\n\n// ============================================================================\n// Public API\n// ============================================================================\n\n/**\n * Detect the {@link FileType} for a file path based on its extension.\n * Returns `null` for unrecognized extensions or files without an extension.\n */\nexport function detectFileType(filePath: string): FileType | null {\n\tconst ext = extname(filePath).toLowerCase();\n\tif (!ext) return null;\n\treturn EXTENSION_MAP.get(ext) ?? null;\n}\n\n/**\n * Scan a project directory and return all indexable files.\n *\n * Walks the tree rooted at {@link projectRoot}, respects `.gitignore` rules,\n * skips binary / oversized files, and optionally includes memory files from\n * a global memory directory.\n */\nexport async function scanProject(projectRoot: string, globalMemoryDir?: string): Promise<ScannedFile[]> {\n\tconst results: ScannedFile[] = [];\n\n\t// Detect if projectRoot is the home directory — use shallow scan mode\n\t// to avoid recursing into the entire home dir (which would be catastrophic).\n\tconst isHomeDir = isHomeDirPath(projectRoot);\n\n\tif (isHomeDir) {\n\t\t// Shallow mode: only scan top-level files and ~/.dreb/memory/\n\t\tscanShallow(projectRoot, results);\n\t} else {\n\t\t// Normal mode: full recursive walk with .gitignore\n\t\tconst ig = ignore();\n\t\tloadGitignore(ig, projectRoot, projectRoot);\n\t\twalkDirectory(projectRoot, projectRoot, ig, results);\n\t}\n\n\t// Include tool-visible .dreb/ subdirs (bypasses gitignore).\n\t// In home dir mode, global memory is already handled separately below,\n\t// and we don't want to double-scan ~/.dreb/memory/.\n\tif (!isHomeDir) {\n\t\tfor (const dir of getDrebToolVisibleDirs(projectRoot)) {\n\t\t\tscanMemoryDir(dir, projectRoot, results);\n\t\t}\n\t}\n\n\t// Include global memory files if the directory exists\n\tif (globalMemoryDir && existsSync(globalMemoryDir)) {\n\t\tscanMemoryDir(globalMemoryDir, projectRoot, results);\n\t}\n\n\treturn results;\n}\n\n/** Check if a path is the user's home directory. */\nfunction isHomeDirPath(dir: string): boolean {\n\ttry {\n\t\tconst home = homedir();\n\t\t// Normalize trailing slashes for comparison\n\t\tconst normalizedDir = dir.replace(/[/\\\\]+$/, \"\");\n\t\tconst normalizedHome = home.replace(/[/\\\\]+$/, \"\");\n\t\treturn normalizedDir === normalizedHome;\n\t} catch {\n\t\treturn false;\n\t}\n}\n\n/**\n * Shallow scan mode for home directory: only index top-level files\n * (no directory recursion) to avoid scanning the entire home directory.\n * Memory files are handled separately via scanMemoryDir.\n */\nfunction scanShallow(dir: string, results: ScannedFile[]): void {\n\tlet entries: string[];\n\ttry {\n\t\tentries = readdirSync(dir);\n\t} catch {\n\t\treturn;\n\t}\n\n\tfor (const entry of entries) {\n\t\t// Skip dotfiles/dotdirs in home dir (except specific ones we want)\n\t\tif (entry.startsWith(\".\")) continue;\n\n\t\tconst fullPath = join(dir, entry);\n\n\t\tlet stats: Stats;\n\t\ttry {\n\t\t\tstats = statSync(fullPath);\n\t\t} catch {\n\t\t\tcontinue;\n\t\t}\n\n\t\t// Only index files, not directories (shallow mode)\n\t\tif (!stats.isFile()) continue;\n\t\tif (stats.size > MAX_FILE_SIZE) continue;\n\t\tif (stats.size === 0) continue;\n\n\t\tconst fileType = detectFileType(entry);\n\t\tif (!fileType) continue;\n\n\t\tresults.push({\n\t\t\tfilePath: entry,\n\t\t\tfileType,\n\t\t\tmtime: stats.mtimeMs,\n\t\t});\n\t}\n}\n\n// ============================================================================\n// Internal helpers\n// ============================================================================\n\ntype IgnoreMatcher = ReturnType<typeof ignore>;\n\n/** Convert an OS path to posix separators for ignore matching. */\nfunction toPosix(p: string): string {\n\treturn p.split(sep).join(\"/\");\n}\n\n/** Load .gitignore rules from a directory into the ignore matcher. */\nfunction loadGitignore(ig: IgnoreMatcher, dir: string, root: string): void {\n\tconst gitignorePath = join(dir, \".gitignore\");\n\tif (!existsSync(gitignorePath)) return;\n\n\ttry {\n\t\tconst content = readFileSync(gitignorePath, \"utf-8\");\n\t\tconst relDir = relative(root, dir);\n\t\tconst prefix = relDir ? `${toPosix(relDir)}/` : \"\";\n\n\t\tconst patterns = content\n\t\t\t.split(/\\r?\\n/)\n\t\t\t.map((line) => prefixPattern(line, prefix))\n\t\t\t.filter((line): line is string => line !== null);\n\n\t\tif (patterns.length > 0) {\n\t\t\tig.add(patterns);\n\t\t}\n\t} catch {\n\t\t// Unreadable .gitignore — skip silently\n\t}\n}\n\n/**\n * Prefix a .gitignore pattern with a directory path so it applies\n * correctly when matching against root-relative paths.\n */\nfunction prefixPattern(line: string, prefix: string): string | null {\n\tconst trimmed = line.trim();\n\tif (!trimmed) return null;\n\tif (trimmed.startsWith(\"#\") && !trimmed.startsWith(\"\\\\#\")) return null;\n\n\tlet pattern = line;\n\tlet negated = false;\n\n\tif (pattern.startsWith(\"!\")) {\n\t\tnegated = true;\n\t\tpattern = pattern.slice(1);\n\t} else if (pattern.startsWith(\"\\\\!\")) {\n\t\tpattern = pattern.slice(1);\n\t}\n\n\tconst prefixed = prefix ? `${prefix}${pattern}` : pattern;\n\treturn negated ? `!${prefixed}` : prefixed;\n}\n\n/**\n * Check if a directory component (relative to root) should be unconditionally skipped.\n * Handles both top-level names (\"node_modules\") and nested paths (\".dreb/index\").\n */\nfunction shouldSkipDir(relPath: string): boolean {\n\tconst posix = toPosix(relPath);\n\n\t// Check the directory name itself\n\tconst parts = posix.split(\"/\");\n\tconst name = parts[parts.length - 1];\n\tif (SKIP_DIRS.has(name)) return true;\n\n\t// Check multi-segment skip patterns (e.g. \".dreb/index\")\n\tfor (const skip of SKIP_DIRS) {\n\t\tif (skip.includes(\"/\") && (posix === skip || posix.endsWith(`/${skip}`))) {\n\t\t\treturn true;\n\t\t}\n\t}\n\n\treturn false;\n}\n\n/** Recursively walk a directory, collecting indexable files. */\nfunction walkDirectory(dir: string, root: string, ig: IgnoreMatcher, results: ScannedFile[]): void {\n\tlet entries: string[];\n\ttry {\n\t\tentries = readdirSync(dir);\n\t} catch {\n\t\treturn; // Permission denied, etc.\n\t}\n\n\tfor (const entry of entries) {\n\t\tconst fullPath = join(dir, entry);\n\t\tconst relPath = relative(root, fullPath);\n\t\tconst posixRel = toPosix(relPath);\n\n\t\tlet stats: Stats;\n\t\ttry {\n\t\t\tstats = statSync(fullPath);\n\t\t} catch {\n\t\t\tcontinue; // Broken symlink, etc.\n\t\t}\n\n\t\tif (stats.isDirectory()) {\n\t\t\t// Hard-coded skip list\n\t\t\tif (shouldSkipDir(relPath)) continue;\n\n\t\t\t// .gitignore check (directories need trailing slash)\n\t\t\tif (ig.ignores(`${posixRel}/`)) continue;\n\n\t\t\t// Load nested .gitignore before descending\n\t\t\tloadGitignore(ig, fullPath, root);\n\n\t\t\twalkDirectory(fullPath, root, ig, results);\n\t\t\tcontinue;\n\t\t}\n\n\t\tif (!stats.isFile()) continue;\n\n\t\t// .gitignore check for files\n\t\tif (ig.ignores(posixRel)) continue;\n\n\t\t// Size gate\n\t\tif (stats.size > MAX_FILE_SIZE) continue;\n\t\tif (stats.size === 0) continue;\n\n\t\t// File type detection\n\t\tconst fileType = detectFileType(entry);\n\t\tif (!fileType) continue;\n\n\t\tresults.push({\n\t\t\tfilePath: posixRel,\n\t\t\tfileType,\n\t\t\tmtime: stats.mtimeMs,\n\t\t});\n\t}\n}\n\n/**\n * Scan a memory directory (project or global) for indexable files.\n *\n * Memory directories are always fully included — no .gitignore filtering —\n * because they live outside the normal project tree or in `.dreb/` which\n * is typically gitignored.\n *\n * Paths for global memory files are stored with a `~memory/` prefix\n * to distinguish them from project files.\n */\nfunction scanMemoryDir(memoryDir: string, projectRoot: string, results: ScannedFile[], baseMemoryDir?: string): void {\n\tlet entries: string[];\n\ttry {\n\t\tentries = readdirSync(memoryDir);\n\t} catch {\n\t\treturn;\n\t}\n\n\tfor (const entry of entries) {\n\t\tconst fullPath = join(memoryDir, entry);\n\n\t\tlet stats: Stats;\n\t\ttry {\n\t\t\tstats = statSync(fullPath);\n\t\t} catch {\n\t\t\tcontinue;\n\t\t}\n\n\t\tif (stats.isDirectory()) {\n\t\t\t// Recurse into subdirectories\n\t\t\tscanMemoryDir(fullPath, projectRoot, results, baseMemoryDir ?? memoryDir);\n\t\t\tcontinue;\n\t\t}\n\n\t\tif (!stats.isFile()) continue;\n\t\tif (stats.size > MAX_FILE_SIZE) continue;\n\t\tif (stats.size === 0) continue;\n\n\t\tconst fileType = detectFileType(entry);\n\t\tif (!fileType) continue;\n\n\t\t// If the memory dir is inside the project root, use normal relative path.\n\t\t// Otherwise, use a ~memory/ prefix so paths remain unique and identifiable.\n\t\tconst rel = relative(projectRoot, fullPath);\n\t\tconst isOutsideProject = rel.startsWith(\"..\") || isAbsolute(rel);\n\t\tconst rootMemoryDir = baseMemoryDir ?? memoryDir;\n\t\tconst filePath = isOutsideProject ? `~memory/${relative(rootMemoryDir, fullPath)}` : rel;\n\n\t\tresults.push({\n\t\t\tfilePath: toPosix(filePath),\n\t\t\tfileType,\n\t\t\tmtime: stats.mtimeMs,\n\t\t});\n\t}\n}\n"]}