@kolisachint/hoocode-agent 0.4.164 → 0.4.166

package/dist/core/search/eval-live.d.ts

@@ -0,0 +1,50 @@
+/**
+ * Live-edit evaluation: the one thing the main gold set structurally cannot
+ * measure.
+ *
+ * The eval corpus is a pinned checkout that exactly matches the embedding
+ * index, so every retriever sees the same content and the grep leg looks
+ * redundant — measurably so, since it is a significant regression once BM25 is
+ * present. But an agent's corpus is *not* a clean checkout: it edits files as
+ * it works, and the daemon has no file watcher, so anything written during a
+ * session is invisible to both dense and BM25 until the next index build.
+ * Grep is the only leg that sees it.
+ *
+ * This applies edits to the corpus **after** indexing and then asks queries
+ * that only those edits can answer. A retriever backed by the index scores 0
+ * by construction; grep scores what it can actually find. That is the number
+ * the decision to keep or drop the grep leg rests on, and nothing in the main
+ * sweep produces it.
+ *
+ * Gold line ranges are resolved from anchors after the edits land, so the
+ * fixture never hardcodes line numbers for content it also defines.
+ */
+import type { EvalQuery } from "./eval.js";
+/** A file written into the corpus after indexing. */
+export interface LiveEdit {
+    /** Repo-relative POSIX path. Created if absent, overwritten if present. */
+    path: string;
+    /** Full file contents. Held inline so the edit is reproducible from the
+     *  fixture alone, with no dependency on what the corpus already held. */
+    content: string;
+    /** Why this edit exists — reporting only. */
+    note?: string;
+}
+export interface LiveEditFixture {
+    edits: LiveEdit[];
+    /** Queries answerable *only* from the edited content. */
+    queries: EvalQuery[];
+}
+export declare function loadLiveEditFixture(fixturePath: string): LiveEditFixture;
+/**
+ * Write the edits into `corpusRoot`, then resolve the queries' gold ranges
+ * against the resulting tree.
+ *
+ * Order matters: resolving before the write would look for anchors in content
+ * that does not exist yet.
+ */
+export declare function applyLiveEdits(corpusRoot: string, fixture: LiveEditFixture): {
+    queries: EvalQuery[];
+    issues: string[];
+};
+//# sourceMappingURL=eval-live.d.ts.map

package/dist/core/search/eval-live.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"eval-live.d.ts","sourceRoot":"","sources":["../../../src/core/search/eval-live.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;GAoBG;AAKH,OAAO,KAAK,EAAE,SAAS,EAAE,MAAM,WAAW,CAAC;AAG3C,qDAAqD;AACrD,MAAM,WAAW,QAAQ;IACxB,2EAA2E;IAC3E,IAAI,EAAE,MAAM,CAAC;IACb;6EACyE;IACzE,OAAO,EAAE,MAAM,CAAC;IAChB,+CAA6C;IAC7C,IAAI,CAAC,EAAE,MAAM,CAAC;CACd;AAED,MAAM,WAAW,eAAe;IAC/B,KAAK,EAAE,QAAQ,EAAE,CAAC;IAClB,yDAAyD;IACzD,OAAO,EAAE,SAAS,EAAE,CAAC;CACrB;AAED,wBAAgB,mBAAmB,CAAC,WAAW,EAAE,MAAM,GAAG,eAAe,CAExE;AAED;;;;;;GAMG;AACH,wBAAgB,cAAc,CAC7B,UAAU,EAAE,MAAM,EAClB,OAAO,EAAE,eAAe,GACtB;IAAE,OAAO,EAAE,SAAS,EAAE,CAAC;IAAC,MAAM,EAAE,MAAM,EAAE,CAAA;CAAE,CAY5C","sourcesContent":["/**\n * Live-edit evaluation: the one thing the main gold set structurally cannot\n * measure.\n *\n * The eval corpus is a pinned checkout that exactly matches the embedding\n * index, so every retriever sees the same content and the grep leg looks\n * redundant — measurably so, since it is a significant regression once BM25 is\n * present. But an agent's corpus is *not* a clean checkout: it edits files as\n * it works, and the daemon has no file watcher, so anything written during a\n * session is invisible to both dense and BM25 until the next index build.\n * Grep is the only leg that sees it.\n *\n * This applies edits to the corpus **after** indexing and then asks queries\n * that only those edits can answer. A retriever backed by the index scores 0\n * by construction; grep scores what it can actually find. That is the number\n * the decision to keep or drop the grep leg rests on, and nothing in the main\n * sweep produces it.\n *\n * Gold line ranges are resolved from anchors after the edits land, so the\n * fixture never hardcodes line numbers for content it also defines.\n */\n\nimport { mkdirSync, writeFileSync } from \"node:fs\";\nimport path from \"node:path\";\nimport { readFileSync } from \"fs\";\nimport type { EvalQuery } from \"./eval.js\";\nimport { resolveGoldSet } from \"./eval-gold.js\";\n\n/** A file written into the corpus after indexing. */\nexport interface LiveEdit {\n\t/** Repo-relative POSIX path. Created if absent, overwritten if present. */\n\tpath: string;\n\t/** Full file contents. Held inline so the edit is reproducible from the\n\t *  fixture alone, with no dependency on what the corpus already held. */\n\tcontent: string;\n\t/** Why this edit exists — reporting only. */\n\tnote?: string;\n}\n\nexport interface LiveEditFixture {\n\tedits: LiveEdit[];\n\t/** Queries answerable *only* from the edited content. */\n\tqueries: EvalQuery[];\n}\n\nexport function loadLiveEditFixture(fixturePath: string): LiveEditFixture {\n\treturn JSON.parse(readFileSync(fixturePath, \"utf-8\")) as LiveEditFixture;\n}\n\n/**\n * Write the edits into `corpusRoot`, then resolve the queries' gold ranges\n * against the resulting tree.\n *\n * Order matters: resolving before the write would look for anchors in content\n * that does not exist yet.\n */\nexport function applyLiveEdits(\n\tcorpusRoot: string,\n\tfixture: LiveEditFixture,\n): { queries: EvalQuery[]; issues: string[] } {\n\tfor (const edit of fixture.edits) {\n\t\tconst target = path.resolve(corpusRoot, edit.path);\n\t\tmkdirSync(path.dirname(target), { recursive: true });\n\t\twriteFileSync(target, edit.content);\n\t}\n\n\tconst { dataset, issues } = resolveGoldSet(corpusRoot, fixture.queries);\n\treturn {\n\t\tqueries: dataset,\n\t\tissues: issues.map((i) => `${i.queryId} [${i.path}]: ${i.problem}`),\n\t};\n}\n"]}

package/dist/core/search/eval-live.js

@@ -0,0 +1,48 @@
+/**
+ * Live-edit evaluation: the one thing the main gold set structurally cannot
+ * measure.
+ *
+ * The eval corpus is a pinned checkout that exactly matches the embedding
+ * index, so every retriever sees the same content and the grep leg looks
+ * redundant — measurably so, since it is a significant regression once BM25 is
+ * present. But an agent's corpus is *not* a clean checkout: it edits files as
+ * it works, and the daemon has no file watcher, so anything written during a
+ * session is invisible to both dense and BM25 until the next index build.
+ * Grep is the only leg that sees it.
+ *
+ * This applies edits to the corpus **after** indexing and then asks queries
+ * that only those edits can answer. A retriever backed by the index scores 0
+ * by construction; grep scores what it can actually find. That is the number
+ * the decision to keep or drop the grep leg rests on, and nothing in the main
+ * sweep produces it.
+ *
+ * Gold line ranges are resolved from anchors after the edits land, so the
+ * fixture never hardcodes line numbers for content it also defines.
+ */
+import { mkdirSync, writeFileSync } from "node:fs";
+import path from "node:path";
+import { readFileSync } from "fs";
+import { resolveGoldSet } from "./eval-gold.js";
+export function loadLiveEditFixture(fixturePath) {
+    return JSON.parse(readFileSync(fixturePath, "utf-8"));
+}
+/**
+ * Write the edits into `corpusRoot`, then resolve the queries' gold ranges
+ * against the resulting tree.
+ *
+ * Order matters: resolving before the write would look for anchors in content
+ * that does not exist yet.
+ */
+export function applyLiveEdits(corpusRoot, fixture) {
+    for (const edit of fixture.edits) {
+        const target = path.resolve(corpusRoot, edit.path);
+        mkdirSync(path.dirname(target), { recursive: true });
+        writeFileSync(target, edit.content);
+    }
+    const { dataset, issues } = resolveGoldSet(corpusRoot, fixture.queries);
+    return {
+        queries: dataset,
+        issues: issues.map((i) => `${i.queryId} [${i.path}]: ${i.problem}`),
+    };
+}
+//# sourceMappingURL=eval-live.js.map

package/dist/core/search/eval-live.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"eval-live.js","sourceRoot":"","sources":["../../../src/core/search/eval-live.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;GAoBG;AAEH,OAAO,EAAE,SAAS,EAAE,aAAa,EAAE,MAAM,SAAS,CAAC;AACnD,OAAO,IAAI,MAAM,WAAW,CAAC;AAC7B,OAAO,EAAE,YAAY,EAAE,MAAM,IAAI,CAAC;AAElC,OAAO,EAAE,cAAc,EAAE,MAAM,gBAAgB,CAAC;AAmBhD,MAAM,UAAU,mBAAmB,CAAC,WAAmB,EAAmB;IACzE,OAAO,IAAI,CAAC,KAAK,CAAC,YAAY,CAAC,WAAW,EAAE,OAAO,CAAC,CAAoB,CAAC;AAAA,CACzE;AAED;;;;;;GAMG;AACH,MAAM,UAAU,cAAc,CAC7B,UAAkB,EAClB,OAAwB,EACqB;IAC7C,KAAK,MAAM,IAAI,IAAI,OAAO,CAAC,KAAK,EAAE,CAAC;QAClC,MAAM,MAAM,GAAG,IAAI,CAAC,OAAO,CAAC,UAAU,EAAE,IAAI,CAAC,IAAI,CAAC,CAAC;QACnD,SAAS,CAAC,IAAI,CAAC,OAAO,CAAC,MAAM,CAAC,EAAE,EAAE,SAAS,EAAE,IAAI,EAAE,CAAC,CAAC;QACrD,aAAa,CAAC,MAAM,EAAE,IAAI,CAAC,OAAO,CAAC,CAAC;IACrC,CAAC;IAED,MAAM,EAAE,OAAO,EAAE,MAAM,EAAE,GAAG,cAAc,CAAC,UAAU,EAAE,OAAO,CAAC,OAAO,CAAC,CAAC;IACxE,OAAO;QACN,OAAO,EAAE,OAAO;QAChB,MAAM,EAAE,MAAM,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,GAAG,CAAC,CAAC,OAAO,KAAK,CAAC,CAAC,IAAI,MAAM,CAAC,CAAC,OAAO,EAAE,CAAC;KACnE,CAAC;AAAA,CACF","sourcesContent":["/**\n * Live-edit evaluation: the one thing the main gold set structurally cannot\n * measure.\n *\n * The eval corpus is a pinned checkout that exactly matches the embedding\n * index, so every retriever sees the same content and the grep leg looks\n * redundant — measurably so, since it is a significant regression once BM25 is\n * present. But an agent's corpus is *not* a clean checkout: it edits files as\n * it works, and the daemon has no file watcher, so anything written during a\n * session is invisible to both dense and BM25 until the next index build.\n * Grep is the only leg that sees it.\n *\n * This applies edits to the corpus **after** indexing and then asks queries\n * that only those edits can answer. A retriever backed by the index scores 0\n * by construction; grep scores what it can actually find. That is the number\n * the decision to keep or drop the grep leg rests on, and nothing in the main\n * sweep produces it.\n *\n * Gold line ranges are resolved from anchors after the edits land, so the\n * fixture never hardcodes line numbers for content it also defines.\n */\n\nimport { mkdirSync, writeFileSync } from \"node:fs\";\nimport path from \"node:path\";\nimport { readFileSync } from \"fs\";\nimport type { EvalQuery } from \"./eval.js\";\nimport { resolveGoldSet } from \"./eval-gold.js\";\n\n/** A file written into the corpus after indexing. */\nexport interface LiveEdit {\n\t/** Repo-relative POSIX path. Created if absent, overwritten if present. */\n\tpath: string;\n\t/** Full file contents. Held inline so the edit is reproducible from the\n\t *  fixture alone, with no dependency on what the corpus already held. */\n\tcontent: string;\n\t/** Why this edit exists — reporting only. */\n\tnote?: string;\n}\n\nexport interface LiveEditFixture {\n\tedits: LiveEdit[];\n\t/** Queries answerable *only* from the edited content. */\n\tqueries: EvalQuery[];\n}\n\nexport function loadLiveEditFixture(fixturePath: string): LiveEditFixture {\n\treturn JSON.parse(readFileSync(fixturePath, \"utf-8\")) as LiveEditFixture;\n}\n\n/**\n * Write the edits into `corpusRoot`, then resolve the queries' gold ranges\n * against the resulting tree.\n *\n * Order matters: resolving before the write would look for anchors in content\n * that does not exist yet.\n */\nexport function applyLiveEdits(\n\tcorpusRoot: string,\n\tfixture: LiveEditFixture,\n): { queries: EvalQuery[]; issues: string[] } {\n\tfor (const edit of fixture.edits) {\n\t\tconst target = path.resolve(corpusRoot, edit.path);\n\t\tmkdirSync(path.dirname(target), { recursive: true });\n\t\twriteFileSync(target, edit.content);\n\t}\n\n\tconst { dataset, issues } = resolveGoldSet(corpusRoot, fixture.queries);\n\treturn {\n\t\tqueries: dataset,\n\t\tissues: issues.map((i) => `${i.queryId} [${i.path}]: ${i.problem}`),\n\t};\n}\n"]}

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

@@ -2,14 +2,19 @@
  * Retrieval evaluation gate (docs/hybrid-retrieval-design.md, step 6 of the
  * shipping order).
  *
- * Measures Recall@K for lexical, semantic, and hybrid retrieval across an
- * RRF `k` sweep, against a gold set keyed by **path + optional line range
- * matched by span overlap** — never by chunkId, which is only stable per
- * index build. Recall@50 doubles as the reranker gate: a gold span that
- * never reaches the fused top-50 cannot be rescued by any reranker.
+ * Measures Recall@K and MRR for lexical, semantic, and hybrid retrieval across
+ * an RRF `k` sweep, against a gold set keyed by **path + line range matched by
+ * span overlap** — never by chunkId, which is only stable per index build.
+ * Recall@50 doubles as the reranker gate: a gold span that never reaches the
+ * fused top-50 cannot be rescued by any reranker.
  *
- * Driven by `scripts/search-eval.mjs`; the scoring logic lives here so it is
- * unit-testable without an embedding index.
+ * Rank-sensitive metrics are the point of Recall@1 and MRR: an agent reads the
+ * top result or two, so "the gold span is somewhere in the top 5" understates
+ * how much ordering matters. Recall@5/10/50 are kept for continuity with the
+ * numbers already published in the design note.
+ *
+ * Driven by `scripts/search-eval.ts` via `eval-harness.ts`; the scoring logic
+ * lives here so it is unit-testable without an embedding index.
  */
 import type { EmbsearchService } from "../embsearch/embsearch-service.js";
 import type { CandidateSpan, ResolvedSearchMode, SearchMode } from "./types.js";
@@ -19,6 +24,22 @@ export interface EvalGoldSpan {
     /** 1-based inclusive; omit both to accept any span in the file. */
     startLine?: number;
     endLine?: number;
+    /**
+     * Literal source text that must occur inside `[startLine, endLine]`. Not
+     * used for scoring — it is how the gold set survives the corpus moving
+     * underneath it: `scripts/search-eval-gold.ts` re-resolves the line range
+     * from this anchor, and the fixture test fails when an anchor no longer
+     * sits inside its recorded range. Omitted for `scope: "file"` entries.
+     */
+    anchor?: string;
+    /**
+     * `"span"` (default) scores by overlap with the recorded line range;
+     * `"file"` accepts any span in the file. File scope is deliberate for
+     * path-class queries, where the whole file *is* the answer — it is
+     * recorded explicitly so a file-level score is never mistaken for a
+     * span-level one.
+     */
+    scope?: "span" | "file";
 }
 export interface EvalQuery {
     id: string;
@@ -33,17 +54,57 @@ export interface EvalConfig {
     mode: SearchMode;
     rrfK?: number;
     rerank?: boolean;
+    /**
+     * Score this config against a daemon-side hybrid store (BM25 fused with
+     * vectors inside the Rust daemon) instead of the dense-only index. Skipped
+     * when the harness has no hybrid service, so records without one simply
+     * omit the row rather than silently scoring it as plain semantic.
+     */
+    daemonHybrid?: boolean;
+    /**
+     * Fetch the daemon's BM25 index as a separate leg and fuse it here with
+     * dense and grep. Like `daemonHybrid` it needs a hybrid store, so it is
+     * skipped when the harness has no hybrid service.
+     */
+    bm25Leg?: boolean;
+    /** Rerank with the daemon's cross-encoder rather than the deterministic
+     *  scorer. Needs a service, so it is skipped without one. */
+    crossEncoder?: boolean;
 }
+/**
+ * The sweep from the design doc — single retrievers, hybrid across k, the
+ * routed auto mode — plus reranked (`+rr`) variants for the step 7 gate.
+ *
+ * Exported because `scripts/search-eval.ts` and `test/search-eval.test.ts`
+ * both consume it. It previously lost its `export` to a knip-driven
+ * dead-export sweep (02efaab): the only importer was a `.mjs` script reading
+ * from `dist/`, which static analysis cannot see, and the eval gate silently
+ * stopped running. The TypeScript importers are the fix — do not "clean up"
+ * this export without checking them.
+ */
+export declare const EVAL_CONFIGS: readonly EvalConfig[];
 export declare function spanMatchesGold(span: CandidateSpan, gold: EvalGoldSpan): boolean;
 /** Fraction of gold spans matched by at least one of the top-`k` candidates. */
 export declare function recallAtK(candidates: readonly CandidateSpan[], gold: readonly EvalGoldSpan[], k: number): number;
+/**
+ * Mean reciprocal rank, averaged over gold spans: for each gold span, `1/rank`
+ * of the first candidate matching it (0 when it never appears).
+ *
+ * Averaging per gold span rather than taking the single first hit keeps
+ * multi-span queries (the cross-file class) honest — finding one of two
+ * required sites should not score like finding both. For single-span queries
+ * this reduces to textbook MRR.
+ */
+export declare function mrr(candidates: readonly CandidateSpan[], gold: readonly EvalGoldSpan[]): number;
 export interface EvalQueryResult {
     label: string;
     resolvedMode: ResolvedSearchMode;
     degraded: boolean;
+    recallAt1: number;
     recallAt5: number;
     recallAt10: number;
     recallAt50: number;
+    mrr: number;
 }
-export declare function evaluateQuery(cwd: string, evalQuery: EvalQuery, configs?: readonly EvalConfig[], service?: EmbsearchService): Promise<EvalQueryResult[]>;
+export declare function evaluateQuery(cwd: string, evalQuery: EvalQuery, configs?: readonly EvalConfig[], service?: EmbsearchService, hybridService?: EmbsearchService): Promise<EvalQueryResult[]>;
 //# sourceMappingURL=eval.d.ts.map

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

@@ -1 +1 @@
-{"version":3,"file":"eval.d.ts","sourceRoot":"","sources":["../../../src/core/search/eval.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;GAYG;AAEH,OAAO,KAAK,EAAE,gBAAgB,EAAE,MAAM,mCAAmC,CAAC;AAE1E,OAAO,KAAK,EAAE,aAAa,EAAE,kBAAkB,EAAE,UAAU,EAAE,MAAM,YAAY,CAAC;AAEhF,MAAM,WAAW,YAAY;IAC5B,gCAAgC;IAChC,IAAI,EAAE,MAAM,CAAC;IACb,mEAAmE;IACnE,SAAS,CAAC,EAAE,MAAM,CAAC;IACnB,OAAO,CAAC,EAAE,MAAM,CAAC;CACjB;AAED,MAAM,WAAW,SAAS;IACzB,EAAE,EAAE,MAAM,CAAC;IACX;6DACyD;IACzD,KAAK,EAAE,MAAM,CAAC;IACd,KAAK,EAAE,MAAM,CAAC;IACd,IAAI,EAAE,YAAY,EAAE,CAAC;CACrB;AAED,MAAM,WAAW,UAAU;IAC1B,KAAK,EAAE,MAAM,CAAC;IACd,IAAI,EAAE,UAAU,CAAC;IACjB,IAAI,CAAC,EAAE,MAAM,CAAC;IACd,MAAM,CAAC,EAAE,OAAO,CAAC;CACjB;AAsBD,wBAAgB,eAAe,CAAC,IAAI,EAAE,aAAa,EAAE,IAAI,EAAE,YAAY,GAAG,OAAO,CAIhF;AAED,gFAAgF;AAChF,wBAAgB,SAAS,CAAC,UAAU,EAAE,SAAS,aAAa,EAAE,EAAE,IAAI,EAAE,SAAS,YAAY,EAAE,EAAE,CAAC,EAAE,MAAM,GAAG,MAAM,CAQhH;AAED,MAAM,WAAW,eAAe;IAC/B,KAAK,EAAE,MAAM,CAAC;IACd,YAAY,EAAE,kBAAkB,CAAC;IACjC,QAAQ,EAAE,OAAO,CAAC;IAClB,SAAS,EAAE,MAAM,CAAC;IAClB,UAAU,EAAE,MAAM,CAAC;IACnB,UAAU,EAAE,MAAM,CAAC;CACnB;AAED,wBAAsB,aAAa,CAClC,GAAG,EAAE,MAAM,EACX,SAAS,EAAE,SAAS,EACpB,OAAO,GAAE,SAAS,UAAU,EAAiB,EAC7C,OAAO,CAAC,EAAE,gBAAgB,GACxB,OAAO,CAAC,eAAe,EAAE,CAAC,CAsB5B","sourcesContent":["/**\n * Retrieval evaluation gate (docs/hybrid-retrieval-design.md, step 6 of the\n * shipping order).\n *\n * Measures Recall@K for lexical, semantic, and hybrid retrieval across an\n * RRF `k` sweep, against a gold set keyed by **path + optional line range\n * matched by span overlap** — never by chunkId, which is only stable per\n * index build. Recall@50 doubles as the reranker gate: a gold span that\n * never reaches the fused top-50 cannot be rescued by any reranker.\n *\n * Driven by `scripts/search-eval.mjs`; the scoring logic lives here so it is\n * unit-testable without an embedding index.\n */\n\nimport type { EmbsearchService } from \"../embsearch/embsearch-service.js\";\nimport { retrieveCandidates } from \"./hybrid-search.js\";\nimport type { CandidateSpan, ResolvedSearchMode, SearchMode } from \"./types.js\";\n\nexport interface EvalGoldSpan {\n\t/** Repo-relative POSIX path. */\n\tpath: string;\n\t/** 1-based inclusive; omit both to accept any span in the file. */\n\tstartLine?: number;\n\tendLine?: number;\n}\n\nexport interface EvalQuery {\n\tid: string;\n\t/** Query class from the design doc (exact-symbol, path, error-fragment,\n\t *  conceptual, cross-file, boundary). Reporting only. */\n\tclass: string;\n\tquery: string;\n\tgold: EvalGoldSpan[];\n}\n\nexport interface EvalConfig {\n\tlabel: string;\n\tmode: SearchMode;\n\trrfK?: number;\n\trerank?: boolean;\n}\n\n/** The sweep from the design doc — single retrievers, hybrid across k, the\n *  routed auto mode — plus reranked (`+rr`) variants for the step 7 gate. */\nconst EVAL_CONFIGS: readonly EvalConfig[] = [\n\t{ label: \"lexical\", mode: \"lexical\" },\n\t{ label: \"semantic\", mode: \"semantic\" },\n\t{ label: \"hybrid k=0\", mode: \"hybrid\", rrfK: 0 },\n\t{ label: \"hybrid k=2\", mode: \"hybrid\", rrfK: 2 },\n\t{ label: \"hybrid k=10\", mode: \"hybrid\", rrfK: 10 },\n\t{ label: \"hybrid k=60\", mode: \"hybrid\", rrfK: 60 },\n\t{ label: \"auto\", mode: \"auto\" },\n\t{ label: \"lexical +rr\", mode: \"lexical\", rerank: true },\n\t{ label: \"semantic +rr\", mode: \"semantic\", rerank: true },\n\t{ label: \"hybrid k=2 +rr\", mode: \"hybrid\", rrfK: 2, rerank: true },\n\t{ label: \"hybrid k=60 +rr\", mode: \"hybrid\", rrfK: 60, rerank: true },\n\t{ label: \"auto +rr\", mode: \"auto\", rerank: true },\n];\n\n/** Candidates fetched per eval query — deep enough for the reranker gate. */\nconst EVAL_FETCH_LIMIT = 50;\n\nexport function spanMatchesGold(span: CandidateSpan, gold: EvalGoldSpan): boolean {\n\tif (span.path !== gold.path) return false;\n\tif (gold.startLine === undefined || gold.endLine === undefined) return true;\n\treturn span.startLine <= gold.endLine && span.endLine >= gold.startLine;\n}\n\n/** Fraction of gold spans matched by at least one of the top-`k` candidates. */\nexport function recallAtK(candidates: readonly CandidateSpan[], gold: readonly EvalGoldSpan[], k: number): number {\n\tif (gold.length === 0) return 0;\n\tconst top = candidates.slice(0, k);\n\tlet matched = 0;\n\tfor (const g of gold) {\n\t\tif (top.some((c) => spanMatchesGold(c, g))) matched++;\n\t}\n\treturn matched / gold.length;\n}\n\nexport interface EvalQueryResult {\n\tlabel: string;\n\tresolvedMode: ResolvedSearchMode;\n\tdegraded: boolean;\n\trecallAt5: number;\n\trecallAt10: number;\n\trecallAt50: number;\n}\n\nexport async function evaluateQuery(\n\tcwd: string,\n\tevalQuery: EvalQuery,\n\tconfigs: readonly EvalConfig[] = EVAL_CONFIGS,\n\tservice?: EmbsearchService,\n): Promise<EvalQueryResult[]> {\n\tconst results: EvalQueryResult[] = [];\n\tfor (const config of configs) {\n\t\tconst retrieved = await retrieveCandidates({\n\t\t\tcwd,\n\t\t\tquery: evalQuery.query,\n\t\t\tmode: config.mode,\n\t\t\trrfK: config.rrfK,\n\t\t\trerank: config.rerank ?? false,\n\t\t\tlimit: EVAL_FETCH_LIMIT,\n\t\t\tservice,\n\t\t});\n\t\tresults.push({\n\t\t\tlabel: config.label,\n\t\t\tresolvedMode: retrieved.resolvedMode,\n\t\t\tdegraded: retrieved.degradedReason !== undefined,\n\t\t\trecallAt5: recallAtK(retrieved.candidates, evalQuery.gold, 5),\n\t\t\trecallAt10: recallAtK(retrieved.candidates, evalQuery.gold, 10),\n\t\t\trecallAt50: recallAtK(retrieved.candidates, evalQuery.gold, 50),\n\t\t});\n\t}\n\treturn results;\n}\n"]}
+{"version":3,"file":"eval.d.ts","sourceRoot":"","sources":["../../../src/core/search/eval.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;GAiBG;AAEH,OAAO,KAAK,EAAE,gBAAgB,EAAE,MAAM,mCAAmC,CAAC;AAE1E,OAAO,KAAK,EAAE,aAAa,EAAE,kBAAkB,EAAE,UAAU,EAAE,MAAM,YAAY,CAAC;AAEhF,MAAM,WAAW,YAAY;IAC5B,gCAAgC;IAChC,IAAI,EAAE,MAAM,CAAC;IACb,mEAAmE;IACnE,SAAS,CAAC,EAAE,MAAM,CAAC;IACnB,OAAO,CAAC,EAAE,MAAM,CAAC;IACjB;;;;;;OAMG;IACH,MAAM,CAAC,EAAE,MAAM,CAAC;IAChB;;;;;;OAMG;IACH,KAAK,CAAC,EAAE,MAAM,GAAG,MAAM,CAAC;CACxB;AAED,MAAM,WAAW,SAAS;IACzB,EAAE,EAAE,MAAM,CAAC;IACX;6DACyD;IACzD,KAAK,EAAE,MAAM,CAAC;IACd,KAAK,EAAE,MAAM,CAAC;IACd,IAAI,EAAE,YAAY,EAAE,CAAC;CACrB;AAED,MAAM,WAAW,UAAU;IAC1B,KAAK,EAAE,MAAM,CAAC;IACd,IAAI,EAAE,UAAU,CAAC;IACjB,IAAI,CAAC,EAAE,MAAM,CAAC;IACd,MAAM,CAAC,EAAE,OAAO,CAAC;IACjB;;;;;OAKG;IACH,YAAY,CAAC,EAAE,OAAO,CAAC;IACvB;;;;OAIG;IACH,OAAO,CAAC,EAAE,OAAO,CAAC;IAClB;iEAC6D;IAC7D,YAAY,CAAC,EAAE,OAAO,CAAC;CACvB;AAED;;;;;;;;;;GAUG;AACH,eAAO,MAAM,YAAY,EAAE,SAAS,UAAU,EA6B7C,CAAC;AAKF,wBAAgB,eAAe,CAAC,IAAI,EAAE,aAAa,EAAE,IAAI,EAAE,YAAY,GAAG,OAAO,CAKhF;AAED,gFAAgF;AAChF,wBAAgB,SAAS,CAAC,UAAU,EAAE,SAAS,aAAa,EAAE,EAAE,IAAI,EAAE,SAAS,YAAY,EAAE,EAAE,CAAC,EAAE,MAAM,GAAG,MAAM,CAQhH;AAED;;;;;;;;GAQG;AACH,wBAAgB,GAAG,CAAC,UAAU,EAAE,SAAS,aAAa,EAAE,EAAE,IAAI,EAAE,SAAS,YAAY,EAAE,GAAG,MAAM,CAQ/F;AAED,MAAM,WAAW,eAAe;IAC/B,KAAK,EAAE,MAAM,CAAC;IACd,YAAY,EAAE,kBAAkB,CAAC;IACjC,QAAQ,EAAE,OAAO,CAAC;IAClB,SAAS,EAAE,MAAM,CAAC;IAClB,SAAS,EAAE,MAAM,CAAC;IAClB,UAAU,EAAE,MAAM,CAAC;IACnB,UAAU,EAAE,MAAM,CAAC;IACnB,GAAG,EAAE,MAAM,CAAC;CACZ;AAED,wBAAsB,aAAa,CAClC,GAAG,EAAE,MAAM,EACX,SAAS,EAAE,SAAS,EACpB,OAAO,GAAE,SAAS,UAAU,EAAiB,EAC7C,OAAO,CAAC,EAAE,gBAAgB,EAC1B,aAAa,CAAC,EAAE,gBAAgB,GAC9B,OAAO,CAAC,eAAe,EAAE,CAAC,CAiC5B","sourcesContent":["/**\n * Retrieval evaluation gate (docs/hybrid-retrieval-design.md, step 6 of the\n * shipping order).\n *\n * Measures Recall@K and MRR for lexical, semantic, and hybrid retrieval across\n * an RRF `k` sweep, against a gold set keyed by **path + line range matched by\n * span overlap** — never by chunkId, which is only stable per index build.\n * Recall@50 doubles as the reranker gate: a gold span that never reaches the\n * fused top-50 cannot be rescued by any reranker.\n *\n * Rank-sensitive metrics are the point of Recall@1 and MRR: an agent reads the\n * top result or two, so \"the gold span is somewhere in the top 5\" understates\n * how much ordering matters. Recall@5/10/50 are kept for continuity with the\n * numbers already published in the design note.\n *\n * Driven by `scripts/search-eval.ts` via `eval-harness.ts`; the scoring logic\n * lives here so it is unit-testable without an embedding index.\n */\n\nimport type { EmbsearchService } from \"../embsearch/embsearch-service.js\";\nimport { retrieveCandidates } from \"./hybrid-search.js\";\nimport type { CandidateSpan, ResolvedSearchMode, SearchMode } from \"./types.js\";\n\nexport interface EvalGoldSpan {\n\t/** Repo-relative POSIX path. */\n\tpath: string;\n\t/** 1-based inclusive; omit both to accept any span in the file. */\n\tstartLine?: number;\n\tendLine?: number;\n\t/**\n\t * Literal source text that must occur inside `[startLine, endLine]`. Not\n\t * used for scoring — it is how the gold set survives the corpus moving\n\t * underneath it: `scripts/search-eval-gold.ts` re-resolves the line range\n\t * from this anchor, and the fixture test fails when an anchor no longer\n\t * sits inside its recorded range. Omitted for `scope: \"file\"` entries.\n\t */\n\tanchor?: string;\n\t/**\n\t * `\"span\"` (default) scores by overlap with the recorded line range;\n\t * `\"file\"` accepts any span in the file. File scope is deliberate for\n\t * path-class queries, where the whole file *is* the answer — it is\n\t * recorded explicitly so a file-level score is never mistaken for a\n\t * span-level one.\n\t */\n\tscope?: \"span\" | \"file\";\n}\n\nexport interface EvalQuery {\n\tid: string;\n\t/** Query class from the design doc (exact-symbol, path, error-fragment,\n\t *  conceptual, cross-file, boundary). Reporting only. */\n\tclass: string;\n\tquery: string;\n\tgold: EvalGoldSpan[];\n}\n\nexport interface EvalConfig {\n\tlabel: string;\n\tmode: SearchMode;\n\trrfK?: number;\n\trerank?: boolean;\n\t/**\n\t * Score this config against a daemon-side hybrid store (BM25 fused with\n\t * vectors inside the Rust daemon) instead of the dense-only index. Skipped\n\t * when the harness has no hybrid service, so records without one simply\n\t * omit the row rather than silently scoring it as plain semantic.\n\t */\n\tdaemonHybrid?: boolean;\n\t/**\n\t * Fetch the daemon's BM25 index as a separate leg and fuse it here with\n\t * dense and grep. Like `daemonHybrid` it needs a hybrid store, so it is\n\t * skipped when the harness has no hybrid service.\n\t */\n\tbm25Leg?: boolean;\n\t/** Rerank with the daemon's cross-encoder rather than the deterministic\n\t *  scorer. Needs a service, so it is skipped without one. */\n\tcrossEncoder?: boolean;\n}\n\n/**\n * The sweep from the design doc — single retrievers, hybrid across k, the\n * routed auto mode — plus reranked (`+rr`) variants for the step 7 gate.\n *\n * Exported because `scripts/search-eval.ts` and `test/search-eval.test.ts`\n * both consume it. It previously lost its `export` to a knip-driven\n * dead-export sweep (02efaab): the only importer was a `.mjs` script reading\n * from `dist/`, which static analysis cannot see, and the eval gate silently\n * stopped running. The TypeScript importers are the fix — do not \"clean up\"\n * this export without checking them.\n */\nexport const EVAL_CONFIGS: readonly EvalConfig[] = [\n\t{ label: \"lexical\", mode: \"lexical\" },\n\t{ label: \"semantic\", mode: \"semantic\" },\n\t{ label: \"hybrid k=0\", mode: \"hybrid\", rrfK: 0 },\n\t{ label: \"hybrid k=2\", mode: \"hybrid\", rrfK: 2 },\n\t{ label: \"hybrid k=10\", mode: \"hybrid\", rrfK: 10 },\n\t{ label: \"hybrid k=60\", mode: \"hybrid\", rrfK: 60 },\n\t{ label: \"auto\", mode: \"auto\" },\n\t{ label: \"lexical +rr\", mode: \"lexical\", rerank: true },\n\t{ label: \"semantic +rr\", mode: \"semantic\", rerank: true },\n\t{ label: \"hybrid k=2 +rr\", mode: \"hybrid\", rrfK: 2, rerank: true },\n\t{ label: \"hybrid k=60 +rr\", mode: \"hybrid\", rrfK: 60, rerank: true },\n\t{ label: \"auto +rr\", mode: \"auto\", rerank: true },\n\t// Is BM25 a better lexical leg than ripgrep? These two run the daemon's own\n\t// vector+BM25 fusion and no ripgrep at all, so comparing them against\n\t// \"semantic\" isolates what BM25 adds, and against \"hybrid k=2\" compares the\n\t// two lexical legs at the system level.\n\t{ label: \"daemon-hybrid\", mode: \"semantic\", daemonHybrid: true },\n\t{ label: \"daemon-hybrid +rr\", mode: \"semantic\", rerank: true, daemonHybrid: true },\n\t// Three-way fusion: dense + BM25 + grep, all fused here so `k` is ours and\n\t// the per-leg ranks survive into the trace. `bm25+dense` isolates what the\n\t// grep leg still contributes once BM25 is present.\n\t{ label: \"bm25+dense +rr\", mode: \"semantic\", rerank: true, bm25Leg: true },\n\t{ label: \"3-way +rr\", mode: \"hybrid\", rerank: true, bm25Leg: true },\n\t// Cross-encoder reranking, against the same retrieval each deterministic\n\t// `+rr` row uses — so the delta is the reranker and nothing else.\n\t{ label: \"semantic +ce\", mode: \"semantic\", crossEncoder: true },\n\t{ label: \"auto +ce\", mode: \"auto\", crossEncoder: true },\n\t{ label: \"bm25+dense +ce\", mode: \"semantic\", bm25Leg: true, crossEncoder: true },\n];\n\n/** Candidates fetched per eval query — deep enough for the reranker gate. */\nconst EVAL_FETCH_LIMIT = 50;\n\nexport function spanMatchesGold(span: CandidateSpan, gold: EvalGoldSpan): boolean {\n\tif (span.path !== gold.path) return false;\n\tif (gold.scope === \"file\") return true;\n\tif (gold.startLine === undefined || gold.endLine === undefined) return true;\n\treturn span.startLine <= gold.endLine && span.endLine >= gold.startLine;\n}\n\n/** Fraction of gold spans matched by at least one of the top-`k` candidates. */\nexport function recallAtK(candidates: readonly CandidateSpan[], gold: readonly EvalGoldSpan[], k: number): number {\n\tif (gold.length === 0) return 0;\n\tconst top = candidates.slice(0, k);\n\tlet matched = 0;\n\tfor (const g of gold) {\n\t\tif (top.some((c) => spanMatchesGold(c, g))) matched++;\n\t}\n\treturn matched / gold.length;\n}\n\n/**\n * Mean reciprocal rank, averaged over gold spans: for each gold span, `1/rank`\n * of the first candidate matching it (0 when it never appears).\n *\n * Averaging per gold span rather than taking the single first hit keeps\n * multi-span queries (the cross-file class) honest — finding one of two\n * required sites should not score like finding both. For single-span queries\n * this reduces to textbook MRR.\n */\nexport function mrr(candidates: readonly CandidateSpan[], gold: readonly EvalGoldSpan[]): number {\n\tif (gold.length === 0) return 0;\n\tlet total = 0;\n\tfor (const g of gold) {\n\t\tconst index = candidates.findIndex((c) => spanMatchesGold(c, g));\n\t\tif (index >= 0) total += 1 / (index + 1);\n\t}\n\treturn total / gold.length;\n}\n\nexport interface EvalQueryResult {\n\tlabel: string;\n\tresolvedMode: ResolvedSearchMode;\n\tdegraded: boolean;\n\trecallAt1: number;\n\trecallAt5: number;\n\trecallAt10: number;\n\trecallAt50: number;\n\tmrr: number;\n}\n\nexport async function evaluateQuery(\n\tcwd: string,\n\tevalQuery: EvalQuery,\n\tconfigs: readonly EvalConfig[] = EVAL_CONFIGS,\n\tservice?: EmbsearchService,\n\thybridService?: EmbsearchService,\n): Promise<EvalQueryResult[]> {\n\tconst results: EvalQueryResult[] = [];\n\tfor (const config of configs) {\n\t\t// A daemon-hybrid config against the plain store would error in the\n\t\t// daemon (`query_hybrid requires a hybrid store`); omit the row instead.\n\t\tif ((config.daemonHybrid || config.bm25Leg) && !hybridService) continue;\n\t\t// Cross-encoder rows need a live daemon; omit rather than silently\n\t\t// falling back to the deterministic reranker under a \"+ce\" label.\n\t\tif (config.crossEncoder && !(config.bm25Leg ? hybridService : service)?.supportsCrossEncoder()) continue;\n\t\tconst retrieved = await retrieveCandidates({\n\t\t\tcwd,\n\t\t\tquery: evalQuery.query,\n\t\t\tmode: config.mode,\n\t\t\trrfK: config.rrfK,\n\t\t\trerank: config.rerank ?? false,\n\t\t\tlimit: EVAL_FETCH_LIMIT,\n\t\t\tservice: config.daemonHybrid || config.bm25Leg ? hybridService : service,\n\t\t\tdaemonHybrid: config.daemonHybrid,\n\t\t\tbm25Leg: config.bm25Leg,\n\t\t\tcrossEncoder: config.crossEncoder,\n\t\t});\n\t\tresults.push({\n\t\t\tlabel: config.label,\n\t\t\tresolvedMode: retrieved.resolvedMode,\n\t\t\tdegraded: retrieved.degradedReason !== undefined,\n\t\t\trecallAt1: recallAtK(retrieved.candidates, evalQuery.gold, 1),\n\t\t\trecallAt5: recallAtK(retrieved.candidates, evalQuery.gold, 5),\n\t\t\trecallAt10: recallAtK(retrieved.candidates, evalQuery.gold, 10),\n\t\t\trecallAt50: recallAtK(retrieved.candidates, evalQuery.gold, 50),\n\t\t\tmrr: mrr(retrieved.candidates, evalQuery.gold),\n\t\t});\n\t}\n\treturn results;\n}\n"]}

package/dist/core/search/eval.js

@@ -2,19 +2,33 @@
  * Retrieval evaluation gate (docs/hybrid-retrieval-design.md, step 6 of the
  * shipping order).
  *
- * Measures Recall@K for lexical, semantic, and hybrid retrieval across an
- * RRF `k` sweep, against a gold set keyed by **path + optional line range
- * matched by span overlap** — never by chunkId, which is only stable per
- * index build. Recall@50 doubles as the reranker gate: a gold span that
- * never reaches the fused top-50 cannot be rescued by any reranker.
+ * Measures Recall@K and MRR for lexical, semantic, and hybrid retrieval across
+ * an RRF `k` sweep, against a gold set keyed by **path + line range matched by
+ * span overlap** — never by chunkId, which is only stable per index build.
+ * Recall@50 doubles as the reranker gate: a gold span that never reaches the
+ * fused top-50 cannot be rescued by any reranker.
  *
- * Driven by `scripts/search-eval.mjs`; the scoring logic lives here so it is
- * unit-testable without an embedding index.
+ * Rank-sensitive metrics are the point of Recall@1 and MRR: an agent reads the
+ * top result or two, so "the gold span is somewhere in the top 5" understates
+ * how much ordering matters. Recall@5/10/50 are kept for continuity with the
+ * numbers already published in the design note.
+ *
+ * Driven by `scripts/search-eval.ts` via `eval-harness.ts`; the scoring logic
+ * lives here so it is unit-testable without an embedding index.
  */
 import { retrieveCandidates } from "./hybrid-search.js";
-/** The sweep from the design doc — single retrievers, hybrid across k, the
- *  routed auto mode — plus reranked (`+rr`) variants for the step 7 gate. */
-const EVAL_CONFIGS = [
+/**
+ * The sweep from the design doc — single retrievers, hybrid across k, the
+ * routed auto mode — plus reranked (`+rr`) variants for the step 7 gate.
+ *
+ * Exported because `scripts/search-eval.ts` and `test/search-eval.test.ts`
+ * both consume it. It previously lost its `export` to a knip-driven
+ * dead-export sweep (02efaab): the only importer was a `.mjs` script reading
+ * from `dist/`, which static analysis cannot see, and the eval gate silently
+ * stopped running. The TypeScript importers are the fix — do not "clean up"
+ * this export without checking them.
+ */
+export const EVAL_CONFIGS = [
     { label: "lexical", mode: "lexical" },
     { label: "semantic", mode: "semantic" },
     { label: "hybrid k=0", mode: "hybrid", rrfK: 0 },
@@ -27,12 +41,30 @@ const EVAL_CONFIGS = [
     { label: "hybrid k=2 +rr", mode: "hybrid", rrfK: 2, rerank: true },
     { label: "hybrid k=60 +rr", mode: "hybrid", rrfK: 60, rerank: true },
     { label: "auto +rr", mode: "auto", rerank: true },
+    // Is BM25 a better lexical leg than ripgrep? These two run the daemon's own
+    // vector+BM25 fusion and no ripgrep at all, so comparing them against
+    // "semantic" isolates what BM25 adds, and against "hybrid k=2" compares the
+    // two lexical legs at the system level.
+    { label: "daemon-hybrid", mode: "semantic", daemonHybrid: true },
+    { label: "daemon-hybrid +rr", mode: "semantic", rerank: true, daemonHybrid: true },
+    // Three-way fusion: dense + BM25 + grep, all fused here so `k` is ours and
+    // the per-leg ranks survive into the trace. `bm25+dense` isolates what the
+    // grep leg still contributes once BM25 is present.
+    { label: "bm25+dense +rr", mode: "semantic", rerank: true, bm25Leg: true },
+    { label: "3-way +rr", mode: "hybrid", rerank: true, bm25Leg: true },
+    // Cross-encoder reranking, against the same retrieval each deterministic
+    // `+rr` row uses — so the delta is the reranker and nothing else.
+    { label: "semantic +ce", mode: "semantic", crossEncoder: true },
+    { label: "auto +ce", mode: "auto", crossEncoder: true },
+    { label: "bm25+dense +ce", mode: "semantic", bm25Leg: true, crossEncoder: true },
 ];
 /** Candidates fetched per eval query — deep enough for the reranker gate. */
 const EVAL_FETCH_LIMIT = 50;
 export function spanMatchesGold(span, gold) {
     if (span.path !== gold.path)
         return false;
+    if (gold.scope === "file")
+        return true;
     if (gold.startLine === undefined || gold.endLine === undefined)
         return true;
     return span.startLine <= gold.endLine && span.endLine >= gold.startLine;
@@ -49,9 +81,37 @@ export function recallAtK(candidates, gold, k) {
     }
     return matched / gold.length;
 }
-export async function evaluateQuery(cwd, evalQuery, configs = EVAL_CONFIGS, service) {
+/**
+ * Mean reciprocal rank, averaged over gold spans: for each gold span, `1/rank`
+ * of the first candidate matching it (0 when it never appears).
+ *
+ * Averaging per gold span rather than taking the single first hit keeps
+ * multi-span queries (the cross-file class) honest — finding one of two
+ * required sites should not score like finding both. For single-span queries
+ * this reduces to textbook MRR.
+ */
+export function mrr(candidates, gold) {
+    if (gold.length === 0)
+        return 0;
+    let total = 0;
+    for (const g of gold) {
+        const index = candidates.findIndex((c) => spanMatchesGold(c, g));
+        if (index >= 0)
+            total += 1 / (index + 1);
+    }
+    return total / gold.length;
+}
+export async function evaluateQuery(cwd, evalQuery, configs = EVAL_CONFIGS, service, hybridService) {
     const results = [];
     for (const config of configs) {
+        // A daemon-hybrid config against the plain store would error in the
+        // daemon (`query_hybrid requires a hybrid store`); omit the row instead.
+        if ((config.daemonHybrid || config.bm25Leg) && !hybridService)
+            continue;
+        // Cross-encoder rows need a live daemon; omit rather than silently
+        // falling back to the deterministic reranker under a "+ce" label.
+        if (config.crossEncoder && !(config.bm25Leg ? hybridService : service)?.supportsCrossEncoder())
+            continue;
         const retrieved = await retrieveCandidates({
             cwd,
             query: evalQuery.query,
@@ -59,15 +119,20 @@ export async function evaluateQuery(cwd, evalQuery, configs = EVAL_CONFIGS, serv
             rrfK: config.rrfK,
             rerank: config.rerank ?? false,
             limit: EVAL_FETCH_LIMIT,
-            service,
+            service: config.daemonHybrid || config.bm25Leg ? hybridService : service,
+            daemonHybrid: config.daemonHybrid,
+            bm25Leg: config.bm25Leg,
+            crossEncoder: config.crossEncoder,
         });
         results.push({
             label: config.label,
             resolvedMode: retrieved.resolvedMode,
             degraded: retrieved.degradedReason !== undefined,
+            recallAt1: recallAtK(retrieved.candidates, evalQuery.gold, 1),
             recallAt5: recallAtK(retrieved.candidates, evalQuery.gold, 5),
             recallAt10: recallAtK(retrieved.candidates, evalQuery.gold, 10),
             recallAt50: recallAtK(retrieved.candidates, evalQuery.gold, 50),
+            mrr: mrr(retrieved.candidates, evalQuery.gold),
         });
     }
     return results;

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

@@ -1 +1 @@
-{"version":3,"file":"eval.js","sourceRoot":"","sources":["../../../src/core/search/eval.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;GAYG;AAGH,OAAO,EAAE,kBAAkB,EAAE,MAAM,oBAAoB,CAAC;AA2BxD;+EAC6E;AAC7E,MAAM,YAAY,GAA0B;IAC3C,EAAE,KAAK,EAAE,SAAS,EAAE,IAAI,EAAE,SAAS,EAAE;IACrC,EAAE,KAAK,EAAE,UAAU,EAAE,IAAI,EAAE,UAAU,EAAE;IACvC,EAAE,KAAK,EAAE,YAAY,EAAE,IAAI,EAAE,QAAQ,EAAE,IAAI,EAAE,CAAC,EAAE;IAChD,EAAE,KAAK,EAAE,YAAY,EAAE,IAAI,EAAE,QAAQ,EAAE,IAAI,EAAE,CAAC,EAAE;IAChD,EAAE,KAAK,EAAE,aAAa,EAAE,IAAI,EAAE,QAAQ,EAAE,IAAI,EAAE,EAAE,EAAE;IAClD,EAAE,KAAK,EAAE,aAAa,EAAE,IAAI,EAAE,QAAQ,EAAE,IAAI,EAAE,EAAE,EAAE;IAClD,EAAE,KAAK,EAAE,MAAM,EAAE,IAAI,EAAE,MAAM,EAAE;IAC/B,EAAE,KAAK,EAAE,aAAa,EAAE,IAAI,EAAE,SAAS,EAAE,MAAM,EAAE,IAAI,EAAE;IACvD,EAAE,KAAK,EAAE,cAAc,EAAE,IAAI,EAAE,UAAU,EAAE,MAAM,EAAE,IAAI,EAAE;IACzD,EAAE,KAAK,EAAE,gBAAgB,EAAE,IAAI,EAAE,QAAQ,EAAE,IAAI,EAAE,CAAC,EAAE,MAAM,EAAE,IAAI,EAAE;IAClE,EAAE,KAAK,EAAE,iBAAiB,EAAE,IAAI,EAAE,QAAQ,EAAE,IAAI,EAAE,EAAE,EAAE,MAAM,EAAE,IAAI,EAAE;IACpE,EAAE,KAAK,EAAE,UAAU,EAAE,IAAI,EAAE,MAAM,EAAE,MAAM,EAAE,IAAI,EAAE;CACjD,CAAC;AAEF,+EAA6E;AAC7E,MAAM,gBAAgB,GAAG,EAAE,CAAC;AAE5B,MAAM,UAAU,eAAe,CAAC,IAAmB,EAAE,IAAkB,EAAW;IACjF,IAAI,IAAI,CAAC,IAAI,KAAK,IAAI,CAAC,IAAI;QAAE,OAAO,KAAK,CAAC;IAC1C,IAAI,IAAI,CAAC,SAAS,KAAK,SAAS,IAAI,IAAI,CAAC,OAAO,KAAK,SAAS;QAAE,OAAO,IAAI,CAAC;IAC5E,OAAO,IAAI,CAAC,SAAS,IAAI,IAAI,CAAC,OAAO,IAAI,IAAI,CAAC,OAAO,IAAI,IAAI,CAAC,SAAS,CAAC;AAAA,CACxE;AAED,gFAAgF;AAChF,MAAM,UAAU,SAAS,CAAC,UAAoC,EAAE,IAA6B,EAAE,CAAS,EAAU;IACjH,IAAI,IAAI,CAAC,MAAM,KAAK,CAAC;QAAE,OAAO,CAAC,CAAC;IAChC,MAAM,GAAG,GAAG,UAAU,CAAC,KAAK,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC;IACnC,IAAI,OAAO,GAAG,CAAC,CAAC;IAChB,KAAK,MAAM,CAAC,IAAI,IAAI,EAAE,CAAC;QACtB,IAAI,GAAG,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,eAAe,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC;YAAE,OAAO,EAAE,CAAC;IACvD,CAAC;IACD,OAAO,OAAO,GAAG,IAAI,CAAC,MAAM,CAAC;AAAA,CAC7B;AAWD,MAAM,CAAC,KAAK,UAAU,aAAa,CAClC,GAAW,EACX,SAAoB,EACpB,OAAO,GAA0B,YAAY,EAC7C,OAA0B,EACG;IAC7B,MAAM,OAAO,GAAsB,EAAE,CAAC;IACtC,KAAK,MAAM,MAAM,IAAI,OAAO,EAAE,CAAC;QAC9B,MAAM,SAAS,GAAG,MAAM,kBAAkB,CAAC;YAC1C,GAAG;YACH,KAAK,EAAE,SAAS,CAAC,KAAK;YACtB,IAAI,EAAE,MAAM,CAAC,IAAI;YACjB,IAAI,EAAE,MAAM,CAAC,IAAI;YACjB,MAAM,EAAE,MAAM,CAAC,MAAM,IAAI,KAAK;YAC9B,KAAK,EAAE,gBAAgB;YACvB,OAAO;SACP,CAAC,CAAC;QACH,OAAO,CAAC,IAAI,CAAC;YACZ,KAAK,EAAE,MAAM,CAAC,KAAK;YACnB,YAAY,EAAE,SAAS,CAAC,YAAY;YACpC,QAAQ,EAAE,SAAS,CAAC,cAAc,KAAK,SAAS;YAChD,SAAS,EAAE,SAAS,CAAC,SAAS,CAAC,UAAU,EAAE,SAAS,CAAC,IAAI,EAAE,CAAC,CAAC;YAC7D,UAAU,EAAE,SAAS,CAAC,SAAS,CAAC,UAAU,EAAE,SAAS,CAAC,IAAI,EAAE,EAAE,CAAC;YAC/D,UAAU,EAAE,SAAS,CAAC,SAAS,CAAC,UAAU,EAAE,SAAS,CAAC,IAAI,EAAE,EAAE,CAAC;SAC/D,CAAC,CAAC;IACJ,CAAC;IACD,OAAO,OAAO,CAAC;AAAA,CACf","sourcesContent":["/**\n * Retrieval evaluation gate (docs/hybrid-retrieval-design.md, step 6 of the\n * shipping order).\n *\n * Measures Recall@K for lexical, semantic, and hybrid retrieval across an\n * RRF `k` sweep, against a gold set keyed by **path + optional line range\n * matched by span overlap** — never by chunkId, which is only stable per\n * index build. Recall@50 doubles as the reranker gate: a gold span that\n * never reaches the fused top-50 cannot be rescued by any reranker.\n *\n * Driven by `scripts/search-eval.mjs`; the scoring logic lives here so it is\n * unit-testable without an embedding index.\n */\n\nimport type { EmbsearchService } from \"../embsearch/embsearch-service.js\";\nimport { retrieveCandidates } from \"./hybrid-search.js\";\nimport type { CandidateSpan, ResolvedSearchMode, SearchMode } from \"./types.js\";\n\nexport interface EvalGoldSpan {\n\t/** Repo-relative POSIX path. */\n\tpath: string;\n\t/** 1-based inclusive; omit both to accept any span in the file. */\n\tstartLine?: number;\n\tendLine?: number;\n}\n\nexport interface EvalQuery {\n\tid: string;\n\t/** Query class from the design doc (exact-symbol, path, error-fragment,\n\t *  conceptual, cross-file, boundary). Reporting only. */\n\tclass: string;\n\tquery: string;\n\tgold: EvalGoldSpan[];\n}\n\nexport interface EvalConfig {\n\tlabel: string;\n\tmode: SearchMode;\n\trrfK?: number;\n\trerank?: boolean;\n}\n\n/** The sweep from the design doc — single retrievers, hybrid across k, the\n *  routed auto mode — plus reranked (`+rr`) variants for the step 7 gate. */\nconst EVAL_CONFIGS: readonly EvalConfig[] = [\n\t{ label: \"lexical\", mode: \"lexical\" },\n\t{ label: \"semantic\", mode: \"semantic\" },\n\t{ label: \"hybrid k=0\", mode: \"hybrid\", rrfK: 0 },\n\t{ label: \"hybrid k=2\", mode: \"hybrid\", rrfK: 2 },\n\t{ label: \"hybrid k=10\", mode: \"hybrid\", rrfK: 10 },\n\t{ label: \"hybrid k=60\", mode: \"hybrid\", rrfK: 60 },\n\t{ label: \"auto\", mode: \"auto\" },\n\t{ label: \"lexical +rr\", mode: \"lexical\", rerank: true },\n\t{ label: \"semantic +rr\", mode: \"semantic\", rerank: true },\n\t{ label: \"hybrid k=2 +rr\", mode: \"hybrid\", rrfK: 2, rerank: true },\n\t{ label: \"hybrid k=60 +rr\", mode: \"hybrid\", rrfK: 60, rerank: true },\n\t{ label: \"auto +rr\", mode: \"auto\", rerank: true },\n];\n\n/** Candidates fetched per eval query — deep enough for the reranker gate. */\nconst EVAL_FETCH_LIMIT = 50;\n\nexport function spanMatchesGold(span: CandidateSpan, gold: EvalGoldSpan): boolean {\n\tif (span.path !== gold.path) return false;\n\tif (gold.startLine === undefined || gold.endLine === undefined) return true;\n\treturn span.startLine <= gold.endLine && span.endLine >= gold.startLine;\n}\n\n/** Fraction of gold spans matched by at least one of the top-`k` candidates. */\nexport function recallAtK(candidates: readonly CandidateSpan[], gold: readonly EvalGoldSpan[], k: number): number {\n\tif (gold.length === 0) return 0;\n\tconst top = candidates.slice(0, k);\n\tlet matched = 0;\n\tfor (const g of gold) {\n\t\tif (top.some((c) => spanMatchesGold(c, g))) matched++;\n\t}\n\treturn matched / gold.length;\n}\n\nexport interface EvalQueryResult {\n\tlabel: string;\n\tresolvedMode: ResolvedSearchMode;\n\tdegraded: boolean;\n\trecallAt5: number;\n\trecallAt10: number;\n\trecallAt50: number;\n}\n\nexport async function evaluateQuery(\n\tcwd: string,\n\tevalQuery: EvalQuery,\n\tconfigs: readonly EvalConfig[] = EVAL_CONFIGS,\n\tservice?: EmbsearchService,\n): Promise<EvalQueryResult[]> {\n\tconst results: EvalQueryResult[] = [];\n\tfor (const config of configs) {\n\t\tconst retrieved = await retrieveCandidates({\n\t\t\tcwd,\n\t\t\tquery: evalQuery.query,\n\t\t\tmode: config.mode,\n\t\t\trrfK: config.rrfK,\n\t\t\trerank: config.rerank ?? false,\n\t\t\tlimit: EVAL_FETCH_LIMIT,\n\t\t\tservice,\n\t\t});\n\t\tresults.push({\n\t\t\tlabel: config.label,\n\t\t\tresolvedMode: retrieved.resolvedMode,\n\t\t\tdegraded: retrieved.degradedReason !== undefined,\n\t\t\trecallAt5: recallAtK(retrieved.candidates, evalQuery.gold, 5),\n\t\t\trecallAt10: recallAtK(retrieved.candidates, evalQuery.gold, 10),\n\t\t\trecallAt50: recallAtK(retrieved.candidates, evalQuery.gold, 50),\n\t\t});\n\t}\n\treturn results;\n}\n"]}
+{"version":3,"file":"eval.js","sourceRoot":"","sources":["../../../src/core/search/eval.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;GAiBG;AAGH,OAAO,EAAE,kBAAkB,EAAE,MAAM,oBAAoB,CAAC;AA2DxD;;;;;;;;;;GAUG;AACH,MAAM,CAAC,MAAM,YAAY,GAA0B;IAClD,EAAE,KAAK,EAAE,SAAS,EAAE,IAAI,EAAE,SAAS,EAAE;IACrC,EAAE,KAAK,EAAE,UAAU,EAAE,IAAI,EAAE,UAAU,EAAE;IACvC,EAAE,KAAK,EAAE,YAAY,EAAE,IAAI,EAAE,QAAQ,EAAE,IAAI,EAAE,CAAC,EAAE;IAChD,EAAE,KAAK,EAAE,YAAY,EAAE,IAAI,EAAE,QAAQ,EAAE,IAAI,EAAE,CAAC,EAAE;IAChD,EAAE,KAAK,EAAE,aAAa,EAAE,IAAI,EAAE,QAAQ,EAAE,IAAI,EAAE,EAAE,EAAE;IAClD,EAAE,KAAK,EAAE,aAAa,EAAE,IAAI,EAAE,QAAQ,EAAE,IAAI,EAAE,EAAE,EAAE;IAClD,EAAE,KAAK,EAAE,MAAM,EAAE,IAAI,EAAE,MAAM,EAAE;IAC/B,EAAE,KAAK,EAAE,aAAa,EAAE,IAAI,EAAE,SAAS,EAAE,MAAM,EAAE,IAAI,EAAE;IACvD,EAAE,KAAK,EAAE,cAAc,EAAE,IAAI,EAAE,UAAU,EAAE,MAAM,EAAE,IAAI,EAAE;IACzD,EAAE,KAAK,EAAE,gBAAgB,EAAE,IAAI,EAAE,QAAQ,EAAE,IAAI,EAAE,CAAC,EAAE,MAAM,EAAE,IAAI,EAAE;IAClE,EAAE,KAAK,EAAE,iBAAiB,EAAE,IAAI,EAAE,QAAQ,EAAE,IAAI,EAAE,EAAE,EAAE,MAAM,EAAE,IAAI,EAAE;IACpE,EAAE,KAAK,EAAE,UAAU,EAAE,IAAI,EAAE,MAAM,EAAE,MAAM,EAAE,IAAI,EAAE;IACjD,4EAA4E;IAC5E,sEAAsE;IACtE,4EAA4E;IAC5E,wCAAwC;IACxC,EAAE,KAAK,EAAE,eAAe,EAAE,IAAI,EAAE,UAAU,EAAE,YAAY,EAAE,IAAI,EAAE;IAChE,EAAE,KAAK,EAAE,mBAAmB,EAAE,IAAI,EAAE,UAAU,EAAE,MAAM,EAAE,IAAI,EAAE,YAAY,EAAE,IAAI,EAAE;IAClF,2EAA2E;IAC3E,2EAA2E;IAC3E,mDAAmD;IACnD,EAAE,KAAK,EAAE,gBAAgB,EAAE,IAAI,EAAE,UAAU,EAAE,MAAM,EAAE,IAAI,EAAE,OAAO,EAAE,IAAI,EAAE;IAC1E,EAAE,KAAK,EAAE,WAAW,EAAE,IAAI,EAAE,QAAQ,EAAE,MAAM,EAAE,IAAI,EAAE,OAAO,EAAE,IAAI,EAAE;IACnE,yEAAyE;IACzE,oEAAkE;IAClE,EAAE,KAAK,EAAE,cAAc,EAAE,IAAI,EAAE,UAAU,EAAE,YAAY,EAAE,IAAI,EAAE;IAC/D,EAAE,KAAK,EAAE,UAAU,EAAE,IAAI,EAAE,MAAM,EAAE,YAAY,EAAE,IAAI,EAAE;IACvD,EAAE,KAAK,EAAE,gBAAgB,EAAE,IAAI,EAAE,UAAU,EAAE,OAAO,EAAE,IAAI,EAAE,YAAY,EAAE,IAAI,EAAE;CAChF,CAAC;AAEF,+EAA6E;AAC7E,MAAM,gBAAgB,GAAG,EAAE,CAAC;AAE5B,MAAM,UAAU,eAAe,CAAC,IAAmB,EAAE,IAAkB,EAAW;IACjF,IAAI,IAAI,CAAC,IAAI,KAAK,IAAI,CAAC,IAAI;QAAE,OAAO,KAAK,CAAC;IAC1C,IAAI,IAAI,CAAC,KAAK,KAAK,MAAM;QAAE,OAAO,IAAI,CAAC;IACvC,IAAI,IAAI,CAAC,SAAS,KAAK,SAAS,IAAI,IAAI,CAAC,OAAO,KAAK,SAAS;QAAE,OAAO,IAAI,CAAC;IAC5E,OAAO,IAAI,CAAC,SAAS,IAAI,IAAI,CAAC,OAAO,IAAI,IAAI,CAAC,OAAO,IAAI,IAAI,CAAC,SAAS,CAAC;AAAA,CACxE;AAED,gFAAgF;AAChF,MAAM,UAAU,SAAS,CAAC,UAAoC,EAAE,IAA6B,EAAE,CAAS,EAAU;IACjH,IAAI,IAAI,CAAC,MAAM,KAAK,CAAC;QAAE,OAAO,CAAC,CAAC;IAChC,MAAM,GAAG,GAAG,UAAU,CAAC,KAAK,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC;IACnC,IAAI,OAAO,GAAG,CAAC,CAAC;IAChB,KAAK,MAAM,CAAC,IAAI,IAAI,EAAE,CAAC;QACtB,IAAI,GAAG,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,eAAe,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC;YAAE,OAAO,EAAE,CAAC;IACvD,CAAC;IACD,OAAO,OAAO,GAAG,IAAI,CAAC,MAAM,CAAC;AAAA,CAC7B;AAED;;;;;;;;GAQG;AACH,MAAM,UAAU,GAAG,CAAC,UAAoC,EAAE,IAA6B,EAAU;IAChG,IAAI,IAAI,CAAC,MAAM,KAAK,CAAC;QAAE,OAAO,CAAC,CAAC;IAChC,IAAI,KAAK,GAAG,CAAC,CAAC;IACd,KAAK,MAAM,CAAC,IAAI,IAAI,EAAE,CAAC;QACtB,MAAM,KAAK,GAAG,UAAU,CAAC,SAAS,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,eAAe,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,CAAC;QACjE,IAAI,KAAK,IAAI,CAAC;YAAE,KAAK,IAAI,CAAC,GAAG,CAAC,KAAK,GAAG,CAAC,CAAC,CAAC;IAC1C,CAAC;IACD,OAAO,KAAK,GAAG,IAAI,CAAC,MAAM,CAAC;AAAA,CAC3B;AAaD,MAAM,CAAC,KAAK,UAAU,aAAa,CAClC,GAAW,EACX,SAAoB,EACpB,OAAO,GAA0B,YAAY,EAC7C,OAA0B,EAC1B,aAAgC,EACH;IAC7B,MAAM,OAAO,GAAsB,EAAE,CAAC;IACtC,KAAK,MAAM,MAAM,IAAI,OAAO,EAAE,CAAC;QAC9B,oEAAoE;QACpE,yEAAyE;QACzE,IAAI,CAAC,MAAM,CAAC,YAAY,IAAI,MAAM,CAAC,OAAO,CAAC,IAAI,CAAC,aAAa;YAAE,SAAS;QACxE,mEAAmE;QACnE,kEAAkE;QAClE,IAAI,MAAM,CAAC,YAAY,IAAI,CAAC,CAAC,MAAM,CAAC,OAAO,CAAC,CAAC,CAAC,aAAa,CAAC,CAAC,CAAC,OAAO,CAAC,EAAE,oBAAoB,EAAE;YAAE,SAAS;QACzG,MAAM,SAAS,GAAG,MAAM,kBAAkB,CAAC;YAC1C,GAAG;YACH,KAAK,EAAE,SAAS,CAAC,KAAK;YACtB,IAAI,EAAE,MAAM,CAAC,IAAI;YACjB,IAAI,EAAE,MAAM,CAAC,IAAI;YACjB,MAAM,EAAE,MAAM,CAAC,MAAM,IAAI,KAAK;YAC9B,KAAK,EAAE,gBAAgB;YACvB,OAAO,EAAE,MAAM,CAAC,YAAY,IAAI,MAAM,CAAC,OAAO,CAAC,CAAC,CAAC,aAAa,CAAC,CAAC,CAAC,OAAO;YACxE,YAAY,EAAE,MAAM,CAAC,YAAY;YACjC,OAAO,EAAE,MAAM,CAAC,OAAO;YACvB,YAAY,EAAE,MAAM,CAAC,YAAY;SACjC,CAAC,CAAC;QACH,OAAO,CAAC,IAAI,CAAC;YACZ,KAAK,EAAE,MAAM,CAAC,KAAK;YACnB,YAAY,EAAE,SAAS,CAAC,YAAY;YACpC,QAAQ,EAAE,SAAS,CAAC,cAAc,KAAK,SAAS;YAChD,SAAS,EAAE,SAAS,CAAC,SAAS,CAAC,UAAU,EAAE,SAAS,CAAC,IAAI,EAAE,CAAC,CAAC;YAC7D,SAAS,EAAE,SAAS,CAAC,SAAS,CAAC,UAAU,EAAE,SAAS,CAAC,IAAI,EAAE,CAAC,CAAC;YAC7D,UAAU,EAAE,SAAS,CAAC,SAAS,CAAC,UAAU,EAAE,SAAS,CAAC,IAAI,EAAE,EAAE,CAAC;YAC/D,UAAU,EAAE,SAAS,CAAC,SAAS,CAAC,UAAU,EAAE,SAAS,CAAC,IAAI,EAAE,EAAE,CAAC;YAC/D,GAAG,EAAE,GAAG,CAAC,SAAS,CAAC,UAAU,EAAE,SAAS,CAAC,IAAI,CAAC;SAC9C,CAAC,CAAC;IACJ,CAAC;IACD,OAAO,OAAO,CAAC;AAAA,CACf","sourcesContent":["/**\n * Retrieval evaluation gate (docs/hybrid-retrieval-design.md, step 6 of the\n * shipping order).\n *\n * Measures Recall@K and MRR for lexical, semantic, and hybrid retrieval across\n * an RRF `k` sweep, against a gold set keyed by **path + line range matched by\n * span overlap** — never by chunkId, which is only stable per index build.\n * Recall@50 doubles as the reranker gate: a gold span that never reaches the\n * fused top-50 cannot be rescued by any reranker.\n *\n * Rank-sensitive metrics are the point of Recall@1 and MRR: an agent reads the\n * top result or two, so \"the gold span is somewhere in the top 5\" understates\n * how much ordering matters. Recall@5/10/50 are kept for continuity with the\n * numbers already published in the design note.\n *\n * Driven by `scripts/search-eval.ts` via `eval-harness.ts`; the scoring logic\n * lives here so it is unit-testable without an embedding index.\n */\n\nimport type { EmbsearchService } from \"../embsearch/embsearch-service.js\";\nimport { retrieveCandidates } from \"./hybrid-search.js\";\nimport type { CandidateSpan, ResolvedSearchMode, SearchMode } from \"./types.js\";\n\nexport interface EvalGoldSpan {\n\t/** Repo-relative POSIX path. */\n\tpath: string;\n\t/** 1-based inclusive; omit both to accept any span in the file. */\n\tstartLine?: number;\n\tendLine?: number;\n\t/**\n\t * Literal source text that must occur inside `[startLine, endLine]`. Not\n\t * used for scoring — it is how the gold set survives the corpus moving\n\t * underneath it: `scripts/search-eval-gold.ts` re-resolves the line range\n\t * from this anchor, and the fixture test fails when an anchor no longer\n\t * sits inside its recorded range. Omitted for `scope: \"file\"` entries.\n\t */\n\tanchor?: string;\n\t/**\n\t * `\"span\"` (default) scores by overlap with the recorded line range;\n\t * `\"file\"` accepts any span in the file. File scope is deliberate for\n\t * path-class queries, where the whole file *is* the answer — it is\n\t * recorded explicitly so a file-level score is never mistaken for a\n\t * span-level one.\n\t */\n\tscope?: \"span\" | \"file\";\n}\n\nexport interface EvalQuery {\n\tid: string;\n\t/** Query class from the design doc (exact-symbol, path, error-fragment,\n\t *  conceptual, cross-file, boundary). Reporting only. */\n\tclass: string;\n\tquery: string;\n\tgold: EvalGoldSpan[];\n}\n\nexport interface EvalConfig {\n\tlabel: string;\n\tmode: SearchMode;\n\trrfK?: number;\n\trerank?: boolean;\n\t/**\n\t * Score this config against a daemon-side hybrid store (BM25 fused with\n\t * vectors inside the Rust daemon) instead of the dense-only index. Skipped\n\t * when the harness has no hybrid service, so records without one simply\n\t * omit the row rather than silently scoring it as plain semantic.\n\t */\n\tdaemonHybrid?: boolean;\n\t/**\n\t * Fetch the daemon's BM25 index as a separate leg and fuse it here with\n\t * dense and grep. Like `daemonHybrid` it needs a hybrid store, so it is\n\t * skipped when the harness has no hybrid service.\n\t */\n\tbm25Leg?: boolean;\n\t/** Rerank with the daemon's cross-encoder rather than the deterministic\n\t *  scorer. Needs a service, so it is skipped without one. */\n\tcrossEncoder?: boolean;\n}\n\n/**\n * The sweep from the design doc — single retrievers, hybrid across k, the\n * routed auto mode — plus reranked (`+rr`) variants for the step 7 gate.\n *\n * Exported because `scripts/search-eval.ts` and `test/search-eval.test.ts`\n * both consume it. It previously lost its `export` to a knip-driven\n * dead-export sweep (02efaab): the only importer was a `.mjs` script reading\n * from `dist/`, which static analysis cannot see, and the eval gate silently\n * stopped running. The TypeScript importers are the fix — do not \"clean up\"\n * this export without checking them.\n */\nexport const EVAL_CONFIGS: readonly EvalConfig[] = [\n\t{ label: \"lexical\", mode: \"lexical\" },\n\t{ label: \"semantic\", mode: \"semantic\" },\n\t{ label: \"hybrid k=0\", mode: \"hybrid\", rrfK: 0 },\n\t{ label: \"hybrid k=2\", mode: \"hybrid\", rrfK: 2 },\n\t{ label: \"hybrid k=10\", mode: \"hybrid\", rrfK: 10 },\n\t{ label: \"hybrid k=60\", mode: \"hybrid\", rrfK: 60 },\n\t{ label: \"auto\", mode: \"auto\" },\n\t{ label: \"lexical +rr\", mode: \"lexical\", rerank: true },\n\t{ label: \"semantic +rr\", mode: \"semantic\", rerank: true },\n\t{ label: \"hybrid k=2 +rr\", mode: \"hybrid\", rrfK: 2, rerank: true },\n\t{ label: \"hybrid k=60 +rr\", mode: \"hybrid\", rrfK: 60, rerank: true },\n\t{ label: \"auto +rr\", mode: \"auto\", rerank: true },\n\t// Is BM25 a better lexical leg than ripgrep? These two run the daemon's own\n\t// vector+BM25 fusion and no ripgrep at all, so comparing them against\n\t// \"semantic\" isolates what BM25 adds, and against \"hybrid k=2\" compares the\n\t// two lexical legs at the system level.\n\t{ label: \"daemon-hybrid\", mode: \"semantic\", daemonHybrid: true },\n\t{ label: \"daemon-hybrid +rr\", mode: \"semantic\", rerank: true, daemonHybrid: true },\n\t// Three-way fusion: dense + BM25 + grep, all fused here so `k` is ours and\n\t// the per-leg ranks survive into the trace. `bm25+dense` isolates what the\n\t// grep leg still contributes once BM25 is present.\n\t{ label: \"bm25+dense +rr\", mode: \"semantic\", rerank: true, bm25Leg: true },\n\t{ label: \"3-way +rr\", mode: \"hybrid\", rerank: true, bm25Leg: true },\n\t// Cross-encoder reranking, against the same retrieval each deterministic\n\t// `+rr` row uses — so the delta is the reranker and nothing else.\n\t{ label: \"semantic +ce\", mode: \"semantic\", crossEncoder: true },\n\t{ label: \"auto +ce\", mode: \"auto\", crossEncoder: true },\n\t{ label: \"bm25+dense +ce\", mode: \"semantic\", bm25Leg: true, crossEncoder: true },\n];\n\n/** Candidates fetched per eval query — deep enough for the reranker gate. */\nconst EVAL_FETCH_LIMIT = 50;\n\nexport function spanMatchesGold(span: CandidateSpan, gold: EvalGoldSpan): boolean {\n\tif (span.path !== gold.path) return false;\n\tif (gold.scope === \"file\") return true;\n\tif (gold.startLine === undefined || gold.endLine === undefined) return true;\n\treturn span.startLine <= gold.endLine && span.endLine >= gold.startLine;\n}\n\n/** Fraction of gold spans matched by at least one of the top-`k` candidates. */\nexport function recallAtK(candidates: readonly CandidateSpan[], gold: readonly EvalGoldSpan[], k: number): number {\n\tif (gold.length === 0) return 0;\n\tconst top = candidates.slice(0, k);\n\tlet matched = 0;\n\tfor (const g of gold) {\n\t\tif (top.some((c) => spanMatchesGold(c, g))) matched++;\n\t}\n\treturn matched / gold.length;\n}\n\n/**\n * Mean reciprocal rank, averaged over gold spans: for each gold span, `1/rank`\n * of the first candidate matching it (0 when it never appears).\n *\n * Averaging per gold span rather than taking the single first hit keeps\n * multi-span queries (the cross-file class) honest — finding one of two\n * required sites should not score like finding both. For single-span queries\n * this reduces to textbook MRR.\n */\nexport function mrr(candidates: readonly CandidateSpan[], gold: readonly EvalGoldSpan[]): number {\n\tif (gold.length === 0) return 0;\n\tlet total = 0;\n\tfor (const g of gold) {\n\t\tconst index = candidates.findIndex((c) => spanMatchesGold(c, g));\n\t\tif (index >= 0) total += 1 / (index + 1);\n\t}\n\treturn total / gold.length;\n}\n\nexport interface EvalQueryResult {\n\tlabel: string;\n\tresolvedMode: ResolvedSearchMode;\n\tdegraded: boolean;\n\trecallAt1: number;\n\trecallAt5: number;\n\trecallAt10: number;\n\trecallAt50: number;\n\tmrr: number;\n}\n\nexport async function evaluateQuery(\n\tcwd: string,\n\tevalQuery: EvalQuery,\n\tconfigs: readonly EvalConfig[] = EVAL_CONFIGS,\n\tservice?: EmbsearchService,\n\thybridService?: EmbsearchService,\n): Promise<EvalQueryResult[]> {\n\tconst results: EvalQueryResult[] = [];\n\tfor (const config of configs) {\n\t\t// A daemon-hybrid config against the plain store would error in the\n\t\t// daemon (`query_hybrid requires a hybrid store`); omit the row instead.\n\t\tif ((config.daemonHybrid || config.bm25Leg) && !hybridService) continue;\n\t\t// Cross-encoder rows need a live daemon; omit rather than silently\n\t\t// falling back to the deterministic reranker under a \"+ce\" label.\n\t\tif (config.crossEncoder && !(config.bm25Leg ? hybridService : service)?.supportsCrossEncoder()) continue;\n\t\tconst retrieved = await retrieveCandidates({\n\t\t\tcwd,\n\t\t\tquery: evalQuery.query,\n\t\t\tmode: config.mode,\n\t\t\trrfK: config.rrfK,\n\t\t\trerank: config.rerank ?? false,\n\t\t\tlimit: EVAL_FETCH_LIMIT,\n\t\t\tservice: config.daemonHybrid || config.bm25Leg ? hybridService : service,\n\t\t\tdaemonHybrid: config.daemonHybrid,\n\t\t\tbm25Leg: config.bm25Leg,\n\t\t\tcrossEncoder: config.crossEncoder,\n\t\t});\n\t\tresults.push({\n\t\t\tlabel: config.label,\n\t\t\tresolvedMode: retrieved.resolvedMode,\n\t\t\tdegraded: retrieved.degradedReason !== undefined,\n\t\t\trecallAt1: recallAtK(retrieved.candidates, evalQuery.gold, 1),\n\t\t\trecallAt5: recallAtK(retrieved.candidates, evalQuery.gold, 5),\n\t\t\trecallAt10: recallAtK(retrieved.candidates, evalQuery.gold, 10),\n\t\t\trecallAt50: recallAtK(retrieved.candidates, evalQuery.gold, 50),\n\t\t\tmrr: mrr(retrieved.candidates, evalQuery.gold),\n\t\t});\n\t}\n\treturn results;\n}\n"]}

package/dist/core/search/hybrid-search.d.ts

@@ -26,6 +26,28 @@ export interface RetrieveOptions {
     rrfK?: number;
     /** Rerank the fused top-50 before slicing to `limit`. Default: true. */
     rerank?: boolean;
+    /**
+     * Ask the daemon to fuse its own BM25 index with the vectors and return one
+     * already-fused ranking, instead of taking a dense-only list. Needs a store
+     * built with `--hybrid`. The fused list arrives as a single "embed" leg,
+     * because a pre-fused ranking has no per-retriever structure left to record.
+     *
+     * Prefer {@link bm25Leg}: fusing here keeps the legs separable in the trace
+     * and lets the grep leg participate.
+     */
+    daemonHybrid?: boolean;
+    /**
+     * Fetch the daemon's BM25 index as its own ranked list and fuse it here,
+     * alongside dense and grep. Needs a store built with `--hybrid` and an
+     * embsearch new enough to serve `retriever: "lexical"`.
+     */
+    bm25Leg?: boolean;
+    /**
+     * Reorder the fused shortlist with the daemon's cross-encoder instead of
+     * the deterministic reranker. Needs embsearch >= 0.3.0; costs one model
+     * pass per scored candidate.
+     */
+    crossEncoder?: boolean;
     service?: EmbsearchService;
     signal?: AbortSignal;
 }

package/dist/core/search/hybrid-search.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"hybrid-search.d.ts","sourceRoot":"","sources":["../../../src/core/search/hybrid-search.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;GAaG;AAEH,OAAO,KAAK,EAAE,gBAAgB,EAAE,MAAM,mCAAmC,CAAC;AAQ1E,OAAO,KAAK,EAAiB,cAAc,EAAa,kBAAkB,EAAE,UAAU,EAAE,WAAW,EAAE,MAAM,YAAY,CAAC;AAcxH,MAAM,WAAW,eAAe;IAC/B,GAAG,EAAE,MAAM,CAAC;IACZ,KAAK,EAAE,MAAM,CAAC;IACd,IAAI,CAAC,EAAE,UAAU,CAAC;IAClB,kDAAkD;IAClD,IAAI,CAAC,EAAE,MAAM,CAAC;IACd,yCAAyC;IACzC,KAAK,CAAC,EAAE,MAAM,CAAC;IACf,wFAAwF;IACxF,IAAI,CAAC,EAAE,MAAM,CAAC;IACd,wEAAwE;IACxE,MAAM,CAAC,EAAE,OAAO,CAAC;IACjB,OAAO,CAAC,EAAE,gBAAgB,CAAC;IAC3B,MAAM,CAAC,EAAE,WAAW,CAAC;CACrB;AAED,MAAM,WAAW,cAAc;IAC9B,UAAU,EAAE,cAAc,EAAE,CAAC;IAC7B,YAAY,EAAE,kBAAkB,CAAC;IACjC,cAAc,CAAC,EAAE,MAAM,CAAC;IACxB,UAAU,EAAE,WAAW,CAAC,YAAY,CAAC,CAAC;IACtC,UAAU,EAAE,WAAW,CAAC,YAAY,CAAC,CAAC;IACtC,uDAAuD;IACvD,QAAQ,CAAC,EAAE;QAAE,IAAI,EAAE,MAAM,CAAC;QAAC,KAAK,EAAE,MAAM,CAAA;KAAE,CAAC;IAC3C,IAAI,EAAE,MAAM,CAAC;IACb,MAAM,CAAC,EAAE,WAAW,CAAC,QAAQ,CAAC,CAAC;CAC/B;AAED,MAAM,WAAW,gBAAiB,SAAQ,eAAe;IACxD,oDAAoD;IACpD,WAAW,CAAC,EAAE,MAAM,CAAC;CACrB;AAED,MAAM,WAAW,eAAe;IAC/B,IAAI,EAAE,MAAM,CAAC;IACb,YAAY,EAAE,kBAAkB,CAAC;IACjC,cAAc,CAAC,EAAE,MAAM,CAAC;IACxB,WAAW,EAAE,MAAM,CAAC;IACpB,uDAAuD;IACvD,QAAQ,CAAC,EAAE;QAAE,IAAI,EAAE,MAAM,CAAC;QAAC,KAAK,EAAE,MAAM,CAAA;KAAE,CAAC;CAC3C;AAYD,wBAAsB,kBAAkB,CAAC,OAAO,EAAE,eAAe,GAAG,OAAO,CAAC,cAAc,CAAC,CA2G1F;AAED,wBAAsB,SAAS,CAAC,OAAO,EAAE,gBAAgB,GAAG,OAAO,CAAC,eAAe,CAAC,CAyBnF","sourcesContent":["/**\n * Hybrid search orchestrator: resolve mode, run retrievers in parallel, fuse\n * by rank, expand within budget, trace everything\n * (docs/hybrid-retrieval-design.md).\n *\n * Single-retriever modes flow through the same pipeline — rrfFuse over one\n * list preserves its order — so lexical, semantic, and hybrid all produce the\n * same result shape and the same trace record.\n *\n * `retrieveCandidates` is the candidate-level core (also used by the eval\n * harness, which needs forced modes, a configurable `k`, and no trace\n * pollution); `runSearch` wraps it with span expansion and tracing for the\n * tool.\n */\n\nimport type { EmbsearchService } from \"../embsearch/embsearch-service.js\";\nimport { adaptGrepHits, type ChunkLookup } from \"./adapter.js\";\nimport { assembleContext } from \"./context-assembler.js\";\nimport { runLexicalRetriever } from \"./lexical-retriever.js\";\nimport { resolveSearchMode } from \"./mode.js\";\nimport { rerankCandidates } from \"./rerank.js\";\nimport { DEFAULT_RRF_K, rrfFuse } from \"./rrf.js\";\nimport { writeSearchTrace } from \"./trace.js\";\nimport type { CandidateSpan, FusedCandidate, RankedHit, ResolvedSearchMode, SearchMode, SearchTrace } from \"./types.js\";\n\n/** Raw grep line-hits fetched per query (pre-collapse). */\nconst LEXICAL_MATCH_LIMIT = 200;\n/** Adapted lexical candidates entering fusion. The eval gate showed the\n *  uncapped lexical tail diluting hybrid below plain semantic: lexical\n *  precision is front-loaded by the adapter's term-evidence ranking, while\n *  RRF weighs a rank-30 lexical candidate like a rank-30 embedding hit. */\nconst LEXICAL_FUSION_CAP = 20;\n/** Embedding hits fetched per query — deep enough for fusion to matter. */\nconst EMBED_TOP_K = 50;\n/** Fused candidates kept for reranking / final slicing. */\nconst FUSED_WINDOW = 50;\n\nexport interface RetrieveOptions {\n\tcwd: string;\n\tquery: string;\n\tmode?: SearchMode;\n\t/** Optional glob filter applied to file paths. */\n\tglob?: string;\n\t/** Maximum fused candidates returned. */\n\tlimit?: number;\n\t/** RRF constant override (eval harness sweeps this). Default: {@link DEFAULT_RRF_K}. */\n\trrfK?: number;\n\t/** Rerank the fused top-50 before slicing to `limit`. Default: true. */\n\trerank?: boolean;\n\tservice?: EmbsearchService;\n\tsignal?: AbortSignal;\n}\n\nexport interface RetrieveResult {\n\tcandidates: FusedCandidate[];\n\tresolvedMode: ResolvedSearchMode;\n\tdegradedReason?: string;\n\tindexPhase: SearchTrace[\"indexPhase\"];\n\tretrievers: SearchTrace[\"retrievers\"];\n\t/** Set while the embedding index is still building. */\n\tindexing?: { done: number; total: number };\n\trrfK: number;\n\trerank?: SearchTrace[\"rerank\"];\n}\n\nexport interface RunSearchOptions extends RetrieveOptions {\n\t/** Approximate token budget for the result text. */\n\ttokenBudget?: number;\n}\n\nexport interface RunSearchResult {\n\ttext: string;\n\tresolvedMode: ResolvedSearchMode;\n\tdegradedReason?: string;\n\tresultCount: number;\n\t/** Set while the embedding index is still building. */\n\tindexing?: { done: number; total: number };\n}\n\nfunction normalizeSearchGlob(glob: string | undefined): string | undefined {\n\tif (!glob) return undefined;\n\t// Match fd/rg semantics: a slash-containing glob is anchored anywhere in\n\t// the tree, so prepend \"**/\" unless it already starts with a slash or \"**/\".\n\tif (glob.includes(\"/\") && !glob.startsWith(\"/\") && !glob.startsWith(\"**/\")) {\n\t\treturn `**/${glob}`;\n\t}\n\treturn glob;\n}\n\nexport async function retrieveCandidates(options: RetrieveOptions): Promise<RetrieveResult> {\n\tconst { cwd, query, service, signal } = options;\n\tconst glob = normalizeSearchGlob(options.glob);\n\tconst requestedMode = options.mode ?? \"auto\";\n\tconst limit = Math.max(1, options.limit ?? 10);\n\tconst rrfK = options.rrfK ?? DEFAULT_RRF_K;\n\n\tconst state = service?.getState();\n\tconst embedAvailable = service?.isAvailable() ?? false;\n\tconst embedUnavailableReason =\n\t\tstate === undefined\n\t\t\t? \"semantic index is not enabled\"\n\t\t\t: state.phase === \"unavailable\" || state.phase === \"skipped\"\n\t\t\t\t? state.reason\n\t\t\t\t: state.phase === \"idle\"\n\t\t\t\t\t? \"semantic index has not started\"\n\t\t\t\t\t: undefined;\n\n\tconst resolution = resolveSearchMode(query, requestedMode, embedAvailable, embedUnavailableReason);\n\tconst mode = resolution.mode;\n\n\t// Map lexical hits onto indexed chunk ids whenever the sidecar is usable,\n\t// even in lexical-only mode, so identities line up across modes.\n\tconst lookupChunk: ChunkLookup | undefined = embedAvailable\n\t\t? (rel, line) => service!.findEnclosingChunk(rel, line)\n\t\t: undefined;\n\n\tconst spans = new Map<string, CandidateSpan>();\n\tconst lists: RankedHit[][] = [];\n\tconst retrieverStats: SearchTrace[\"retrievers\"] = {};\n\tconst errors: Error[] = [];\n\n\tconst runLexical = async (): Promise<void> => {\n\t\tconst startedMs = Date.now();\n\t\ttry {\n\t\t\tconst lineHits = await runLexicalRetriever({ cwd, query, limit: LEXICAL_MATCH_LIMIT, glob, signal });\n\t\t\tconst adapted = adaptGrepHits(lineHits, lookupChunk);\n\t\t\t// In single-retriever lexical mode the full list is the result; in\n\t\t\t// hybrid, only the front-loaded head is trustworthy enough to vote.\n\t\t\tconst hits = mode === \"hybrid\" ? adapted.hits.slice(0, LEXICAL_FUSION_CAP) : adapted.hits;\n\t\t\tfor (const [id, span] of adapted.spans) if (!spans.has(id)) spans.set(id, span);\n\t\t\tlists.push(hits);\n\t\t\tretrieverStats.grep = { latencyMs: Date.now() - startedMs, hitCount: hits.length };\n\t\t} catch (e) {\n\t\t\terrors.push(e instanceof Error ? e : new Error(String(e)));\n\t\t\tretrieverStats.grep = { latencyMs: Date.now() - startedMs, hitCount: 0 };\n\t\t}\n\t};\n\n\tconst runEmbed = async (): Promise<void> => {\n\t\tconst startedMs = Date.now();\n\t\ttry {\n\t\t\t// The flat index pads top-k with whatever exists; with the cosine\n\t\t\t// metric the store uses, score <= 0 means \"no relation at all\", so\n\t\t\t// those padding hits would cast RRF votes on pure noise.\n\t\t\tconst chunkHits = (await service!.searchChunks(query, EMBED_TOP_K, glob)).filter((hit) => hit.score > 0);\n\t\t\tconst hits: RankedHit[] = chunkHits.map((hit, i) => ({\n\t\t\t\tid: hit.id,\n\t\t\t\trank: i + 1,\n\t\t\t\tscore: hit.score,\n\t\t\t\tsource: \"embed\",\n\t\t\t}));\n\t\t\tfor (const hit of chunkHits) {\n\t\t\t\tspans.set(hit.id, { path: hit.path, startLine: hit.startLine, endLine: hit.endLine });\n\t\t\t}\n\t\t\tlists.push(hits);\n\t\t\tretrieverStats.embed = { latencyMs: Date.now() - startedMs, hitCount: hits.length };\n\t\t} catch (e) {\n\t\t\terrors.push(e instanceof Error ? e : new Error(String(e)));\n\t\t\tretrieverStats.embed = { latencyMs: Date.now() - startedMs, hitCount: 0 };\n\t\t}\n\t};\n\n\tconst runs: Promise<void>[] = [];\n\tif (mode === \"lexical\" || mode === \"hybrid\") runs.push(runLexical());\n\tif (mode === \"semantic\" || mode === \"hybrid\") runs.push(runEmbed());\n\tawait Promise.all(runs);\n\tif (signal?.aborted) throw new Error(\"Operation aborted\");\n\t// A partial failure in hybrid degrades to whichever retriever survived;\n\t// only a total loss is an error.\n\tif (lists.length === 0) throw errors[0] ?? new Error(\"search produced no retriever results\");\n\n\tconst fused = rrfFuse(lists, rrfK).slice(0, FUSED_WINDOW);\n\tlet candidates: FusedCandidate[] = [];\n\tfor (const hit of fused) {\n\t\tconst span = spans.get(hit.id);\n\t\tif (span) candidates.push({ ...hit, ...span });\n\t}\n\n\tlet rerankInfo: SearchTrace[\"rerank\"];\n\tif (options.rerank !== false) {\n\t\tconst reranked = rerankCandidates(query, candidates, cwd);\n\t\trerankInfo = { applied: true, candidateCount: candidates.length, latencyMs: reranked.latencyMs };\n\t\tcandidates = reranked.candidates;\n\t}\n\tcandidates = candidates.slice(0, limit);\n\n\treturn {\n\t\tcandidates,\n\t\tresolvedMode: mode,\n\t\tdegradedReason: resolution.degradedReason,\n\t\tindexPhase: state?.phase === \"ready\" ? \"ready\" : state?.phase === \"indexing\" ? \"indexing\" : \"unavailable\",\n\t\tretrievers: retrieverStats,\n\t\tindexing: state?.phase === \"indexing\" ? { done: state.done, total: state.total } : undefined,\n\t\trrfK,\n\t\trerank: rerankInfo,\n\t};\n}\n\nexport async function runSearch(options: RunSearchOptions): Promise<RunSearchResult> {\n\tconst retrieved = await retrieveCandidates(options);\n\n\tconst assembled = assembleContext(retrieved.candidates, { cwd: options.cwd, tokenBudget: options.tokenBudget });\n\n\twriteSearchTrace(options.cwd, {\n\t\ttimestampMs: Date.now(),\n\t\tquery: options.query,\n\t\trequestedMode: options.mode ?? \"auto\",\n\t\tresolvedMode: retrieved.resolvedMode,\n\t\tdegradedReason: retrieved.degradedReason,\n\t\tindexPhase: retrieved.indexPhase,\n\t\trrfK: retrieved.resolvedMode === \"hybrid\" ? retrieved.rrfK : undefined,\n\t\tretrievers: retrieved.retrievers,\n\t\tfused: retrieved.candidates.map(({ id, rrfScore, ranks, rawScores }) => ({ id, rrfScore, ranks, rawScores })),\n\t\trerank: retrieved.rerank,\n\t});\n\n\treturn {\n\t\ttext: assembled.text,\n\t\tresolvedMode: retrieved.resolvedMode,\n\t\tdegradedReason: retrieved.degradedReason,\n\t\tresultCount: retrieved.candidates.length,\n\t\tindexing: retrieved.indexing,\n\t};\n}\n"]}
+{"version":3,"file":"hybrid-search.d.ts","sourceRoot":"","sources":["../../../src/core/search/hybrid-search.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;GAaG;AAEH,OAAO,KAAK,EAAE,gBAAgB,EAAE,MAAM,mCAAmC,CAAC;AAS1E,OAAO,KAAK,EAAiB,cAAc,EAAa,kBAAkB,EAAE,UAAU,EAAE,WAAW,EAAE,MAAM,YAAY,CAAC;AAmCxH,MAAM,WAAW,eAAe;IAC/B,GAAG,EAAE,MAAM,CAAC;IACZ,KAAK,EAAE,MAAM,CAAC;IACd,IAAI,CAAC,EAAE,UAAU,CAAC;IAClB,kDAAkD;IAClD,IAAI,CAAC,EAAE,MAAM,CAAC;IACd,yCAAyC;IACzC,KAAK,CAAC,EAAE,MAAM,CAAC;IACf,wFAAwF;IACxF,IAAI,CAAC,EAAE,MAAM,CAAC;IACd,wEAAwE;IACxE,MAAM,CAAC,EAAE,OAAO,CAAC;IACjB;;;;;;;;OAQG;IACH,YAAY,CAAC,EAAE,OAAO,CAAC;IACvB;;;;OAIG;IACH,OAAO,CAAC,EAAE,OAAO,CAAC;IAClB;;;;OAIG;IACH,YAAY,CAAC,EAAE,OAAO,CAAC;IACvB,OAAO,CAAC,EAAE,gBAAgB,CAAC;IAC3B,MAAM,CAAC,EAAE,WAAW,CAAC;CACrB;AAED,MAAM,WAAW,cAAc;IAC9B,UAAU,EAAE,cAAc,EAAE,CAAC;IAC7B,YAAY,EAAE,kBAAkB,CAAC;IACjC,cAAc,CAAC,EAAE,MAAM,CAAC;IACxB,UAAU,EAAE,WAAW,CAAC,YAAY,CAAC,CAAC;IACtC,UAAU,EAAE,WAAW,CAAC,YAAY,CAAC,CAAC;IACtC,uDAAuD;IACvD,QAAQ,CAAC,EAAE;QAAE,IAAI,EAAE,MAAM,CAAC;QAAC,KAAK,EAAE,MAAM,CAAA;KAAE,CAAC;IAC3C,IAAI,EAAE,MAAM,CAAC;IACb,MAAM,CAAC,EAAE,WAAW,CAAC,QAAQ,CAAC,CAAC;CAC/B;AAED,MAAM,WAAW,gBAAiB,SAAQ,eAAe;IACxD,oDAAoD;IACpD,WAAW,CAAC,EAAE,MAAM,CAAC;CACrB;AAED,MAAM,WAAW,eAAe;IAC/B,IAAI,EAAE,MAAM,CAAC;IACb,YAAY,EAAE,kBAAkB,CAAC;IACjC,cAAc,CAAC,EAAE,MAAM,CAAC;IACxB,WAAW,EAAE,MAAM,CAAC;IACpB,uDAAuD;IACvD,QAAQ,CAAC,EAAE;QAAE,IAAI,EAAE,MAAM,CAAC;QAAC,KAAK,EAAE,MAAM,CAAA;KAAE,CAAC;CAC3C;AAYD,wBAAsB,kBAAkB,CAAC,OAAO,EAAE,eAAe,GAAG,OAAO,CAAC,cAAc,CAAC,CA+I1F;AAED,wBAAsB,SAAS,CAAC,OAAO,EAAE,gBAAgB,GAAG,OAAO,CAAC,eAAe,CAAC,CAyBnF","sourcesContent":["/**\n * Hybrid search orchestrator: resolve mode, run retrievers in parallel, fuse\n * by rank, expand within budget, trace everything\n * (docs/hybrid-retrieval-design.md).\n *\n * Single-retriever modes flow through the same pipeline — rrfFuse over one\n * list preserves its order — so lexical, semantic, and hybrid all produce the\n * same result shape and the same trace record.\n *\n * `retrieveCandidates` is the candidate-level core (also used by the eval\n * harness, which needs forced modes, a configurable `k`, and no trace\n * pollution); `runSearch` wraps it with span expansion and tracing for the\n * tool.\n */\n\nimport type { EmbsearchService } from \"../embsearch/embsearch-service.js\";\nimport { adaptGrepHits, type ChunkLookup } from \"./adapter.js\";\nimport { assembleContext } from \"./context-assembler.js\";\nimport { crossEncoderRerank } from \"./cross-rerank.js\";\nimport { runLexicalRetriever } from \"./lexical-retriever.js\";\nimport { resolveSearchMode } from \"./mode.js\";\nimport { rerankCandidates } from \"./rerank.js\";\nimport { DEFAULT_RRF_K, rrfFuse } from \"./rrf.js\";\nimport { writeSearchTrace } from \"./trace.js\";\nimport type { CandidateSpan, FusedCandidate, RankedHit, ResolvedSearchMode, SearchMode, SearchTrace } from \"./types.js\";\n\n/** Raw grep line-hits fetched per query (pre-collapse). */\nconst LEXICAL_MATCH_LIMIT = 200;\n/** Adapted lexical candidates entering fusion. The eval gate showed the\n *  uncapped lexical tail diluting hybrid below plain semantic: lexical\n *  precision is front-loaded by the adapter's term-evidence ranking, while\n *  RRF weighs a rank-30 lexical candidate like a rank-30 embedding hit. */\nconst LEXICAL_FUSION_CAP = 20;\n/**\n * Embedding hits fetched per query when fusing with the grep leg.\n *\n * Deliberately shallow. The grep list is capped at {@link LEXICAL_FUSION_CAP}\n * because its tail is unranked noise, so a deep dense pool here simply\n * outnumbers it: raising this to 200 alongside a 20-candidate grep list cost\n * `auto +rr` 0.472 -> 0.450 MRR.\n */\nconst EMBED_TOP_K = 50;\n/**\n * Per-leg depth when both retrievers are ranked ones (dense + BM25).\n *\n * Four times {@link FUSED_WINDOW}, matching what the daemon's `query_hybrid`\n * does internally (`pool = 4·k`). Fusing at the same depth as the window loses\n * any candidate ranked well by one retriever but just outside the other's\n * top-50 — worth 12pp R@10 and 8pp R@50 for `bm25+dense`, which is exactly\n * what closed the gap to daemon-side fusion.\n */\nconst FUSION_POOL_TOP_K = 200;\n/** BM25 depth, matching {@link FUSION_POOL_TOP_K} so neither ranked leg is\n *  structurally advantaged by pool size. The daemon returns only documents\n *  sharing a query term, so this is an upper bound, not a fill. */\nconst BM25_TOP_K = FUSION_POOL_TOP_K;\n/** Fused candidates kept for reranking / final slicing. */\nconst FUSED_WINDOW = 50;\n\nexport interface RetrieveOptions {\n\tcwd: string;\n\tquery: string;\n\tmode?: SearchMode;\n\t/** Optional glob filter applied to file paths. */\n\tglob?: string;\n\t/** Maximum fused candidates returned. */\n\tlimit?: number;\n\t/** RRF constant override (eval harness sweeps this). Default: {@link DEFAULT_RRF_K}. */\n\trrfK?: number;\n\t/** Rerank the fused top-50 before slicing to `limit`. Default: true. */\n\trerank?: boolean;\n\t/**\n\t * Ask the daemon to fuse its own BM25 index with the vectors and return one\n\t * already-fused ranking, instead of taking a dense-only list. Needs a store\n\t * built with `--hybrid`. The fused list arrives as a single \"embed\" leg,\n\t * because a pre-fused ranking has no per-retriever structure left to record.\n\t *\n\t * Prefer {@link bm25Leg}: fusing here keeps the legs separable in the trace\n\t * and lets the grep leg participate.\n\t */\n\tdaemonHybrid?: boolean;\n\t/**\n\t * Fetch the daemon's BM25 index as its own ranked list and fuse it here,\n\t * alongside dense and grep. Needs a store built with `--hybrid` and an\n\t * embsearch new enough to serve `retriever: \"lexical\"`.\n\t */\n\tbm25Leg?: boolean;\n\t/**\n\t * Reorder the fused shortlist with the daemon's cross-encoder instead of\n\t * the deterministic reranker. Needs embsearch >= 0.3.0; costs one model\n\t * pass per scored candidate.\n\t */\n\tcrossEncoder?: boolean;\n\tservice?: EmbsearchService;\n\tsignal?: AbortSignal;\n}\n\nexport interface RetrieveResult {\n\tcandidates: FusedCandidate[];\n\tresolvedMode: ResolvedSearchMode;\n\tdegradedReason?: string;\n\tindexPhase: SearchTrace[\"indexPhase\"];\n\tretrievers: SearchTrace[\"retrievers\"];\n\t/** Set while the embedding index is still building. */\n\tindexing?: { done: number; total: number };\n\trrfK: number;\n\trerank?: SearchTrace[\"rerank\"];\n}\n\nexport interface RunSearchOptions extends RetrieveOptions {\n\t/** Approximate token budget for the result text. */\n\ttokenBudget?: number;\n}\n\nexport interface RunSearchResult {\n\ttext: string;\n\tresolvedMode: ResolvedSearchMode;\n\tdegradedReason?: string;\n\tresultCount: number;\n\t/** Set while the embedding index is still building. */\n\tindexing?: { done: number; total: number };\n}\n\nfunction normalizeSearchGlob(glob: string | undefined): string | undefined {\n\tif (!glob) return undefined;\n\t// Match fd/rg semantics: a slash-containing glob is anchored anywhere in\n\t// the tree, so prepend \"**/\" unless it already starts with a slash or \"**/\".\n\tif (glob.includes(\"/\") && !glob.startsWith(\"/\") && !glob.startsWith(\"**/\")) {\n\t\treturn `**/${glob}`;\n\t}\n\treturn glob;\n}\n\nexport async function retrieveCandidates(options: RetrieveOptions): Promise<RetrieveResult> {\n\tconst { cwd, query, service, signal } = options;\n\tconst glob = normalizeSearchGlob(options.glob);\n\tconst requestedMode = options.mode ?? \"auto\";\n\tconst limit = Math.max(1, options.limit ?? 10);\n\tconst rrfK = options.rrfK ?? DEFAULT_RRF_K;\n\n\tconst state = service?.getState();\n\tconst embedAvailable = service?.isAvailable() ?? false;\n\tconst embedUnavailableReason =\n\t\tstate === undefined\n\t\t\t? \"semantic index is not enabled\"\n\t\t\t: state.phase === \"unavailable\" || state.phase === \"skipped\"\n\t\t\t\t? state.reason\n\t\t\t\t: state.phase === \"idle\"\n\t\t\t\t\t? \"semantic index has not started\"\n\t\t\t\t\t: undefined;\n\n\tconst resolution = resolveSearchMode(query, requestedMode, embedAvailable, embedUnavailableReason);\n\tconst mode = resolution.mode;\n\n\t// Map lexical hits onto indexed chunk ids whenever the sidecar is usable,\n\t// even in lexical-only mode, so identities line up across modes.\n\tconst lookupChunk: ChunkLookup | undefined = embedAvailable\n\t\t? (rel, line) => service!.findEnclosingChunk(rel, line)\n\t\t: undefined;\n\n\tconst spans = new Map<string, CandidateSpan>();\n\tconst lists: RankedHit[][] = [];\n\tconst retrieverStats: SearchTrace[\"retrievers\"] = {};\n\tconst errors: Error[] = [];\n\n\tconst runLexical = async (): Promise<void> => {\n\t\tconst startedMs = Date.now();\n\t\ttry {\n\t\t\tconst lineHits = await runLexicalRetriever({ cwd, query, limit: LEXICAL_MATCH_LIMIT, glob, signal });\n\t\t\tconst adapted = adaptGrepHits(lineHits, lookupChunk);\n\t\t\t// In single-retriever lexical mode the full list is the result; in\n\t\t\t// hybrid, only the front-loaded head is trustworthy enough to vote.\n\t\t\tconst hits = mode === \"hybrid\" ? adapted.hits.slice(0, LEXICAL_FUSION_CAP) : adapted.hits;\n\t\t\tfor (const [id, span] of adapted.spans) if (!spans.has(id)) spans.set(id, span);\n\t\t\tlists.push(hits);\n\t\t\tretrieverStats.grep = { latencyMs: Date.now() - startedMs, hitCount: hits.length };\n\t\t} catch (e) {\n\t\t\terrors.push(e instanceof Error ? e : new Error(String(e)));\n\t\t\tretrieverStats.grep = { latencyMs: Date.now() - startedMs, hitCount: 0 };\n\t\t}\n\t};\n\n\tconst runEmbed = async (): Promise<void> => {\n\t\tconst startedMs = Date.now();\n\t\ttry {\n\t\t\t// The flat index pads top-k with whatever exists; with the cosine\n\t\t\t// metric the store uses, score <= 0 means \"no relation at all\", so\n\t\t\t// those padding hits would cast RRF votes on pure noise.\n\t\t\t// A BM25 leg is itself ranked, so the pair can afford — and needs — the\n\t\t\t// deeper pool; the grep leg cannot (see EMBED_TOP_K).\n\t\t\tconst topK = options.bm25Leg ? FUSION_POOL_TOP_K : EMBED_TOP_K;\n\t\t\tconst chunkHits = (\n\t\t\t\tawait service!.searchChunks(query, topK, glob, options.daemonHybrid ? \"hybrid\" : \"dense\")\n\t\t\t).filter((hit) => hit.score > 0);\n\t\t\tconst hits: RankedHit[] = chunkHits.map((hit, i) => ({\n\t\t\t\tid: hit.id,\n\t\t\t\trank: i + 1,\n\t\t\t\tscore: hit.score,\n\t\t\t\tsource: \"embed\",\n\t\t\t}));\n\t\t\tfor (const hit of chunkHits) {\n\t\t\t\tspans.set(hit.id, { path: hit.path, startLine: hit.startLine, endLine: hit.endLine });\n\t\t\t}\n\t\t\tlists.push(hits);\n\t\t\tretrieverStats.embed = { latencyMs: Date.now() - startedMs, hitCount: hits.length };\n\t\t} catch (e) {\n\t\t\terrors.push(e instanceof Error ? e : new Error(String(e)));\n\t\t\tretrieverStats.embed = { latencyMs: Date.now() - startedMs, hitCount: 0 };\n\t\t}\n\t};\n\n\tconst runBm25 = async (): Promise<void> => {\n\t\tconst startedMs = Date.now();\n\t\ttry {\n\t\t\t// Raw BM25 sums; only the ordering enters fusion, the score is a\n\t\t\t// diagnostic. Unlike the dense leg there is no zero-score padding to\n\t\t\t// filter — the daemon omits documents sharing no query term.\n\t\t\tconst chunkHits = await service!.searchChunks(query, BM25_TOP_K, glob, \"lexical\");\n\t\t\tconst hits: RankedHit[] = chunkHits.map((hit, i) => ({\n\t\t\t\tid: hit.id,\n\t\t\t\trank: i + 1,\n\t\t\t\tscore: hit.score,\n\t\t\t\tsource: \"bm25\",\n\t\t\t}));\n\t\t\tfor (const hit of chunkHits) {\n\t\t\t\tif (!spans.has(hit.id)) {\n\t\t\t\t\tspans.set(hit.id, { path: hit.path, startLine: hit.startLine, endLine: hit.endLine });\n\t\t\t\t}\n\t\t\t}\n\t\t\tlists.push(hits);\n\t\t\tretrieverStats.bm25 = { latencyMs: Date.now() - startedMs, hitCount: hits.length };\n\t\t} catch (e) {\n\t\t\terrors.push(e instanceof Error ? e : new Error(String(e)));\n\t\t\tretrieverStats.bm25 = { latencyMs: Date.now() - startedMs, hitCount: 0 };\n\t\t}\n\t};\n\n\tconst runs: Promise<void>[] = [];\n\tif (mode === \"lexical\" || mode === \"hybrid\") runs.push(runLexical());\n\tif (mode === \"semantic\" || mode === \"hybrid\") runs.push(runEmbed());\n\tif (options.bm25Leg && embedAvailable && mode !== \"lexical\") runs.push(runBm25());\n\tawait Promise.all(runs);\n\tif (signal?.aborted) throw new Error(\"Operation aborted\");\n\t// A partial failure in hybrid degrades to whichever retriever survived;\n\t// only a total loss is an error.\n\tif (lists.length === 0) throw errors[0] ?? new Error(\"search produced no retriever results\");\n\n\tconst fused = rrfFuse(lists, rrfK).slice(0, FUSED_WINDOW);\n\tlet candidates: FusedCandidate[] = [];\n\tfor (const hit of fused) {\n\t\tconst span = spans.get(hit.id);\n\t\tif (span) candidates.push({ ...hit, ...span });\n\t}\n\n\tlet rerankInfo: SearchTrace[\"rerank\"];\n\tif (options.crossEncoder && service?.supportsCrossEncoder()) {\n\t\tconst reranked = await crossEncoderRerank(query, candidates, cwd, service);\n\t\trerankInfo = { applied: true, candidateCount: reranked.scored, latencyMs: reranked.latencyMs };\n\t\tcandidates = reranked.candidates;\n\t} else if (options.rerank !== false) {\n\t\tconst reranked = rerankCandidates(query, candidates, cwd);\n\t\trerankInfo = { applied: true, candidateCount: candidates.length, latencyMs: reranked.latencyMs };\n\t\tcandidates = reranked.candidates;\n\t}\n\tcandidates = candidates.slice(0, limit);\n\n\treturn {\n\t\tcandidates,\n\t\tresolvedMode: mode,\n\t\tdegradedReason: resolution.degradedReason,\n\t\tindexPhase: state?.phase === \"ready\" ? \"ready\" : state?.phase === \"indexing\" ? \"indexing\" : \"unavailable\",\n\t\tretrievers: retrieverStats,\n\t\tindexing: state?.phase === \"indexing\" ? { done: state.done, total: state.total } : undefined,\n\t\trrfK,\n\t\trerank: rerankInfo,\n\t};\n}\n\nexport async function runSearch(options: RunSearchOptions): Promise<RunSearchResult> {\n\tconst retrieved = await retrieveCandidates(options);\n\n\tconst assembled = assembleContext(retrieved.candidates, { cwd: options.cwd, tokenBudget: options.tokenBudget });\n\n\twriteSearchTrace(options.cwd, {\n\t\ttimestampMs: Date.now(),\n\t\tquery: options.query,\n\t\trequestedMode: options.mode ?? \"auto\",\n\t\tresolvedMode: retrieved.resolvedMode,\n\t\tdegradedReason: retrieved.degradedReason,\n\t\tindexPhase: retrieved.indexPhase,\n\t\trrfK: retrieved.resolvedMode === \"hybrid\" ? retrieved.rrfK : undefined,\n\t\tretrievers: retrieved.retrievers,\n\t\tfused: retrieved.candidates.map(({ id, rrfScore, ranks, rawScores }) => ({ id, rrfScore, ranks, rawScores })),\n\t\trerank: retrieved.rerank,\n\t});\n\n\treturn {\n\t\ttext: assembled.text,\n\t\tresolvedMode: retrieved.resolvedMode,\n\t\tdegradedReason: retrieved.degradedReason,\n\t\tresultCount: retrieved.candidates.length,\n\t\tindexing: retrieved.indexing,\n\t};\n}\n"]}