@kolisachint/hoocode-agent 0.5.59 → 0.5.60

package/CHANGELOG.md

@@ -1,5 +1,35 @@
 # Changelog
 
+## [0.5.60] - 2026-09-05
+
+### Fixed
+
+- **A semantic index built by an older embedding model now rebuilds instead of
+  switching semantic search off for good.** The embsearch daemon refuses to open
+  a store whose recorded model disagrees with its own — rightly, since vectors
+  from two models are not comparable — but refusing was where it stopped: the
+  store stayed on disk, the daemon never came up, and the service parked itself
+  in `unavailable` with a rebuild one directory-removal away. The check meant to
+  catch this ran *after* the spawn and reset only the sidecar, never the store,
+  so on a model change it was unreachable anyway. Recovery now catches the
+  refusal and confirms, via embsearch's `store-info` — which reads a store's
+  manifest without loading a model, and so is the only thing that can still
+  identify a store nothing will open — that a readable store is what the daemon
+  choked on, then rebuilds. A store it cannot read is left alone and the original
+  error stands, because deleting an index is the wrong answer to an unknown
+  fault. This was about to bite: embsearch 0.3.2 changed its model id, and only
+  the fact that an already-installed binary is never upgraded in place kept
+  anyone from hitting it. `store-info` ships in **embsearch 0.3.3**; against an
+  older binary the probe simply fails and the previous behaviour stands, so no
+  version floor is needed and nothing regresses on an older daemon.
+- **A chunker change no longer leaves dead text in the index.** Bumping
+  `CHUNKER_VERSION` resets the sidecar, and the sidecar is the only record of how
+  many chunks each file produced — so it reported zero for every file, and the
+  loop that drops superseded chunks counts down from that number. Any file that
+  re-chunked into fewer pieces left its tail vectors behind holding text that no
+  longer existed anywhere, retrievable forever. Upserts hid it: chunk counts
+  looked right, the sidecar looked right, only the search results were wrong.
+
 ## [0.5.59] - 2026-09-05
 
 ### Fixed

package/dist/core/embsearch/chunker.d.ts

@@ -11,6 +11,17 @@
  * sidecar triggers a clean rebuild of the store.
  */
 export declare const CHUNKER_VERSION = 2;
+/**
+ * Hard character cap per chunk.
+ *
+ * The "~256 tokens ≈ 1000 chars" this was set from does not hold for code:
+ * measured against the bundled tokenizer over this repo, 1000 chars is **313
+ * tokens** at the median, so MiniLM (256) truncates 85.8% of chunks and drops
+ * 20.3% of the corpus's tokens, while bge-small (512) drops 0.1%. Raising this
+ * is therefore not free in the way the old comment implied — it spends a
+ * budget that is already overdrawn on one model and nearly full on the other.
+ */
+export declare const CHUNK_MAX_CHARS = 1000;
 export interface Chunk {
     /** `relpath#index` — the id stored in the vector index. */
     id: string;
@@ -24,6 +35,11 @@ export interface Chunk {
 /**
  * Split `content` into chunks. `relPath` becomes the id prefix. Returns an
  * empty array for empty or binary-looking content.
+ *
+ * `maxChars` overrides {@link CHUNK_MAX_CHARS} for eval arms that sweep the
+ * window. Production never passes it; a caller that does is changing what the
+ * index contains and owes the store a distinct key, since nothing about a
+ * stored vector records the cap it was built under.
  */
-export declare function chunkFile(relPath: string, content: string): Chunk[];
+export declare function chunkFile(relPath: string, content: string, maxChars?: number): Chunk[];
 //# sourceMappingURL=chunker.d.ts.map

package/dist/core/embsearch/chunker.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"chunker.d.ts","sourceRoot":"","sources":["../../../src/core/embsearch/chunker.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;GAWG;AAEH,eAAO,MAAM,eAAe,IAAI,CAAC;AAmBjC,MAAM,WAAW,KAAK;IACrB,6DAA2D;IAC3D,EAAE,EAAE,MAAM,CAAC;IACX,iCAAiC;IACjC,IAAI,EAAE,MAAM,CAAC;IACb,oCAAoC;IACpC,SAAS,EAAE,MAAM,CAAC;IAClB,kCAAkC;IAClC,OAAO,EAAE,MAAM,CAAC;CAChB;AAQD;;;GAGG;AACH,wBAAgB,SAAS,CAAC,OAAO,EAAE,MAAM,EAAE,OAAO,EAAE,MAAM,GAAG,KAAK,EAAE,CA8CnE","sourcesContent":["/**\n * File chunking for semantic indexing.\n *\n * Splits a file into overlapping line windows, each capped by characters so a\n * chunk stays within the embedding model's effective token window (MiniLM\n * truncates around 256 tokens ≈ 1000 chars). Chunk ids are `relpath#index`;\n * the id → line-range mapping is kept in the sidecar metadata (index-meta.ts)\n * so search hits can be rendered as `path:start-end`.\n *\n * Bump CHUNKER_VERSION when the strategy changes — a version mismatch in the\n * sidecar triggers a clean rebuild of the store.\n */\n\nexport const CHUNKER_VERSION = 2;\n\n/**\n * Target lines per chunk.\n *\n * Halving this (30 lines / 500 chars) was measured and rejected. The theory\n * was dilution — a four-line answer sharing one vector with five neighbouring\n * methods — but sharper chunks cost reach: Recall@50 fell 79% -> 69% on\n * `auto +rr` and the boundary class went from 2 of 4 findable to 0 of 4.\n * A fixed top-k over smaller chunks retrieves less *content*: 50 chunks at\n * ~460 chars sees half the corpus that 50 at ~930 does. Recall@1 ticked up\n * (53% -> 55%), which is the sharpening, and it did not pay for the loss.\n */\nconst CHUNK_LINES = 60;\n/** Overlapping lines between consecutive chunks, for context continuity. */\nconst CHUNK_OVERLAP_LINES = 10;\n/** Hard character cap per chunk (MiniLM truncates ~256 tokens ≈ 1000 chars). */\nconst CHUNK_MAX_CHARS = 1000;\n\nexport interface Chunk {\n\t/** `relpath#index` — the id stored in the vector index. */\n\tid: string;\n\t/** Text sent to the embedder. */\n\ttext: string;\n\t/** 1-based inclusive start line. */\n\tstartLine: number;\n\t/** 1-based inclusive end line. */\n\tendLine: number;\n}\n\n/** Heuristic binary sniff: NUL byte in the first 8KB. */\nfunction looksBinary(content: string): boolean {\n\tconst probe = content.slice(0, 8192);\n\treturn probe.includes(\"\\u0000\");\n}\n\n/**\n * Split `content` into chunks. `relPath` becomes the id prefix. Returns an\n * empty array for empty or binary-looking content.\n */\nexport function chunkFile(relPath: string, content: string): Chunk[] {\n\tif (!content.trim() || looksBinary(content)) return [];\n\tconst lines = content.replace(/\\r\\n/g, \"\\n\").replace(/\\r/g, \"\\n\").split(\"\\n\");\n\tconst chunks: Chunk[] = [];\n\tlet start = 0; // 0-based\n\tlet index = 0;\n\n\twhile (start < lines.length) {\n\t\tlet end = start; // exclusive\n\t\tlet chars = 0;\n\t\twhile (end < lines.length && end - start < CHUNK_LINES) {\n\t\t\tconst lineLen = lines[end].length + 1;\n\t\t\tif (chars + lineLen > CHUNK_MAX_CHARS && end > start) break;\n\t\t\tchars += lineLen;\n\t\t\tend++;\n\t\t}\n\t\t// Trim whole blank lines off both ends, and narrow the recorded range to\n\t\t// match. Trimming the joined string instead left `startLine`/`endLine`\n\t\t// claiming lines whose text was never embedded — 31% of chunks in this\n\t\t// repo, up to 3 lines each. Those lines were unfindable by the vector\n\t\t// leg while still being handed to the context assembler as if they were\n\t\t// part of the chunk.\n\t\tlet from = start;\n\t\tlet to = end; // exclusive\n\t\twhile (from < to && lines[from].trim() === \"\") from++;\n\t\twhile (to > from && lines[to - 1].trim() === \"\") to--;\n\t\tlet text = lines.slice(from, to).join(\"\\n\").trim();\n\t\tif (text.length > CHUNK_MAX_CHARS) {\n\t\t\t// Oversized chunk (e.g. long minified line): keep the prefix. The\n\t\t\t// underlying model would truncate anyway, so this stays bounded.\n\t\t\ttext = text.slice(0, CHUNK_MAX_CHARS);\n\t\t}\n\t\tif (text) {\n\t\t\tchunks.push({\n\t\t\t\tid: `${relPath}#${index}`,\n\t\t\t\ttext,\n\t\t\t\tstartLine: from + 1,\n\t\t\t\tendLine: to,\n\t\t\t});\n\t\t\tindex++;\n\t\t}\n\t\tif (end >= lines.length) break;\n\t\t// Step forward with overlap, but always make progress.\n\t\tstart = Math.max(end - CHUNK_OVERLAP_LINES, start + 1);\n\t}\n\treturn chunks;\n}\n"]}
+{"version":3,"file":"chunker.d.ts","sourceRoot":"","sources":["../../../src/core/embsearch/chunker.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;GAWG;AAEH,eAAO,MAAM,eAAe,IAAI,CAAC;AAgBjC;;;;;;;;;GASG;AACH,eAAO,MAAM,eAAe,OAAO,CAAC;AAEpC,MAAM,WAAW,KAAK;IACrB,6DAA2D;IAC3D,EAAE,EAAE,MAAM,CAAC;IACX,iCAAiC;IACjC,IAAI,EAAE,MAAM,CAAC;IACb,oCAAoC;IACpC,SAAS,EAAE,MAAM,CAAC;IAClB,kCAAkC;IAClC,OAAO,EAAE,MAAM,CAAC;CAChB;AAQD;;;;;;;;GAQG;AACH,wBAAgB,SAAS,CAAC,OAAO,EAAE,MAAM,EAAE,OAAO,EAAE,MAAM,EAAE,QAAQ,GAAE,MAAwB,GAAG,KAAK,EAAE,CA8CvG","sourcesContent":["/**\n * File chunking for semantic indexing.\n *\n * Splits a file into overlapping line windows, each capped by characters so a\n * chunk stays within the embedding model's effective token window (MiniLM\n * truncates around 256 tokens ≈ 1000 chars). Chunk ids are `relpath#index`;\n * the id → line-range mapping is kept in the sidecar metadata (index-meta.ts)\n * so search hits can be rendered as `path:start-end`.\n *\n * Bump CHUNKER_VERSION when the strategy changes — a version mismatch in the\n * sidecar triggers a clean rebuild of the store.\n */\n\nexport const CHUNKER_VERSION = 2;\n\n/**\n * Target lines per chunk.\n *\n * Halving this (30 lines / 500 chars) was measured and rejected. The theory\n * was dilution — a four-line answer sharing one vector with five neighbouring\n * methods — but sharper chunks cost reach: Recall@50 fell 79% -> 69% on\n * `auto +rr` and the boundary class went from 2 of 4 findable to 0 of 4.\n * A fixed top-k over smaller chunks retrieves less *content*: 50 chunks at\n * ~460 chars sees half the corpus that 50 at ~930 does. Recall@1 ticked up\n * (53% -> 55%), which is the sharpening, and it did not pay for the loss.\n */\nconst CHUNK_LINES = 60;\n/** Overlapping lines between consecutive chunks, for context continuity. */\nconst CHUNK_OVERLAP_LINES = 10;\n/**\n * Hard character cap per chunk.\n *\n * The \"~256 tokens ≈ 1000 chars\" this was set from does not hold for code:\n * measured against the bundled tokenizer over this repo, 1000 chars is **313\n * tokens** at the median, so MiniLM (256) truncates 85.8% of chunks and drops\n * 20.3% of the corpus's tokens, while bge-small (512) drops 0.1%. Raising this\n * is therefore not free in the way the old comment implied — it spends a\n * budget that is already overdrawn on one model and nearly full on the other.\n */\nexport const CHUNK_MAX_CHARS = 1000;\n\nexport interface Chunk {\n\t/** `relpath#index` — the id stored in the vector index. */\n\tid: string;\n\t/** Text sent to the embedder. */\n\ttext: string;\n\t/** 1-based inclusive start line. */\n\tstartLine: number;\n\t/** 1-based inclusive end line. */\n\tendLine: number;\n}\n\n/** Heuristic binary sniff: NUL byte in the first 8KB. */\nfunction looksBinary(content: string): boolean {\n\tconst probe = content.slice(0, 8192);\n\treturn probe.includes(\"\\u0000\");\n}\n\n/**\n * Split `content` into chunks. `relPath` becomes the id prefix. Returns an\n * empty array for empty or binary-looking content.\n *\n * `maxChars` overrides {@link CHUNK_MAX_CHARS} for eval arms that sweep the\n * window. Production never passes it; a caller that does is changing what the\n * index contains and owes the store a distinct key, since nothing about a\n * stored vector records the cap it was built under.\n */\nexport function chunkFile(relPath: string, content: string, maxChars: number = CHUNK_MAX_CHARS): Chunk[] {\n\tif (!content.trim() || looksBinary(content)) return [];\n\tconst lines = content.replace(/\\r\\n/g, \"\\n\").replace(/\\r/g, \"\\n\").split(\"\\n\");\n\tconst chunks: Chunk[] = [];\n\tlet start = 0; // 0-based\n\tlet index = 0;\n\n\twhile (start < lines.length) {\n\t\tlet end = start; // exclusive\n\t\tlet chars = 0;\n\t\twhile (end < lines.length && end - start < CHUNK_LINES) {\n\t\t\tconst lineLen = lines[end].length + 1;\n\t\t\tif (chars + lineLen > maxChars && end > start) break;\n\t\t\tchars += lineLen;\n\t\t\tend++;\n\t\t}\n\t\t// Trim whole blank lines off both ends, and narrow the recorded range to\n\t\t// match. Trimming the joined string instead left `startLine`/`endLine`\n\t\t// claiming lines whose text was never embedded — 31% of chunks in this\n\t\t// repo, up to 3 lines each. Those lines were unfindable by the vector\n\t\t// leg while still being handed to the context assembler as if they were\n\t\t// part of the chunk.\n\t\tlet from = start;\n\t\tlet to = end; // exclusive\n\t\twhile (from < to && lines[from].trim() === \"\") from++;\n\t\twhile (to > from && lines[to - 1].trim() === \"\") to--;\n\t\tlet text = lines.slice(from, to).join(\"\\n\").trim();\n\t\tif (text.length > maxChars) {\n\t\t\t// Oversized chunk (e.g. long minified line): keep the prefix. The\n\t\t\t// underlying model would truncate anyway, so this stays bounded.\n\t\t\ttext = text.slice(0, maxChars);\n\t\t}\n\t\tif (text) {\n\t\t\tchunks.push({\n\t\t\t\tid: `${relPath}#${index}`,\n\t\t\t\ttext,\n\t\t\t\tstartLine: from + 1,\n\t\t\t\tendLine: to,\n\t\t\t});\n\t\t\tindex++;\n\t\t}\n\t\tif (end >= lines.length) break;\n\t\t// Step forward with overlap, but always make progress.\n\t\tstart = Math.max(end - CHUNK_OVERLAP_LINES, start + 1);\n\t}\n\treturn chunks;\n}\n"]}

package/dist/core/embsearch/chunker.js

@@ -25,8 +25,17 @@ export const CHUNKER_VERSION = 2;
 const CHUNK_LINES = 60;
 /** Overlapping lines between consecutive chunks, for context continuity. */
 const CHUNK_OVERLAP_LINES = 10;
-/** Hard character cap per chunk (MiniLM truncates ~256 tokens ≈ 1000 chars). */
-const CHUNK_MAX_CHARS = 1000;
+/**
+ * Hard character cap per chunk.
+ *
+ * The "~256 tokens ≈ 1000 chars" this was set from does not hold for code:
+ * measured against the bundled tokenizer over this repo, 1000 chars is **313
+ * tokens** at the median, so MiniLM (256) truncates 85.8% of chunks and drops
+ * 20.3% of the corpus's tokens, while bge-small (512) drops 0.1%. Raising this
+ * is therefore not free in the way the old comment implied — it spends a
+ * budget that is already overdrawn on one model and nearly full on the other.
+ */
+export const CHUNK_MAX_CHARS = 1000;
 /** Heuristic binary sniff: NUL byte in the first 8KB. */
 function looksBinary(content) {
     const probe = content.slice(0, 8192);
@@ -35,8 +44,13 @@ function looksBinary(content) {
 /**
  * Split `content` into chunks. `relPath` becomes the id prefix. Returns an
  * empty array for empty or binary-looking content.
+ *
+ * `maxChars` overrides {@link CHUNK_MAX_CHARS} for eval arms that sweep the
+ * window. Production never passes it; a caller that does is changing what the
+ * index contains and owes the store a distinct key, since nothing about a
+ * stored vector records the cap it was built under.
  */
-export function chunkFile(relPath, content) {
+export function chunkFile(relPath, content, maxChars = CHUNK_MAX_CHARS) {
     if (!content.trim() || looksBinary(content))
         return [];
     const lines = content.replace(/\r\n/g, "\n").replace(/\r/g, "\n").split("\n");
@@ -48,7 +62,7 @@ export function chunkFile(relPath, content) {
         let chars = 0;
         while (end < lines.length && end - start < CHUNK_LINES) {
             const lineLen = lines[end].length + 1;
-            if (chars + lineLen > CHUNK_MAX_CHARS && end > start)
+            if (chars + lineLen > maxChars && end > start)
                 break;
             chars += lineLen;
             end++;
@@ -66,10 +80,10 @@ export function chunkFile(relPath, content) {
         while (to > from && lines[to - 1].trim() === "")
             to--;
         let text = lines.slice(from, to).join("\n").trim();
-        if (text.length > CHUNK_MAX_CHARS) {
+        if (text.length > maxChars) {
             // Oversized chunk (e.g. long minified line): keep the prefix. The
             // underlying model would truncate anyway, so this stays bounded.
-            text = text.slice(0, CHUNK_MAX_CHARS);
+            text = text.slice(0, maxChars);
         }
         if (text) {
             chunks.push({

package/dist/core/embsearch/chunker.js.map

@@ -1 +1 @@
-{"version":3,"file":"chunker.js","sourceRoot":"","sources":["../../../src/core/embsearch/chunker.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;GAWG;AAEH,MAAM,CAAC,MAAM,eAAe,GAAG,CAAC,CAAC;AAEjC;;;;;;;;;;GAUG;AACH,MAAM,WAAW,GAAG,EAAE,CAAC;AACvB,4EAA4E;AAC5E,MAAM,mBAAmB,GAAG,EAAE,CAAC;AAC/B,kFAAgF;AAChF,MAAM,eAAe,GAAG,IAAI,CAAC;AAa7B,yDAAyD;AACzD,SAAS,WAAW,CAAC,OAAe,EAAW;IAC9C,MAAM,KAAK,GAAG,OAAO,CAAC,KAAK,CAAC,CAAC,EAAE,IAAI,CAAC,CAAC;IACrC,OAAO,KAAK,CAAC,QAAQ,CAAC,QAAQ,CAAC,CAAC;AAAA,CAChC;AAED;;;GAGG;AACH,MAAM,UAAU,SAAS,CAAC,OAAe,EAAE,OAAe,EAAW;IACpE,IAAI,CAAC,OAAO,CAAC,IAAI,EAAE,IAAI,WAAW,CAAC,OAAO,CAAC;QAAE,OAAO,EAAE,CAAC;IACvD,MAAM,KAAK,GAAG,OAAO,CAAC,OAAO,CAAC,OAAO,EAAE,IAAI,CAAC,CAAC,OAAO,CAAC,KAAK,EAAE,IAAI,CAAC,CAAC,KAAK,CAAC,IAAI,CAAC,CAAC;IAC9E,MAAM,MAAM,GAAY,EAAE,CAAC;IAC3B,IAAI,KAAK,GAAG,CAAC,CAAC,CAAC,UAAU;IACzB,IAAI,KAAK,GAAG,CAAC,CAAC;IAEd,OAAO,KAAK,GAAG,KAAK,CAAC,MAAM,EAAE,CAAC;QAC7B,IAAI,GAAG,GAAG,KAAK,CAAC,CAAC,YAAY;QAC7B,IAAI,KAAK,GAAG,CAAC,CAAC;QACd,OAAO,GAAG,GAAG,KAAK,CAAC,MAAM,IAAI,GAAG,GAAG,KAAK,GAAG,WAAW,EAAE,CAAC;YACxD,MAAM,OAAO,GAAG,KAAK,CAAC,GAAG,CAAC,CAAC,MAAM,GAAG,CAAC,CAAC;YACtC,IAAI,KAAK,GAAG,OAAO,GAAG,eAAe,IAAI,GAAG,GAAG,KAAK;gBAAE,MAAM;YAC5D,KAAK,IAAI,OAAO,CAAC;YACjB,GAAG,EAAE,CAAC;QACP,CAAC;QACD,yEAAyE;QACzE,uEAAuE;QACvE,yEAAuE;QACvE,sEAAsE;QACtE,wEAAwE;QACxE,qBAAqB;QACrB,IAAI,IAAI,GAAG,KAAK,CAAC;QACjB,IAAI,EAAE,GAAG,GAAG,CAAC,CAAC,YAAY;QAC1B,OAAO,IAAI,GAAG,EAAE,IAAI,KAAK,CAAC,IAAI,CAAC,CAAC,IAAI,EAAE,KAAK,EAAE;YAAE,IAAI,EAAE,CAAC;QACtD,OAAO,EAAE,GAAG,IAAI,IAAI,KAAK,CAAC,EAAE,GAAG,CAAC,CAAC,CAAC,IAAI,EAAE,KAAK,EAAE;YAAE,EAAE,EAAE,CAAC;QACtD,IAAI,IAAI,GAAG,KAAK,CAAC,KAAK,CAAC,IAAI,EAAE,EAAE,CAAC,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC,IAAI,EAAE,CAAC;QACnD,IAAI,IAAI,CAAC,MAAM,GAAG,eAAe,EAAE,CAAC;YACnC,kEAAkE;YAClE,iEAAiE;YACjE,IAAI,GAAG,IAAI,CAAC,KAAK,CAAC,CAAC,EAAE,eAAe,CAAC,CAAC;QACvC,CAAC;QACD,IAAI,IAAI,EAAE,CAAC;YACV,MAAM,CAAC,IAAI,CAAC;gBACX,EAAE,EAAE,GAAG,OAAO,IAAI,KAAK,EAAE;gBACzB,IAAI;gBACJ,SAAS,EAAE,IAAI,GAAG,CAAC;gBACnB,OAAO,EAAE,EAAE;aACX,CAAC,CAAC;YACH,KAAK,EAAE,CAAC;QACT,CAAC;QACD,IAAI,GAAG,IAAI,KAAK,CAAC,MAAM;YAAE,MAAM;QAC/B,uDAAuD;QACvD,KAAK,GAAG,IAAI,CAAC,GAAG,CAAC,GAAG,GAAG,mBAAmB,EAAE,KAAK,GAAG,CAAC,CAAC,CAAC;IACxD,CAAC;IACD,OAAO,MAAM,CAAC;AAAA,CACd","sourcesContent":["/**\n * File chunking for semantic indexing.\n *\n * Splits a file into overlapping line windows, each capped by characters so a\n * chunk stays within the embedding model's effective token window (MiniLM\n * truncates around 256 tokens ≈ 1000 chars). Chunk ids are `relpath#index`;\n * the id → line-range mapping is kept in the sidecar metadata (index-meta.ts)\n * so search hits can be rendered as `path:start-end`.\n *\n * Bump CHUNKER_VERSION when the strategy changes — a version mismatch in the\n * sidecar triggers a clean rebuild of the store.\n */\n\nexport const CHUNKER_VERSION = 2;\n\n/**\n * Target lines per chunk.\n *\n * Halving this (30 lines / 500 chars) was measured and rejected. The theory\n * was dilution — a four-line answer sharing one vector with five neighbouring\n * methods — but sharper chunks cost reach: Recall@50 fell 79% -> 69% on\n * `auto +rr` and the boundary class went from 2 of 4 findable to 0 of 4.\n * A fixed top-k over smaller chunks retrieves less *content*: 50 chunks at\n * ~460 chars sees half the corpus that 50 at ~930 does. Recall@1 ticked up\n * (53% -> 55%), which is the sharpening, and it did not pay for the loss.\n */\nconst CHUNK_LINES = 60;\n/** Overlapping lines between consecutive chunks, for context continuity. */\nconst CHUNK_OVERLAP_LINES = 10;\n/** Hard character cap per chunk (MiniLM truncates ~256 tokens ≈ 1000 chars). */\nconst CHUNK_MAX_CHARS = 1000;\n\nexport interface Chunk {\n\t/** `relpath#index` — the id stored in the vector index. */\n\tid: string;\n\t/** Text sent to the embedder. */\n\ttext: string;\n\t/** 1-based inclusive start line. */\n\tstartLine: number;\n\t/** 1-based inclusive end line. */\n\tendLine: number;\n}\n\n/** Heuristic binary sniff: NUL byte in the first 8KB. */\nfunction looksBinary(content: string): boolean {\n\tconst probe = content.slice(0, 8192);\n\treturn probe.includes(\"\\u0000\");\n}\n\n/**\n * Split `content` into chunks. `relPath` becomes the id prefix. Returns an\n * empty array for empty or binary-looking content.\n */\nexport function chunkFile(relPath: string, content: string): Chunk[] {\n\tif (!content.trim() || looksBinary(content)) return [];\n\tconst lines = content.replace(/\\r\\n/g, \"\\n\").replace(/\\r/g, \"\\n\").split(\"\\n\");\n\tconst chunks: Chunk[] = [];\n\tlet start = 0; // 0-based\n\tlet index = 0;\n\n\twhile (start < lines.length) {\n\t\tlet end = start; // exclusive\n\t\tlet chars = 0;\n\t\twhile (end < lines.length && end - start < CHUNK_LINES) {\n\t\t\tconst lineLen = lines[end].length + 1;\n\t\t\tif (chars + lineLen > CHUNK_MAX_CHARS && end > start) break;\n\t\t\tchars += lineLen;\n\t\t\tend++;\n\t\t}\n\t\t// Trim whole blank lines off both ends, and narrow the recorded range to\n\t\t// match. Trimming the joined string instead left `startLine`/`endLine`\n\t\t// claiming lines whose text was never embedded — 31% of chunks in this\n\t\t// repo, up to 3 lines each. Those lines were unfindable by the vector\n\t\t// leg while still being handed to the context assembler as if they were\n\t\t// part of the chunk.\n\t\tlet from = start;\n\t\tlet to = end; // exclusive\n\t\twhile (from < to && lines[from].trim() === \"\") from++;\n\t\twhile (to > from && lines[to - 1].trim() === \"\") to--;\n\t\tlet text = lines.slice(from, to).join(\"\\n\").trim();\n\t\tif (text.length > CHUNK_MAX_CHARS) {\n\t\t\t// Oversized chunk (e.g. long minified line): keep the prefix. The\n\t\t\t// underlying model would truncate anyway, so this stays bounded.\n\t\t\ttext = text.slice(0, CHUNK_MAX_CHARS);\n\t\t}\n\t\tif (text) {\n\t\t\tchunks.push({\n\t\t\t\tid: `${relPath}#${index}`,\n\t\t\t\ttext,\n\t\t\t\tstartLine: from + 1,\n\t\t\t\tendLine: to,\n\t\t\t});\n\t\t\tindex++;\n\t\t}\n\t\tif (end >= lines.length) break;\n\t\t// Step forward with overlap, but always make progress.\n\t\tstart = Math.max(end - CHUNK_OVERLAP_LINES, start + 1);\n\t}\n\treturn chunks;\n}\n"]}
+{"version":3,"file":"chunker.js","sourceRoot":"","sources":["../../../src/core/embsearch/chunker.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;GAWG;AAEH,MAAM,CAAC,MAAM,eAAe,GAAG,CAAC,CAAC;AAEjC;;;;;;;;;;GAUG;AACH,MAAM,WAAW,GAAG,EAAE,CAAC;AACvB,4EAA4E;AAC5E,MAAM,mBAAmB,GAAG,EAAE,CAAC;AAC/B;;;;;;;;;GASG;AACH,MAAM,CAAC,MAAM,eAAe,GAAG,IAAI,CAAC;AAapC,yDAAyD;AACzD,SAAS,WAAW,CAAC,OAAe,EAAW;IAC9C,MAAM,KAAK,GAAG,OAAO,CAAC,KAAK,CAAC,CAAC,EAAE,IAAI,CAAC,CAAC;IACrC,OAAO,KAAK,CAAC,QAAQ,CAAC,QAAQ,CAAC,CAAC;AAAA,CAChC;AAED;;;;;;;;GAQG;AACH,MAAM,UAAU,SAAS,CAAC,OAAe,EAAE,OAAe,EAAE,QAAQ,GAAW,eAAe,EAAW;IACxG,IAAI,CAAC,OAAO,CAAC,IAAI,EAAE,IAAI,WAAW,CAAC,OAAO,CAAC;QAAE,OAAO,EAAE,CAAC;IACvD,MAAM,KAAK,GAAG,OAAO,CAAC,OAAO,CAAC,OAAO,EAAE,IAAI,CAAC,CAAC,OAAO,CAAC,KAAK,EAAE,IAAI,CAAC,CAAC,KAAK,CAAC,IAAI,CAAC,CAAC;IAC9E,MAAM,MAAM,GAAY,EAAE,CAAC;IAC3B,IAAI,KAAK,GAAG,CAAC,CAAC,CAAC,UAAU;IACzB,IAAI,KAAK,GAAG,CAAC,CAAC;IAEd,OAAO,KAAK,GAAG,KAAK,CAAC,MAAM,EAAE,CAAC;QAC7B,IAAI,GAAG,GAAG,KAAK,CAAC,CAAC,YAAY;QAC7B,IAAI,KAAK,GAAG,CAAC,CAAC;QACd,OAAO,GAAG,GAAG,KAAK,CAAC,MAAM,IAAI,GAAG,GAAG,KAAK,GAAG,WAAW,EAAE,CAAC;YACxD,MAAM,OAAO,GAAG,KAAK,CAAC,GAAG,CAAC,CAAC,MAAM,GAAG,CAAC,CAAC;YACtC,IAAI,KAAK,GAAG,OAAO,GAAG,QAAQ,IAAI,GAAG,GAAG,KAAK;gBAAE,MAAM;YACrD,KAAK,IAAI,OAAO,CAAC;YACjB,GAAG,EAAE,CAAC;QACP,CAAC;QACD,yEAAyE;QACzE,uEAAuE;QACvE,yEAAuE;QACvE,sEAAsE;QACtE,wEAAwE;QACxE,qBAAqB;QACrB,IAAI,IAAI,GAAG,KAAK,CAAC;QACjB,IAAI,EAAE,GAAG,GAAG,CAAC,CAAC,YAAY;QAC1B,OAAO,IAAI,GAAG,EAAE,IAAI,KAAK,CAAC,IAAI,CAAC,CAAC,IAAI,EAAE,KAAK,EAAE;YAAE,IAAI,EAAE,CAAC;QACtD,OAAO,EAAE,GAAG,IAAI,IAAI,KAAK,CAAC,EAAE,GAAG,CAAC,CAAC,CAAC,IAAI,EAAE,KAAK,EAAE;YAAE,EAAE,EAAE,CAAC;QACtD,IAAI,IAAI,GAAG,KAAK,CAAC,KAAK,CAAC,IAAI,EAAE,EAAE,CAAC,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC,IAAI,EAAE,CAAC;QACnD,IAAI,IAAI,CAAC,MAAM,GAAG,QAAQ,EAAE,CAAC;YAC5B,kEAAkE;YAClE,iEAAiE;YACjE,IAAI,GAAG,IAAI,CAAC,KAAK,CAAC,CAAC,EAAE,QAAQ,CAAC,CAAC;QAChC,CAAC;QACD,IAAI,IAAI,EAAE,CAAC;YACV,MAAM,CAAC,IAAI,CAAC;gBACX,EAAE,EAAE,GAAG,OAAO,IAAI,KAAK,EAAE;gBACzB,IAAI;gBACJ,SAAS,EAAE,IAAI,GAAG,CAAC;gBACnB,OAAO,EAAE,EAAE;aACX,CAAC,CAAC;YACH,KAAK,EAAE,CAAC;QACT,CAAC;QACD,IAAI,GAAG,IAAI,KAAK,CAAC,MAAM;YAAE,MAAM;QAC/B,uDAAuD;QACvD,KAAK,GAAG,IAAI,CAAC,GAAG,CAAC,GAAG,GAAG,mBAAmB,EAAE,KAAK,GAAG,CAAC,CAAC,CAAC;IACxD,CAAC;IACD,OAAO,MAAM,CAAC;AAAA,CACd","sourcesContent":["/**\n * File chunking for semantic indexing.\n *\n * Splits a file into overlapping line windows, each capped by characters so a\n * chunk stays within the embedding model's effective token window (MiniLM\n * truncates around 256 tokens ≈ 1000 chars). Chunk ids are `relpath#index`;\n * the id → line-range mapping is kept in the sidecar metadata (index-meta.ts)\n * so search hits can be rendered as `path:start-end`.\n *\n * Bump CHUNKER_VERSION when the strategy changes — a version mismatch in the\n * sidecar triggers a clean rebuild of the store.\n */\n\nexport const CHUNKER_VERSION = 2;\n\n/**\n * Target lines per chunk.\n *\n * Halving this (30 lines / 500 chars) was measured and rejected. The theory\n * was dilution — a four-line answer sharing one vector with five neighbouring\n * methods — but sharper chunks cost reach: Recall@50 fell 79% -> 69% on\n * `auto +rr` and the boundary class went from 2 of 4 findable to 0 of 4.\n * A fixed top-k over smaller chunks retrieves less *content*: 50 chunks at\n * ~460 chars sees half the corpus that 50 at ~930 does. Recall@1 ticked up\n * (53% -> 55%), which is the sharpening, and it did not pay for the loss.\n */\nconst CHUNK_LINES = 60;\n/** Overlapping lines between consecutive chunks, for context continuity. */\nconst CHUNK_OVERLAP_LINES = 10;\n/**\n * Hard character cap per chunk.\n *\n * The \"~256 tokens ≈ 1000 chars\" this was set from does not hold for code:\n * measured against the bundled tokenizer over this repo, 1000 chars is **313\n * tokens** at the median, so MiniLM (256) truncates 85.8% of chunks and drops\n * 20.3% of the corpus's tokens, while bge-small (512) drops 0.1%. Raising this\n * is therefore not free in the way the old comment implied — it spends a\n * budget that is already overdrawn on one model and nearly full on the other.\n */\nexport const CHUNK_MAX_CHARS = 1000;\n\nexport interface Chunk {\n\t/** `relpath#index` — the id stored in the vector index. */\n\tid: string;\n\t/** Text sent to the embedder. */\n\ttext: string;\n\t/** 1-based inclusive start line. */\n\tstartLine: number;\n\t/** 1-based inclusive end line. */\n\tendLine: number;\n}\n\n/** Heuristic binary sniff: NUL byte in the first 8KB. */\nfunction looksBinary(content: string): boolean {\n\tconst probe = content.slice(0, 8192);\n\treturn probe.includes(\"\\u0000\");\n}\n\n/**\n * Split `content` into chunks. `relPath` becomes the id prefix. Returns an\n * empty array for empty or binary-looking content.\n *\n * `maxChars` overrides {@link CHUNK_MAX_CHARS} for eval arms that sweep the\n * window. Production never passes it; a caller that does is changing what the\n * index contains and owes the store a distinct key, since nothing about a\n * stored vector records the cap it was built under.\n */\nexport function chunkFile(relPath: string, content: string, maxChars: number = CHUNK_MAX_CHARS): Chunk[] {\n\tif (!content.trim() || looksBinary(content)) return [];\n\tconst lines = content.replace(/\\r\\n/g, \"\\n\").replace(/\\r/g, \"\\n\").split(\"\\n\");\n\tconst chunks: Chunk[] = [];\n\tlet start = 0; // 0-based\n\tlet index = 0;\n\n\twhile (start < lines.length) {\n\t\tlet end = start; // exclusive\n\t\tlet chars = 0;\n\t\twhile (end < lines.length && end - start < CHUNK_LINES) {\n\t\t\tconst lineLen = lines[end].length + 1;\n\t\t\tif (chars + lineLen > maxChars && end > start) break;\n\t\t\tchars += lineLen;\n\t\t\tend++;\n\t\t}\n\t\t// Trim whole blank lines off both ends, and narrow the recorded range to\n\t\t// match. Trimming the joined string instead left `startLine`/`endLine`\n\t\t// claiming lines whose text was never embedded — 31% of chunks in this\n\t\t// repo, up to 3 lines each. Those lines were unfindable by the vector\n\t\t// leg while still being handed to the context assembler as if they were\n\t\t// part of the chunk.\n\t\tlet from = start;\n\t\tlet to = end; // exclusive\n\t\twhile (from < to && lines[from].trim() === \"\") from++;\n\t\twhile (to > from && lines[to - 1].trim() === \"\") to--;\n\t\tlet text = lines.slice(from, to).join(\"\\n\").trim();\n\t\tif (text.length > maxChars) {\n\t\t\t// Oversized chunk (e.g. long minified line): keep the prefix. The\n\t\t\t// underlying model would truncate anyway, so this stays bounded.\n\t\t\ttext = text.slice(0, maxChars);\n\t\t}\n\t\tif (text) {\n\t\t\tchunks.push({\n\t\t\t\tid: `${relPath}#${index}`,\n\t\t\t\ttext,\n\t\t\t\tstartLine: from + 1,\n\t\t\t\tendLine: to,\n\t\t\t});\n\t\t\tindex++;\n\t\t}\n\t\tif (end >= lines.length) break;\n\t\t// Step forward with overlap, but always make progress.\n\t\tstart = Math.max(end - CHUNK_OVERLAP_LINES, start + 1);\n\t}\n\treturn chunks;\n}\n"]}

package/dist/core/embsearch/client.d.ts

@@ -57,6 +57,15 @@ export interface EmbSearchClientOptions {
     binaryPath: string;
     /** Store directory passed as `--path`. */
     storePath: string;
+    /**
+     * Model directory passed as `--model` (onnx builds only): a dir holding
+     * `model.onnx`, `tokenizer.json` and `model.json`.
+     *
+     * Omitted, the daemon uses the model bundled into the binary. Set, the
+     * binary no longer determines which model produced a vector — which is why
+     * the eval harness records `info.model_id` rather than the binary version.
+     */
+    modelDir?: string;
     /** Metric for a freshly created store. Default: "cosine". */
     metric?: "cosine" | "dot" | "euclidean";
 }

package/dist/core/embsearch/client.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"client.d.ts","sourceRoot":"","sources":["../../../src/core/embsearch/client.ts"],"names":[],"mappings":"AAAA;;;;;;;GAOG;AAIH,MAAM,WAAW,eAAe;IAC/B,EAAE,EAAE,MAAM,CAAC;IACX,KAAK,EAAE,MAAM,CAAC;CACd;AAED,MAAM,WAAW,mBAAmB;IACnC,OAAO,EAAE,MAAM,CAAC;IAChB,GAAG,EAAE,MAAM,CAAC;IACZ,KAAK,EAAE,MAAM,CAAC;IACd;;;;;;OAMG;IACH,MAAM,CAAC,EAAE,OAAO,CAAC;IACjB;;;;;;;;OAQG;IACH,MAAM,CAAC,EAAE,OAAO,CAAC;CACjB;AAED,oDAAoD;AACpD,MAAM,WAAW,sBAAsB;IACtC,EAAE,EAAE,MAAM,CAAC;IACX,IAAI,EAAE,MAAM,CAAC;CACb;AAED;wDACwD;AACxD,MAAM,WAAW,qBAAqB;IACrC,EAAE,EAAE,MAAM,CAAC;IACX,KAAK,EAAE,MAAM,CAAC;CACd;AAED,MAAM,WAAW,mBAAmB;IACnC,QAAQ,EAAE,MAAM,CAAC;IACjB,OAAO,EAAE,MAAM,CAAC;CAChB;AAED,wEAAwE;AACxE,MAAM,MAAM,eAAe,GAAG,OAAO,GAAG,SAAS,GAAG,QAAQ,CAAC;AAE7D,MAAM,WAAW,sBAAsB;IACtC,6EAA6E;IAC7E,MAAM,CAAC,EAAE,OAAO,CAAC;IACjB,sCAAsC;IACtC,UAAU,EAAE,MAAM,CAAC;IACnB,0CAA0C;IAC1C,SAAS,EAAE,MAAM,CAAC;IAClB,6DAA6D;IAC7D,MAAM,CAAC,EAAE,QAAQ,GAAG,KAAK,GAAG,WAAW,CAAC;CACxC;AA4BD,qBAAa,eAAe;IAC3B,OAAO,CAAC,IAAI,CAAiC;IAC7C,OAAO,CAAC,KAAK,CAAiB;IAC9B,OAAO,CAAC,MAAM,CAAM;IACpB,OAAO,CAAC,MAAM,CAAS;IACvB,OAAO,CAAC,YAAY,CAAgB;IAEpC,YAAY,IAAI,EAAE,sBAAsB,EA+BvC;IAED,6DAA6D;IAC7D,KAAK,IAAI,OAAO,CAAC,IAAI,CAAC,CAErB;IAED,IAAI,QAAQ,IAAI,OAAO,CAEtB;IAED,OAAO,CAAC,QAAQ;IAqBhB,OAAO,CAAC,IAAI;IAQZ;;;;;;;;;;;;OAYG;IACG,KAAK,CAAC,IAAI,EAAE,MAAM,EAAE,CAAC,SAAK,EAAE,SAAS,GAAE,eAAyB,GAAG,OAAO,CAAC,eAAe,EAAE,CAAC,CAKlG;IAED;;;;;;;;OAQG;IACG,MAAM,CAAC,KAAK,EAAE,MAAM,EAAE,QAAQ,EAAE,sBAAsB,EAAE,EAAE,CAAC,EAAE,MAAM,GAAG,OAAO,CAAC,qBAAqB,EAAE,CAAC,CAG3G;IAED;;;;OAIG;IACG,IAAI,CAAC,KAAK,EAAE,KAAK,CAAC;QAAE,EAAE,EAAE,MAAM,CAAC;QAAC,IAAI,EAAE,MAAM,CAAA;KAAE,CAAC,GAAG,OAAO,CAAC,mBAAmB,CAAC,CAGnF;IAED,qEAAqE;IAC/D,IAAI,IAAI,OAAO,CAAC,mBAAmB,CAAC,CASzC;IAED,yDAAyD;IACnD,MAAM,CAAC,EAAE,EAAE,MAAM,GAAG,OAAO,CAAC,OAAO,CAAC,CAGzC;IAED,uDAAuD;IACjD,OAAO,IAAI,OAAO,CAAC,IAAI,CAAC,CAE7B;IAED,gDAAgD;IAC1C,IAAI,IAAI,OAAO,CAAC,IAAI,CAAC,CAE1B;IAED,+DAA+D;IACzD,KAAK,IAAI,OAAO,CAAC,IAAI,CAAC,CAI3B;CACD","sourcesContent":["/**\n * Client for the `embsearch` stdio daemon (vendored from\n * github.com/kolisachint/embeddingsearchtools ts/client.ts).\n *\n * Spawns `embsearch serve` once and talks newline-delimited JSON over its\n * stdin/stdout. The process stays alive so the model and index load a single\n * time; every query after startup is hot.\n */\n\nimport { type ChildProcessWithoutNullStreams, spawn } from \"child_process\";\n\nexport interface EmbSearchResult {\n\tid: string;\n\tscore: number;\n}\n\nexport interface EmbSearchDaemonInfo {\n\tmodelId: string;\n\tdim: number;\n\tcount: number;\n\t/**\n\t * Whether the open store carries a BM25 index alongside its vectors.\n\t *\n\t * Fixed when the store is created — passing `--hybrid` at a non-hybrid\n\t * store only warns — so this is the only reliable way to find out, and the\n\t * signal that an existing store has to be rebuilt rather than reused.\n\t */\n\thybrid?: boolean;\n\t/**\n\t * Whether this daemon can actually serve `rerank`, or `undefined` from a\n\t * daemon too old to say.\n\t *\n\t * Reported separately from the version because the two came apart:\n\t * released binaries carry the `rerank` op but no longer bundle the ~23 MB\n\t * cross-encoder weights, so a version check alone would advertise a\n\t * reranker that fails on first call.\n\t */\n\trerank?: boolean;\n}\n\n/** One candidate sent for cross-encoder scoring. */\nexport interface EmbSearchRerankPassage {\n\tid: string;\n\ttext: string;\n}\n\n/** A cross-encoder relevance logit. Higher is more relevant, but the scale is\n *  unnormalized and comparable only within one call. */\nexport interface EmbSearchRerankResult {\n\tid: string;\n\tscore: number;\n}\n\nexport interface EmbSearchBulkResult {\n\tinserted: number;\n\tupdated: number;\n}\n\n/** Which daemon-side retriever answers a query (embsearch >= 0.2.0). */\nexport type DaemonRetriever = \"dense\" | \"lexical\" | \"hybrid\";\n\nexport interface EmbSearchClientOptions {\n\t/** Open/create the store with a BM25 lexical index alongside the vectors. */\n\thybrid?: boolean;\n\t/** Path to the `embsearch` binary. */\n\tbinaryPath: string;\n\t/** Store directory passed as `--path`. */\n\tstorePath: string;\n\t/** Metric for a freshly created store. Default: \"cosine\". */\n\tmetric?: \"cosine\" | \"dot\" | \"euclidean\";\n}\n\ninterface Pending {\n\tresolve: (value: EmbSearchRawResponse) => void;\n\treject: (err: Error) => void;\n}\n\ninterface EmbSearchRawResponse {\n\tok: boolean;\n\terror?: string;\n\tresults?: EmbSearchResult[];\n\tinserted?: boolean;\n\tremoved?: boolean;\n\tcount?: number;\n\tinserted_count?: number;\n\tupdated_count?: number;\n\tmodel_id?: string;\n\tdim?: number;\n\t/** Cross-encoder scores. Separate from `results` because the daemon's\n\t *  logits are not comparable to retrieval scores. */\n\treranked?: EmbSearchRerankResult[];\n\t/** `info` only: whether the open store has a BM25 index. */\n\thybrid?: boolean;\n\t/** `info` only: whether `rerank` will work. Absent on daemons too old to\n\t *  report it. */\n\trerank?: boolean;\n}\n\nexport class EmbSearchClient {\n\tprivate proc: ChildProcessWithoutNullStreams;\n\tprivate queue: Pending[] = [];\n\tprivate buffer = \"\";\n\tprivate closed = false;\n\tprivate readyPromise: Promise<void>;\n\n\tconstructor(opts: EmbSearchClientOptions) {\n\t\tconst args = [\"serve\", \"--path\", opts.storePath];\n\t\tif (opts.metric) args.push(\"--metric\", opts.metric);\n\t\t// Hybrid-ness is fixed when a store is created: passing --hybrid against\n\t\t// an existing non-hybrid store warns and is ignored daemon-side, so a\n\t\t// hybrid store needs its own directory.\n\t\tif (opts.hybrid) args.push(\"--hybrid\");\n\n\t\tthis.proc = spawn(opts.binaryPath, args, { stdio: [\"pipe\", \"pipe\", \"pipe\"] });\n\n\t\tthis.proc.stdout.setEncoding(\"utf8\");\n\t\tthis.proc.stdout.on(\"data\", (chunk: string) => this.onStdout(chunk));\n\t\t// Swallow the informational stderr banner; errors surface via responses/exit.\n\t\tthis.proc.stderr.resume();\n\n\t\t// Readiness = the daemon answering a ping (probes the actual request loop).\n\t\tthis.readyPromise = this.send({ op: \"ping\" }).then(() => undefined);\n\t\t// Spawn failures reject the pending ping via the exit handler; mark the\n\t\t// promise handled so a failed spawn doesn't raise an unhandled rejection\n\t\t// before ready() is awaited.\n\t\tthis.readyPromise.catch(() => {});\n\n\t\tthis.proc.on(\"error\", (err) => {\n\t\t\tthis.closed = true;\n\t\t\tfor (const p of this.queue.splice(0)) p.reject(err);\n\t\t});\n\t\tthis.proc.on(\"exit\", (code) => {\n\t\t\tthis.closed = true;\n\t\t\tconst err = new Error(`embsearch daemon exited (code ${code})`);\n\t\t\tfor (const p of this.queue.splice(0)) p.reject(err);\n\t\t});\n\t}\n\n\t/** Resolves once the daemon has loaded the model + index. */\n\tready(): Promise<void> {\n\t\treturn this.readyPromise;\n\t}\n\n\tget isClosed(): boolean {\n\t\treturn this.closed;\n\t}\n\n\tprivate onStdout(chunk: string): void {\n\t\tthis.buffer += chunk;\n\t\t// Each response is one line; dispatch FIFO against the pending queue.\n\t\tfor (let nl = this.buffer.indexOf(\"\\n\"); nl !== -1; nl = this.buffer.indexOf(\"\\n\")) {\n\t\t\tconst line = this.buffer.slice(0, nl).trim();\n\t\t\tthis.buffer = this.buffer.slice(nl + 1);\n\t\t\tif (!line) continue;\n\t\t\tconst pending = this.queue.shift();\n\t\t\tif (!pending) continue;\n\t\t\tlet msg: EmbSearchRawResponse;\n\t\t\ttry {\n\t\t\t\tmsg = JSON.parse(line) as EmbSearchRawResponse;\n\t\t\t} catch {\n\t\t\t\tpending.reject(new Error(`bad response: ${line}`));\n\t\t\t\tcontinue;\n\t\t\t}\n\t\t\tif (msg.ok) pending.resolve(msg);\n\t\t\telse pending.reject(new Error(msg.error ?? \"unknown error\"));\n\t\t}\n\t}\n\n\tprivate send(req: Record<string, unknown>): Promise<EmbSearchRawResponse> {\n\t\tif (this.closed) return Promise.reject(new Error(\"embsearch client is closed\"));\n\t\treturn new Promise<EmbSearchRawResponse>((resolve, reject) => {\n\t\t\tthis.queue.push({ resolve, reject });\n\t\t\tthis.proc.stdin.write(`${JSON.stringify(req)}\\n`);\n\t\t});\n\t}\n\n\t/**\n\t * Search for the top-`k` matches for `text`.\n\t *\n\t * `retriever` selects which leg answers:\n\t *  - `dense` (default) — vector search, the historical behaviour;\n\t *  - `lexical` — BM25 only, raw scores, no embedding computed;\n\t *  - `hybrid` — both, pre-fused by the daemon's own RRF constant.\n\t *\n\t * `lexical` and `hybrid` need a store created with `--hybrid`. Prefer\n\t * `lexical` over `hybrid` when fusing here: `hybrid` collapses both legs\n\t * into one RRF score, discarding the per-retriever ranks the trace records\n\t * and preventing n-way fusion with the grep leg.\n\t */\n\tasync query(text: string, k = 10, retriever: DaemonRetriever = \"dense\"): Promise<EmbSearchResult[]> {\n\t\tconst res = await this.send(\n\t\t\tretriever === \"dense\" ? { op: \"query\", text, k } : { op: \"query\", text, k, retriever },\n\t\t);\n\t\treturn res.results ?? [];\n\t}\n\n\t/**\n\t * Score `passages` against `query` with the daemon's cross-encoder and\n\t * return the best `k`, best first.\n\t *\n\t * Passages are sent inline rather than referenced by id: the caller has the\n\t * exact spans it intends to show the model, and a cross-encoder scores the\n\t * text it is given, so sending anything else would score the wrong thing.\n\t * Requires an onnx build with reranker weights (embsearch >= 0.3.0).\n\t */\n\tasync rerank(query: string, passages: EmbSearchRerankPassage[], k: number): Promise<EmbSearchRerankResult[]> {\n\t\tconst res = await this.send({ op: \"rerank\", query, passages, k });\n\t\treturn res.reranked ?? [];\n\t}\n\n\t/**\n\t * Batched insert-or-replace. One embedding inference for the whole batch —\n\t * the fast path for bulk indexing. Keep batches modest (e.g. 32–64) so a\n\t * concurrent query is not stuck behind a huge inference.\n\t */\n\tasync bulk(items: Array<{ id: string; text: string }>): Promise<EmbSearchBulkResult> {\n\t\tconst res = await this.send({ op: \"bulk\", items });\n\t\treturn { inserted: res.inserted_count ?? 0, updated: res.updated_count ?? 0 };\n\t}\n\n\t/** Model id, dimensionality, and live vector count of the daemon. */\n\tasync info(): Promise<EmbSearchDaemonInfo> {\n\t\tconst res = await this.send({ op: \"info\" });\n\t\treturn {\n\t\t\tmodelId: res.model_id ?? \"\",\n\t\t\tdim: res.dim ?? 0,\n\t\t\tcount: res.count ?? 0,\n\t\t\thybrid: res.hybrid,\n\t\t\trerank: res.rerank,\n\t\t};\n\t}\n\n\t/** Remove a record. Resolves to `true` if it existed. */\n\tasync remove(id: string): Promise<boolean> {\n\t\tconst res = await this.send({ op: \"remove\", id });\n\t\treturn res.removed === true;\n\t}\n\n\t/** Reclaim tombstoned rows left behind by `remove`. */\n\tasync compact(): Promise<void> {\n\t\tawait this.send({ op: \"compact\" });\n\t}\n\n\t/** Persist the index to the store directory. */\n\tasync save(): Promise<void> {\n\t\tawait this.send({ op: \"save\" });\n\t}\n\n\t/** Shut the daemon down, closing stdin so it exits cleanly. */\n\tasync close(): Promise<void> {\n\t\tif (this.closed) return;\n\t\tthis.proc.stdin.end();\n\t\tawait new Promise<void>((resolve) => this.proc.on(\"exit\", () => resolve()));\n\t}\n}\n"]}
+{"version":3,"file":"client.d.ts","sourceRoot":"","sources":["../../../src/core/embsearch/client.ts"],"names":[],"mappings":"AAAA;;;;;;;GAOG;AAIH,MAAM,WAAW,eAAe;IAC/B,EAAE,EAAE,MAAM,CAAC;IACX,KAAK,EAAE,MAAM,CAAC;CACd;AAED,MAAM,WAAW,mBAAmB;IACnC,OAAO,EAAE,MAAM,CAAC;IAChB,GAAG,EAAE,MAAM,CAAC;IACZ,KAAK,EAAE,MAAM,CAAC;IACd;;;;;;OAMG;IACH,MAAM,CAAC,EAAE,OAAO,CAAC;IACjB;;;;;;;;OAQG;IACH,MAAM,CAAC,EAAE,OAAO,CAAC;CACjB;AAED,oDAAoD;AACpD,MAAM,WAAW,sBAAsB;IACtC,EAAE,EAAE,MAAM,CAAC;IACX,IAAI,EAAE,MAAM,CAAC;CACb;AAED;wDACwD;AACxD,MAAM,WAAW,qBAAqB;IACrC,EAAE,EAAE,MAAM,CAAC;IACX,KAAK,EAAE,MAAM,CAAC;CACd;AAED,MAAM,WAAW,mBAAmB;IACnC,QAAQ,EAAE,MAAM,CAAC;IACjB,OAAO,EAAE,MAAM,CAAC;CAChB;AAED,wEAAwE;AACxE,MAAM,MAAM,eAAe,GAAG,OAAO,GAAG,SAAS,GAAG,QAAQ,CAAC;AAE7D,MAAM,WAAW,sBAAsB;IACtC,6EAA6E;IAC7E,MAAM,CAAC,EAAE,OAAO,CAAC;IACjB,sCAAsC;IACtC,UAAU,EAAE,MAAM,CAAC;IACnB,0CAA0C;IAC1C,SAAS,EAAE,MAAM,CAAC;IAClB;;;;;;;OAOG;IACH,QAAQ,CAAC,EAAE,MAAM,CAAC;IAClB,6DAA6D;IAC7D,MAAM,CAAC,EAAE,QAAQ,GAAG,KAAK,GAAG,WAAW,CAAC;CACxC;AA4BD,qBAAa,eAAe;IAC3B,OAAO,CAAC,IAAI,CAAiC;IAC7C,OAAO,CAAC,KAAK,CAAiB;IAC9B,OAAO,CAAC,MAAM,CAAM;IACpB,OAAO,CAAC,MAAM,CAAS;IACvB,OAAO,CAAC,YAAY,CAAgB;IAEpC,YAAY,IAAI,EAAE,sBAAsB,EAgCvC;IAED,6DAA6D;IAC7D,KAAK,IAAI,OAAO,CAAC,IAAI,CAAC,CAErB;IAED,IAAI,QAAQ,IAAI,OAAO,CAEtB;IAED,OAAO,CAAC,QAAQ;IAqBhB,OAAO,CAAC,IAAI;IAQZ;;;;;;;;;;;;OAYG;IACG,KAAK,CAAC,IAAI,EAAE,MAAM,EAAE,CAAC,SAAK,EAAE,SAAS,GAAE,eAAyB,GAAG,OAAO,CAAC,eAAe,EAAE,CAAC,CAKlG;IAED;;;;;;;;OAQG;IACG,MAAM,CAAC,KAAK,EAAE,MAAM,EAAE,QAAQ,EAAE,sBAAsB,EAAE,EAAE,CAAC,EAAE,MAAM,GAAG,OAAO,CAAC,qBAAqB,EAAE,CAAC,CAG3G;IAED;;;;OAIG;IACG,IAAI,CAAC,KAAK,EAAE,KAAK,CAAC;QAAE,EAAE,EAAE,MAAM,CAAC;QAAC,IAAI,EAAE,MAAM,CAAA;KAAE,CAAC,GAAG,OAAO,CAAC,mBAAmB,CAAC,CAGnF;IAED,qEAAqE;IAC/D,IAAI,IAAI,OAAO,CAAC,mBAAmB,CAAC,CASzC;IAED,yDAAyD;IACnD,MAAM,CAAC,EAAE,EAAE,MAAM,GAAG,OAAO,CAAC,OAAO,CAAC,CAGzC;IAED,uDAAuD;IACjD,OAAO,IAAI,OAAO,CAAC,IAAI,CAAC,CAE7B;IAED,gDAAgD;IAC1C,IAAI,IAAI,OAAO,CAAC,IAAI,CAAC,CAE1B;IAED,+DAA+D;IACzD,KAAK,IAAI,OAAO,CAAC,IAAI,CAAC,CAI3B;CACD","sourcesContent":["/**\n * Client for the `embsearch` stdio daemon (vendored from\n * github.com/kolisachint/embeddingsearchtools ts/client.ts).\n *\n * Spawns `embsearch serve` once and talks newline-delimited JSON over its\n * stdin/stdout. The process stays alive so the model and index load a single\n * time; every query after startup is hot.\n */\n\nimport { type ChildProcessWithoutNullStreams, spawn } from \"child_process\";\n\nexport interface EmbSearchResult {\n\tid: string;\n\tscore: number;\n}\n\nexport interface EmbSearchDaemonInfo {\n\tmodelId: string;\n\tdim: number;\n\tcount: number;\n\t/**\n\t * Whether the open store carries a BM25 index alongside its vectors.\n\t *\n\t * Fixed when the store is created — passing `--hybrid` at a non-hybrid\n\t * store only warns — so this is the only reliable way to find out, and the\n\t * signal that an existing store has to be rebuilt rather than reused.\n\t */\n\thybrid?: boolean;\n\t/**\n\t * Whether this daemon can actually serve `rerank`, or `undefined` from a\n\t * daemon too old to say.\n\t *\n\t * Reported separately from the version because the two came apart:\n\t * released binaries carry the `rerank` op but no longer bundle the ~23 MB\n\t * cross-encoder weights, so a version check alone would advertise a\n\t * reranker that fails on first call.\n\t */\n\trerank?: boolean;\n}\n\n/** One candidate sent for cross-encoder scoring. */\nexport interface EmbSearchRerankPassage {\n\tid: string;\n\ttext: string;\n}\n\n/** A cross-encoder relevance logit. Higher is more relevant, but the scale is\n *  unnormalized and comparable only within one call. */\nexport interface EmbSearchRerankResult {\n\tid: string;\n\tscore: number;\n}\n\nexport interface EmbSearchBulkResult {\n\tinserted: number;\n\tupdated: number;\n}\n\n/** Which daemon-side retriever answers a query (embsearch >= 0.2.0). */\nexport type DaemonRetriever = \"dense\" | \"lexical\" | \"hybrid\";\n\nexport interface EmbSearchClientOptions {\n\t/** Open/create the store with a BM25 lexical index alongside the vectors. */\n\thybrid?: boolean;\n\t/** Path to the `embsearch` binary. */\n\tbinaryPath: string;\n\t/** Store directory passed as `--path`. */\n\tstorePath: string;\n\t/**\n\t * Model directory passed as `--model` (onnx builds only): a dir holding\n\t * `model.onnx`, `tokenizer.json` and `model.json`.\n\t *\n\t * Omitted, the daemon uses the model bundled into the binary. Set, the\n\t * binary no longer determines which model produced a vector — which is why\n\t * the eval harness records `info.model_id` rather than the binary version.\n\t */\n\tmodelDir?: string;\n\t/** Metric for a freshly created store. Default: \"cosine\". */\n\tmetric?: \"cosine\" | \"dot\" | \"euclidean\";\n}\n\ninterface Pending {\n\tresolve: (value: EmbSearchRawResponse) => void;\n\treject: (err: Error) => void;\n}\n\ninterface EmbSearchRawResponse {\n\tok: boolean;\n\terror?: string;\n\tresults?: EmbSearchResult[];\n\tinserted?: boolean;\n\tremoved?: boolean;\n\tcount?: number;\n\tinserted_count?: number;\n\tupdated_count?: number;\n\tmodel_id?: string;\n\tdim?: number;\n\t/** Cross-encoder scores. Separate from `results` because the daemon's\n\t *  logits are not comparable to retrieval scores. */\n\treranked?: EmbSearchRerankResult[];\n\t/** `info` only: whether the open store has a BM25 index. */\n\thybrid?: boolean;\n\t/** `info` only: whether `rerank` will work. Absent on daemons too old to\n\t *  report it. */\n\trerank?: boolean;\n}\n\nexport class EmbSearchClient {\n\tprivate proc: ChildProcessWithoutNullStreams;\n\tprivate queue: Pending[] = [];\n\tprivate buffer = \"\";\n\tprivate closed = false;\n\tprivate readyPromise: Promise<void>;\n\n\tconstructor(opts: EmbSearchClientOptions) {\n\t\tconst args = [\"serve\", \"--path\", opts.storePath];\n\t\tif (opts.metric) args.push(\"--metric\", opts.metric);\n\t\tif (opts.modelDir) args.push(\"--model\", opts.modelDir);\n\t\t// Hybrid-ness is fixed when a store is created: passing --hybrid against\n\t\t// an existing non-hybrid store warns and is ignored daemon-side, so a\n\t\t// hybrid store needs its own directory.\n\t\tif (opts.hybrid) args.push(\"--hybrid\");\n\n\t\tthis.proc = spawn(opts.binaryPath, args, { stdio: [\"pipe\", \"pipe\", \"pipe\"] });\n\n\t\tthis.proc.stdout.setEncoding(\"utf8\");\n\t\tthis.proc.stdout.on(\"data\", (chunk: string) => this.onStdout(chunk));\n\t\t// Swallow the informational stderr banner; errors surface via responses/exit.\n\t\tthis.proc.stderr.resume();\n\n\t\t// Readiness = the daemon answering a ping (probes the actual request loop).\n\t\tthis.readyPromise = this.send({ op: \"ping\" }).then(() => undefined);\n\t\t// Spawn failures reject the pending ping via the exit handler; mark the\n\t\t// promise handled so a failed spawn doesn't raise an unhandled rejection\n\t\t// before ready() is awaited.\n\t\tthis.readyPromise.catch(() => {});\n\n\t\tthis.proc.on(\"error\", (err) => {\n\t\t\tthis.closed = true;\n\t\t\tfor (const p of this.queue.splice(0)) p.reject(err);\n\t\t});\n\t\tthis.proc.on(\"exit\", (code) => {\n\t\t\tthis.closed = true;\n\t\t\tconst err = new Error(`embsearch daemon exited (code ${code})`);\n\t\t\tfor (const p of this.queue.splice(0)) p.reject(err);\n\t\t});\n\t}\n\n\t/** Resolves once the daemon has loaded the model + index. */\n\tready(): Promise<void> {\n\t\treturn this.readyPromise;\n\t}\n\n\tget isClosed(): boolean {\n\t\treturn this.closed;\n\t}\n\n\tprivate onStdout(chunk: string): void {\n\t\tthis.buffer += chunk;\n\t\t// Each response is one line; dispatch FIFO against the pending queue.\n\t\tfor (let nl = this.buffer.indexOf(\"\\n\"); nl !== -1; nl = this.buffer.indexOf(\"\\n\")) {\n\t\t\tconst line = this.buffer.slice(0, nl).trim();\n\t\t\tthis.buffer = this.buffer.slice(nl + 1);\n\t\t\tif (!line) continue;\n\t\t\tconst pending = this.queue.shift();\n\t\t\tif (!pending) continue;\n\t\t\tlet msg: EmbSearchRawResponse;\n\t\t\ttry {\n\t\t\t\tmsg = JSON.parse(line) as EmbSearchRawResponse;\n\t\t\t} catch {\n\t\t\t\tpending.reject(new Error(`bad response: ${line}`));\n\t\t\t\tcontinue;\n\t\t\t}\n\t\t\tif (msg.ok) pending.resolve(msg);\n\t\t\telse pending.reject(new Error(msg.error ?? \"unknown error\"));\n\t\t}\n\t}\n\n\tprivate send(req: Record<string, unknown>): Promise<EmbSearchRawResponse> {\n\t\tif (this.closed) return Promise.reject(new Error(\"embsearch client is closed\"));\n\t\treturn new Promise<EmbSearchRawResponse>((resolve, reject) => {\n\t\t\tthis.queue.push({ resolve, reject });\n\t\t\tthis.proc.stdin.write(`${JSON.stringify(req)}\\n`);\n\t\t});\n\t}\n\n\t/**\n\t * Search for the top-`k` matches for `text`.\n\t *\n\t * `retriever` selects which leg answers:\n\t *  - `dense` (default) — vector search, the historical behaviour;\n\t *  - `lexical` — BM25 only, raw scores, no embedding computed;\n\t *  - `hybrid` — both, pre-fused by the daemon's own RRF constant.\n\t *\n\t * `lexical` and `hybrid` need a store created with `--hybrid`. Prefer\n\t * `lexical` over `hybrid` when fusing here: `hybrid` collapses both legs\n\t * into one RRF score, discarding the per-retriever ranks the trace records\n\t * and preventing n-way fusion with the grep leg.\n\t */\n\tasync query(text: string, k = 10, retriever: DaemonRetriever = \"dense\"): Promise<EmbSearchResult[]> {\n\t\tconst res = await this.send(\n\t\t\tretriever === \"dense\" ? { op: \"query\", text, k } : { op: \"query\", text, k, retriever },\n\t\t);\n\t\treturn res.results ?? [];\n\t}\n\n\t/**\n\t * Score `passages` against `query` with the daemon's cross-encoder and\n\t * return the best `k`, best first.\n\t *\n\t * Passages are sent inline rather than referenced by id: the caller has the\n\t * exact spans it intends to show the model, and a cross-encoder scores the\n\t * text it is given, so sending anything else would score the wrong thing.\n\t * Requires an onnx build with reranker weights (embsearch >= 0.3.0).\n\t */\n\tasync rerank(query: string, passages: EmbSearchRerankPassage[], k: number): Promise<EmbSearchRerankResult[]> {\n\t\tconst res = await this.send({ op: \"rerank\", query, passages, k });\n\t\treturn res.reranked ?? [];\n\t}\n\n\t/**\n\t * Batched insert-or-replace. One embedding inference for the whole batch —\n\t * the fast path for bulk indexing. Keep batches modest (e.g. 32–64) so a\n\t * concurrent query is not stuck behind a huge inference.\n\t */\n\tasync bulk(items: Array<{ id: string; text: string }>): Promise<EmbSearchBulkResult> {\n\t\tconst res = await this.send({ op: \"bulk\", items });\n\t\treturn { inserted: res.inserted_count ?? 0, updated: res.updated_count ?? 0 };\n\t}\n\n\t/** Model id, dimensionality, and live vector count of the daemon. */\n\tasync info(): Promise<EmbSearchDaemonInfo> {\n\t\tconst res = await this.send({ op: \"info\" });\n\t\treturn {\n\t\t\tmodelId: res.model_id ?? \"\",\n\t\t\tdim: res.dim ?? 0,\n\t\t\tcount: res.count ?? 0,\n\t\t\thybrid: res.hybrid,\n\t\t\trerank: res.rerank,\n\t\t};\n\t}\n\n\t/** Remove a record. Resolves to `true` if it existed. */\n\tasync remove(id: string): Promise<boolean> {\n\t\tconst res = await this.send({ op: \"remove\", id });\n\t\treturn res.removed === true;\n\t}\n\n\t/** Reclaim tombstoned rows left behind by `remove`. */\n\tasync compact(): Promise<void> {\n\t\tawait this.send({ op: \"compact\" });\n\t}\n\n\t/** Persist the index to the store directory. */\n\tasync save(): Promise<void> {\n\t\tawait this.send({ op: \"save\" });\n\t}\n\n\t/** Shut the daemon down, closing stdin so it exits cleanly. */\n\tasync close(): Promise<void> {\n\t\tif (this.closed) return;\n\t\tthis.proc.stdin.end();\n\t\tawait new Promise<void>((resolve) => this.proc.on(\"exit\", () => resolve()));\n\t}\n}\n"]}

package/dist/core/embsearch/client.js

@@ -17,6 +17,8 @@ export class EmbSearchClient {
         const args = ["serve", "--path", opts.storePath];
         if (opts.metric)
             args.push("--metric", opts.metric);
+        if (opts.modelDir)
+            args.push("--model", opts.modelDir);
         // Hybrid-ness is fixed when a store is created: passing --hybrid against
         // an existing non-hybrid store warns and is ignored daemon-side, so a
         // hybrid store needs its own directory.

package/dist/core/embsearch/client.js.map

@@ -1 +1 @@
-{"version":3,"file":"client.js","sourceRoot":"","sources":["../../../src/core/embsearch/client.ts"],"names":[],"mappings":"AAAA;;;;;;;GAOG;AAEH,OAAO,EAAuC,KAAK,EAAE,MAAM,eAAe,CAAC;AAyF3E,MAAM,OAAO,eAAe;IACnB,IAAI,CAAiC;IACrC,KAAK,GAAc,EAAE,CAAC;IACtB,MAAM,GAAG,EAAE,CAAC;IACZ,MAAM,GAAG,KAAK,CAAC;IACf,YAAY,CAAgB;IAEpC,YAAY,IAA4B,EAAE;QACzC,MAAM,IAAI,GAAG,CAAC,OAAO,EAAE,QAAQ,EAAE,IAAI,CAAC,SAAS,CAAC,CAAC;QACjD,IAAI,IAAI,CAAC,MAAM;YAAE,IAAI,CAAC,IAAI,CAAC,UAAU,EAAE,IAAI,CAAC,MAAM,CAAC,CAAC;QACpD,yEAAyE;QACzE,sEAAsE;QACtE,wCAAwC;QACxC,IAAI,IAAI,CAAC,MAAM;YAAE,IAAI,CAAC,IAAI,CAAC,UAAU,CAAC,CAAC;QAEvC,IAAI,CAAC,IAAI,GAAG,KAAK,CAAC,IAAI,CAAC,UAAU,EAAE,IAAI,EAAE,EAAE,KAAK,EAAE,CAAC,MAAM,EAAE,MAAM,EAAE,MAAM,CAAC,EAAE,CAAC,CAAC;QAE9E,IAAI,CAAC,IAAI,CAAC,MAAM,CAAC,WAAW,CAAC,MAAM,CAAC,CAAC;QACrC,IAAI,CAAC,IAAI,CAAC,MAAM,CAAC,EAAE,CAAC,MAAM,EAAE,CAAC,KAAa,EAAE,EAAE,CAAC,IAAI,CAAC,QAAQ,CAAC,KAAK,CAAC,CAAC,CAAC;QACrE,8EAA8E;QAC9E,IAAI,CAAC,IAAI,CAAC,MAAM,CAAC,MAAM,EAAE,CAAC;QAE1B,4EAA4E;QAC5E,IAAI,CAAC,YAAY,GAAG,IAAI,CAAC,IAAI,CAAC,EAAE,EAAE,EAAE,MAAM,EAAE,CAAC,CAAC,IAAI,CAAC,GAAG,EAAE,CAAC,SAAS,CAAC,CAAC;QACpE,wEAAwE;QACxE,yEAAyE;QACzE,6BAA6B;QAC7B,IAAI,CAAC,YAAY,CAAC,KAAK,CAAC,GAAG,EAAE,CAAC,EAAC,CAAC,CAAC,CAAC;QAElC,IAAI,CAAC,IAAI,CAAC,EAAE,CAAC,OAAO,EAAE,CAAC,GAAG,EAAE,EAAE,CAAC;YAC9B,IAAI,CAAC,MAAM,GAAG,IAAI,CAAC;YACnB,KAAK,MAAM,CAAC,IAAI,IAAI,CAAC,KAAK,CAAC,MAAM,CAAC,CAAC,CAAC;gBAAE,CAAC,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC;QAAA,CACpD,CAAC,CAAC;QACH,IAAI,CAAC,IAAI,CAAC,EAAE,CAAC,MAAM,EAAE,CAAC,IAAI,EAAE,EAAE,CAAC;YAC9B,IAAI,CAAC,MAAM,GAAG,IAAI,CAAC;YACnB,MAAM,GAAG,GAAG,IAAI,KAAK,CAAC,iCAAiC,IAAI,GAAG,CAAC,CAAC;YAChE,KAAK,MAAM,CAAC,IAAI,IAAI,CAAC,KAAK,CAAC,MAAM,CAAC,CAAC,CAAC;gBAAE,CAAC,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC;QAAA,CACpD,CAAC,CAAC;IAAA,CACH;IAED,6DAA6D;IAC7D,KAAK,GAAkB;QACtB,OAAO,IAAI,CAAC,YAAY,CAAC;IAAA,CACzB;IAED,IAAI,QAAQ,GAAY;QACvB,OAAO,IAAI,CAAC,MAAM,CAAC;IAAA,CACnB;IAEO,QAAQ,CAAC,KAAa,EAAQ;QACrC,IAAI,CAAC,MAAM,IAAI,KAAK,CAAC;QACrB,sEAAsE;QACtE,KAAK,IAAI,EAAE,GAAG,IAAI,CAAC,MAAM,CAAC,OAAO,CAAC,IAAI,CAAC,EAAE,EAAE,KAAK,CAAC,CAAC,EAAE,EAAE,GAAG,IAAI,CAAC,MAAM,CAAC,OAAO,CAAC,IAAI,CAAC,EAAE,CAAC;YACpF,MAAM,IAAI,GAAG,IAAI,CAAC,MAAM,CAAC,KAAK,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,IAAI,EAAE,CAAC;YAC7C,IAAI,CAAC,MAAM,GAAG,IAAI,CAAC,MAAM,CAAC,KAAK,CAAC,EAAE,GAAG,CAAC,CAAC,CAAC;YACxC,IAAI,CAAC,IAAI;gBAAE,SAAS;YACpB,MAAM,OAAO,GAAG,IAAI,CAAC,KAAK,CAAC,KAAK,EAAE,CAAC;YACnC,IAAI,CAAC,OAAO;gBAAE,SAAS;YACvB,IAAI,GAAyB,CAAC;YAC9B,IAAI,CAAC;gBACJ,GAAG,GAAG,IAAI,CAAC,KAAK,CAAC,IAAI,CAAyB,CAAC;YAChD,CAAC;YAAC,MAAM,CAAC;gBACR,OAAO,CAAC,MAAM,CAAC,IAAI,KAAK,CAAC,iBAAiB,IAAI,EAAE,CAAC,CAAC,CAAC;gBACnD,SAAS;YACV,CAAC;YACD,IAAI,GAAG,CAAC,EAAE;gBAAE,OAAO,CAAC,OAAO,CAAC,GAAG,CAAC,CAAC;;gBAC5B,OAAO,CAAC,MAAM,CAAC,IAAI,KAAK,CAAC,GAAG,CAAC,KAAK,IAAI,eAAe,CAAC,CAAC,CAAC;QAC9D,CAAC;IAAA,CACD;IAEO,IAAI,CAAC,GAA4B,EAAiC;QACzE,IAAI,IAAI,CAAC,MAAM;YAAE,OAAO,OAAO,CAAC,MAAM,CAAC,IAAI,KAAK,CAAC,4BAA4B,CAAC,CAAC,CAAC;QAChF,OAAO,IAAI,OAAO,CAAuB,CAAC,OAAO,EAAE,MAAM,EAAE,EAAE,CAAC;YAC7D,IAAI,CAAC,KAAK,CAAC,IAAI,CAAC,EAAE,OAAO,EAAE,MAAM,EAAE,CAAC,CAAC;YACrC,IAAI,CAAC,IAAI,CAAC,KAAK,CAAC,KAAK,CAAC,GAAG,IAAI,CAAC,SAAS,CAAC,GAAG,CAAC,IAAI,CAAC,CAAC;QAAA,CAClD,CAAC,CAAC;IAAA,CACH;IAED;;;;;;;;;;;;OAYG;IACH,KAAK,CAAC,KAAK,CAAC,IAAY,EAAE,CAAC,GAAG,EAAE,EAAE,SAAS,GAAoB,OAAO,EAA8B;QACnG,MAAM,GAAG,GAAG,MAAM,IAAI,CAAC,IAAI,CAC1B,SAAS,KAAK,OAAO,CAAC,CAAC,CAAC,EAAE,EAAE,EAAE,OAAO,EAAE,IAAI,EAAE,CAAC,EAAE,CAAC,CAAC,CAAC,EAAE,EAAE,EAAE,OAAO,EAAE,IAAI,EAAE,CAAC,EAAE,SAAS,EAAE,CACtF,CAAC;QACF,OAAO,GAAG,CAAC,OAAO,IAAI,EAAE,CAAC;IAAA,CACzB;IAED;;;;;;;;OAQG;IACH,KAAK,CAAC,MAAM,CAAC,KAAa,EAAE,QAAkC,EAAE,CAAS,EAAoC;QAC5G,MAAM,GAAG,GAAG,MAAM,IAAI,CAAC,IAAI,CAAC,EAAE,EAAE,EAAE,QAAQ,EAAE,KAAK,EAAE,QAAQ,EAAE,CAAC,EAAE,CAAC,CAAC;QAClE,OAAO,GAAG,CAAC,QAAQ,IAAI,EAAE,CAAC;IAAA,CAC1B;IAED;;;;OAIG;IACH,KAAK,CAAC,IAAI,CAAC,KAA0C,EAAgC;QACpF,MAAM,GAAG,GAAG,MAAM,IAAI,CAAC,IAAI,CAAC,EAAE,EAAE,EAAE,MAAM,EAAE,KAAK,EAAE,CAAC,CAAC;QACnD,OAAO,EAAE,QAAQ,EAAE,GAAG,CAAC,cAAc,IAAI,CAAC,EAAE,OAAO,EAAE,GAAG,CAAC,aAAa,IAAI,CAAC,EAAE,CAAC;IAAA,CAC9E;IAED,qEAAqE;IACrE,KAAK,CAAC,IAAI,GAAiC;QAC1C,MAAM,GAAG,GAAG,MAAM,IAAI,CAAC,IAAI,CAAC,EAAE,EAAE,EAAE,MAAM,EAAE,CAAC,CAAC;QAC5C,OAAO;YACN,OAAO,EAAE,GAAG,CAAC,QAAQ,IAAI,EAAE;YAC3B,GAAG,EAAE,GAAG,CAAC,GAAG,IAAI,CAAC;YACjB,KAAK,EAAE,GAAG,CAAC,KAAK,IAAI,CAAC;YACrB,MAAM,EAAE,GAAG,CAAC,MAAM;YAClB,MAAM,EAAE,GAAG,CAAC,MAAM;SAClB,CAAC;IAAA,CACF;IAED,yDAAyD;IACzD,KAAK,CAAC,MAAM,CAAC,EAAU,EAAoB;QAC1C,MAAM,GAAG,GAAG,MAAM,IAAI,CAAC,IAAI,CAAC,EAAE,EAAE,EAAE,QAAQ,EAAE,EAAE,EAAE,CAAC,CAAC;QAClD,OAAO,GAAG,CAAC,OAAO,KAAK,IAAI,CAAC;IAAA,CAC5B;IAED,uDAAuD;IACvD,KAAK,CAAC,OAAO,GAAkB;QAC9B,MAAM,IAAI,CAAC,IAAI,CAAC,EAAE,EAAE,EAAE,SAAS,EAAE,CAAC,CAAC;IAAA,CACnC;IAED,gDAAgD;IAChD,KAAK,CAAC,IAAI,GAAkB;QAC3B,MAAM,IAAI,CAAC,IAAI,CAAC,EAAE,EAAE,EAAE,MAAM,EAAE,CAAC,CAAC;IAAA,CAChC;IAED,+DAA+D;IAC/D,KAAK,CAAC,KAAK,GAAkB;QAC5B,IAAI,IAAI,CAAC,MAAM;YAAE,OAAO;QACxB,IAAI,CAAC,IAAI,CAAC,KAAK,CAAC,GAAG,EAAE,CAAC;QACtB,MAAM,IAAI,OAAO,CAAO,CAAC,OAAO,EAAE,EAAE,CAAC,IAAI,CAAC,IAAI,CAAC,EAAE,CAAC,MAAM,EAAE,GAAG,EAAE,CAAC,OAAO,EAAE,CAAC,CAAC,CAAC;IAAA,CAC5E;CACD","sourcesContent":["/**\n * Client for the `embsearch` stdio daemon (vendored from\n * github.com/kolisachint/embeddingsearchtools ts/client.ts).\n *\n * Spawns `embsearch serve` once and talks newline-delimited JSON over its\n * stdin/stdout. The process stays alive so the model and index load a single\n * time; every query after startup is hot.\n */\n\nimport { type ChildProcessWithoutNullStreams, spawn } from \"child_process\";\n\nexport interface EmbSearchResult {\n\tid: string;\n\tscore: number;\n}\n\nexport interface EmbSearchDaemonInfo {\n\tmodelId: string;\n\tdim: number;\n\tcount: number;\n\t/**\n\t * Whether the open store carries a BM25 index alongside its vectors.\n\t *\n\t * Fixed when the store is created — passing `--hybrid` at a non-hybrid\n\t * store only warns — so this is the only reliable way to find out, and the\n\t * signal that an existing store has to be rebuilt rather than reused.\n\t */\n\thybrid?: boolean;\n\t/**\n\t * Whether this daemon can actually serve `rerank`, or `undefined` from a\n\t * daemon too old to say.\n\t *\n\t * Reported separately from the version because the two came apart:\n\t * released binaries carry the `rerank` op but no longer bundle the ~23 MB\n\t * cross-encoder weights, so a version check alone would advertise a\n\t * reranker that fails on first call.\n\t */\n\trerank?: boolean;\n}\n\n/** One candidate sent for cross-encoder scoring. */\nexport interface EmbSearchRerankPassage {\n\tid: string;\n\ttext: string;\n}\n\n/** A cross-encoder relevance logit. Higher is more relevant, but the scale is\n *  unnormalized and comparable only within one call. */\nexport interface EmbSearchRerankResult {\n\tid: string;\n\tscore: number;\n}\n\nexport interface EmbSearchBulkResult {\n\tinserted: number;\n\tupdated: number;\n}\n\n/** Which daemon-side retriever answers a query (embsearch >= 0.2.0). */\nexport type DaemonRetriever = \"dense\" | \"lexical\" | \"hybrid\";\n\nexport interface EmbSearchClientOptions {\n\t/** Open/create the store with a BM25 lexical index alongside the vectors. */\n\thybrid?: boolean;\n\t/** Path to the `embsearch` binary. */\n\tbinaryPath: string;\n\t/** Store directory passed as `--path`. */\n\tstorePath: string;\n\t/** Metric for a freshly created store. Default: \"cosine\". */\n\tmetric?: \"cosine\" | \"dot\" | \"euclidean\";\n}\n\ninterface Pending {\n\tresolve: (value: EmbSearchRawResponse) => void;\n\treject: (err: Error) => void;\n}\n\ninterface EmbSearchRawResponse {\n\tok: boolean;\n\terror?: string;\n\tresults?: EmbSearchResult[];\n\tinserted?: boolean;\n\tremoved?: boolean;\n\tcount?: number;\n\tinserted_count?: number;\n\tupdated_count?: number;\n\tmodel_id?: string;\n\tdim?: number;\n\t/** Cross-encoder scores. Separate from `results` because the daemon's\n\t *  logits are not comparable to retrieval scores. */\n\treranked?: EmbSearchRerankResult[];\n\t/** `info` only: whether the open store has a BM25 index. */\n\thybrid?: boolean;\n\t/** `info` only: whether `rerank` will work. Absent on daemons too old to\n\t *  report it. */\n\trerank?: boolean;\n}\n\nexport class EmbSearchClient {\n\tprivate proc: ChildProcessWithoutNullStreams;\n\tprivate queue: Pending[] = [];\n\tprivate buffer = \"\";\n\tprivate closed = false;\n\tprivate readyPromise: Promise<void>;\n\n\tconstructor(opts: EmbSearchClientOptions) {\n\t\tconst args = [\"serve\", \"--path\", opts.storePath];\n\t\tif (opts.metric) args.push(\"--metric\", opts.metric);\n\t\t// Hybrid-ness is fixed when a store is created: passing --hybrid against\n\t\t// an existing non-hybrid store warns and is ignored daemon-side, so a\n\t\t// hybrid store needs its own directory.\n\t\tif (opts.hybrid) args.push(\"--hybrid\");\n\n\t\tthis.proc = spawn(opts.binaryPath, args, { stdio: [\"pipe\", \"pipe\", \"pipe\"] });\n\n\t\tthis.proc.stdout.setEncoding(\"utf8\");\n\t\tthis.proc.stdout.on(\"data\", (chunk: string) => this.onStdout(chunk));\n\t\t// Swallow the informational stderr banner; errors surface via responses/exit.\n\t\tthis.proc.stderr.resume();\n\n\t\t// Readiness = the daemon answering a ping (probes the actual request loop).\n\t\tthis.readyPromise = this.send({ op: \"ping\" }).then(() => undefined);\n\t\t// Spawn failures reject the pending ping via the exit handler; mark the\n\t\t// promise handled so a failed spawn doesn't raise an unhandled rejection\n\t\t// before ready() is awaited.\n\t\tthis.readyPromise.catch(() => {});\n\n\t\tthis.proc.on(\"error\", (err) => {\n\t\t\tthis.closed = true;\n\t\t\tfor (const p of this.queue.splice(0)) p.reject(err);\n\t\t});\n\t\tthis.proc.on(\"exit\", (code) => {\n\t\t\tthis.closed = true;\n\t\t\tconst err = new Error(`embsearch daemon exited (code ${code})`);\n\t\t\tfor (const p of this.queue.splice(0)) p.reject(err);\n\t\t});\n\t}\n\n\t/** Resolves once the daemon has loaded the model + index. */\n\tready(): Promise<void> {\n\t\treturn this.readyPromise;\n\t}\n\n\tget isClosed(): boolean {\n\t\treturn this.closed;\n\t}\n\n\tprivate onStdout(chunk: string): void {\n\t\tthis.buffer += chunk;\n\t\t// Each response is one line; dispatch FIFO against the pending queue.\n\t\tfor (let nl = this.buffer.indexOf(\"\\n\"); nl !== -1; nl = this.buffer.indexOf(\"\\n\")) {\n\t\t\tconst line = this.buffer.slice(0, nl).trim();\n\t\t\tthis.buffer = this.buffer.slice(nl + 1);\n\t\t\tif (!line) continue;\n\t\t\tconst pending = this.queue.shift();\n\t\t\tif (!pending) continue;\n\t\t\tlet msg: EmbSearchRawResponse;\n\t\t\ttry {\n\t\t\t\tmsg = JSON.parse(line) as EmbSearchRawResponse;\n\t\t\t} catch {\n\t\t\t\tpending.reject(new Error(`bad response: ${line}`));\n\t\t\t\tcontinue;\n\t\t\t}\n\t\t\tif (msg.ok) pending.resolve(msg);\n\t\t\telse pending.reject(new Error(msg.error ?? \"unknown error\"));\n\t\t}\n\t}\n\n\tprivate send(req: Record<string, unknown>): Promise<EmbSearchRawResponse> {\n\t\tif (this.closed) return Promise.reject(new Error(\"embsearch client is closed\"));\n\t\treturn new Promise<EmbSearchRawResponse>((resolve, reject) => {\n\t\t\tthis.queue.push({ resolve, reject });\n\t\t\tthis.proc.stdin.write(`${JSON.stringify(req)}\\n`);\n\t\t});\n\t}\n\n\t/**\n\t * Search for the top-`k` matches for `text`.\n\t *\n\t * `retriever` selects which leg answers:\n\t *  - `dense` (default) — vector search, the historical behaviour;\n\t *  - `lexical` — BM25 only, raw scores, no embedding computed;\n\t *  - `hybrid` — both, pre-fused by the daemon's own RRF constant.\n\t *\n\t * `lexical` and `hybrid` need a store created with `--hybrid`. Prefer\n\t * `lexical` over `hybrid` when fusing here: `hybrid` collapses both legs\n\t * into one RRF score, discarding the per-retriever ranks the trace records\n\t * and preventing n-way fusion with the grep leg.\n\t */\n\tasync query(text: string, k = 10, retriever: DaemonRetriever = \"dense\"): Promise<EmbSearchResult[]> {\n\t\tconst res = await this.send(\n\t\t\tretriever === \"dense\" ? { op: \"query\", text, k } : { op: \"query\", text, k, retriever },\n\t\t);\n\t\treturn res.results ?? [];\n\t}\n\n\t/**\n\t * Score `passages` against `query` with the daemon's cross-encoder and\n\t * return the best `k`, best first.\n\t *\n\t * Passages are sent inline rather than referenced by id: the caller has the\n\t * exact spans it intends to show the model, and a cross-encoder scores the\n\t * text it is given, so sending anything else would score the wrong thing.\n\t * Requires an onnx build with reranker weights (embsearch >= 0.3.0).\n\t */\n\tasync rerank(query: string, passages: EmbSearchRerankPassage[], k: number): Promise<EmbSearchRerankResult[]> {\n\t\tconst res = await this.send({ op: \"rerank\", query, passages, k });\n\t\treturn res.reranked ?? [];\n\t}\n\n\t/**\n\t * Batched insert-or-replace. One embedding inference for the whole batch —\n\t * the fast path for bulk indexing. Keep batches modest (e.g. 32–64) so a\n\t * concurrent query is not stuck behind a huge inference.\n\t */\n\tasync bulk(items: Array<{ id: string; text: string }>): Promise<EmbSearchBulkResult> {\n\t\tconst res = await this.send({ op: \"bulk\", items });\n\t\treturn { inserted: res.inserted_count ?? 0, updated: res.updated_count ?? 0 };\n\t}\n\n\t/** Model id, dimensionality, and live vector count of the daemon. */\n\tasync info(): Promise<EmbSearchDaemonInfo> {\n\t\tconst res = await this.send({ op: \"info\" });\n\t\treturn {\n\t\t\tmodelId: res.model_id ?? \"\",\n\t\t\tdim: res.dim ?? 0,\n\t\t\tcount: res.count ?? 0,\n\t\t\thybrid: res.hybrid,\n\t\t\trerank: res.rerank,\n\t\t};\n\t}\n\n\t/** Remove a record. Resolves to `true` if it existed. */\n\tasync remove(id: string): Promise<boolean> {\n\t\tconst res = await this.send({ op: \"remove\", id });\n\t\treturn res.removed === true;\n\t}\n\n\t/** Reclaim tombstoned rows left behind by `remove`. */\n\tasync compact(): Promise<void> {\n\t\tawait this.send({ op: \"compact\" });\n\t}\n\n\t/** Persist the index to the store directory. */\n\tasync save(): Promise<void> {\n\t\tawait this.send({ op: \"save\" });\n\t}\n\n\t/** Shut the daemon down, closing stdin so it exits cleanly. */\n\tasync close(): Promise<void> {\n\t\tif (this.closed) return;\n\t\tthis.proc.stdin.end();\n\t\tawait new Promise<void>((resolve) => this.proc.on(\"exit\", () => resolve()));\n\t}\n}\n"]}
+{"version":3,"file":"client.js","sourceRoot":"","sources":["../../../src/core/embsearch/client.ts"],"names":[],"mappings":"AAAA;;;;;;;GAOG;AAEH,OAAO,EAAuC,KAAK,EAAE,MAAM,eAAe,CAAC;AAkG3E,MAAM,OAAO,eAAe;IACnB,IAAI,CAAiC;IACrC,KAAK,GAAc,EAAE,CAAC;IACtB,MAAM,GAAG,EAAE,CAAC;IACZ,MAAM,GAAG,KAAK,CAAC;IACf,YAAY,CAAgB;IAEpC,YAAY,IAA4B,EAAE;QACzC,MAAM,IAAI,GAAG,CAAC,OAAO,EAAE,QAAQ,EAAE,IAAI,CAAC,SAAS,CAAC,CAAC;QACjD,IAAI,IAAI,CAAC,MAAM;YAAE,IAAI,CAAC,IAAI,CAAC,UAAU,EAAE,IAAI,CAAC,MAAM,CAAC,CAAC;QACpD,IAAI,IAAI,CAAC,QAAQ;YAAE,IAAI,CAAC,IAAI,CAAC,SAAS,EAAE,IAAI,CAAC,QAAQ,CAAC,CAAC;QACvD,yEAAyE;QACzE,sEAAsE;QACtE,wCAAwC;QACxC,IAAI,IAAI,CAAC,MAAM;YAAE,IAAI,CAAC,IAAI,CAAC,UAAU,CAAC,CAAC;QAEvC,IAAI,CAAC,IAAI,GAAG,KAAK,CAAC,IAAI,CAAC,UAAU,EAAE,IAAI,EAAE,EAAE,KAAK,EAAE,CAAC,MAAM,EAAE,MAAM,EAAE,MAAM,CAAC,EAAE,CAAC,CAAC;QAE9E,IAAI,CAAC,IAAI,CAAC,MAAM,CAAC,WAAW,CAAC,MAAM,CAAC,CAAC;QACrC,IAAI,CAAC,IAAI,CAAC,MAAM,CAAC,EAAE,CAAC,MAAM,EAAE,CAAC,KAAa,EAAE,EAAE,CAAC,IAAI,CAAC,QAAQ,CAAC,KAAK,CAAC,CAAC,CAAC;QACrE,8EAA8E;QAC9E,IAAI,CAAC,IAAI,CAAC,MAAM,CAAC,MAAM,EAAE,CAAC;QAE1B,4EAA4E;QAC5E,IAAI,CAAC,YAAY,GAAG,IAAI,CAAC,IAAI,CAAC,EAAE,EAAE,EAAE,MAAM,EAAE,CAAC,CAAC,IAAI,CAAC,GAAG,EAAE,CAAC,SAAS,CAAC,CAAC;QACpE,wEAAwE;QACxE,yEAAyE;QACzE,6BAA6B;QAC7B,IAAI,CAAC,YAAY,CAAC,KAAK,CAAC,GAAG,EAAE,CAAC,EAAC,CAAC,CAAC,CAAC;QAElC,IAAI,CAAC,IAAI,CAAC,EAAE,CAAC,OAAO,EAAE,CAAC,GAAG,EAAE,EAAE,CAAC;YAC9B,IAAI,CAAC,MAAM,GAAG,IAAI,CAAC;YACnB,KAAK,MAAM,CAAC,IAAI,IAAI,CAAC,KAAK,CAAC,MAAM,CAAC,CAAC,CAAC;gBAAE,CAAC,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC;QAAA,CACpD,CAAC,CAAC;QACH,IAAI,CAAC,IAAI,CAAC,EAAE,CAAC,MAAM,EAAE,CAAC,IAAI,EAAE,EAAE,CAAC;YAC9B,IAAI,CAAC,MAAM,GAAG,IAAI,CAAC;YACnB,MAAM,GAAG,GAAG,IAAI,KAAK,CAAC,iCAAiC,IAAI,GAAG,CAAC,CAAC;YAChE,KAAK,MAAM,CAAC,IAAI,IAAI,CAAC,KAAK,CAAC,MAAM,CAAC,CAAC,CAAC;gBAAE,CAAC,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC;QAAA,CACpD,CAAC,CAAC;IAAA,CACH;IAED,6DAA6D;IAC7D,KAAK,GAAkB;QACtB,OAAO,IAAI,CAAC,YAAY,CAAC;IAAA,CACzB;IAED,IAAI,QAAQ,GAAY;QACvB,OAAO,IAAI,CAAC,MAAM,CAAC;IAAA,CACnB;IAEO,QAAQ,CAAC,KAAa,EAAQ;QACrC,IAAI,CAAC,MAAM,IAAI,KAAK,CAAC;QACrB,sEAAsE;QACtE,KAAK,IAAI,EAAE,GAAG,IAAI,CAAC,MAAM,CAAC,OAAO,CAAC,IAAI,CAAC,EAAE,EAAE,KAAK,CAAC,CAAC,EAAE,EAAE,GAAG,IAAI,CAAC,MAAM,CAAC,OAAO,CAAC,IAAI,CAAC,EAAE,CAAC;YACpF,MAAM,IAAI,GAAG,IAAI,CAAC,MAAM,CAAC,KAAK,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,IAAI,EAAE,CAAC;YAC7C,IAAI,CAAC,MAAM,GAAG,IAAI,CAAC,MAAM,CAAC,KAAK,CAAC,EAAE,GAAG,CAAC,CAAC,CAAC;YACxC,IAAI,CAAC,IAAI;gBAAE,SAAS;YACpB,MAAM,OAAO,GAAG,IAAI,CAAC,KAAK,CAAC,KAAK,EAAE,CAAC;YACnC,IAAI,CAAC,OAAO;gBAAE,SAAS;YACvB,IAAI,GAAyB,CAAC;YAC9B,IAAI,CAAC;gBACJ,GAAG,GAAG,IAAI,CAAC,KAAK,CAAC,IAAI,CAAyB,CAAC;YAChD,CAAC;YAAC,MAAM,CAAC;gBACR,OAAO,CAAC,MAAM,CAAC,IAAI,KAAK,CAAC,iBAAiB,IAAI,EAAE,CAAC,CAAC,CAAC;gBACnD,SAAS;YACV,CAAC;YACD,IAAI,GAAG,CAAC,EAAE;gBAAE,OAAO,CAAC,OAAO,CAAC,GAAG,CAAC,CAAC;;gBAC5B,OAAO,CAAC,MAAM,CAAC,IAAI,KAAK,CAAC,GAAG,CAAC,KAAK,IAAI,eAAe,CAAC,CAAC,CAAC;QAC9D,CAAC;IAAA,CACD;IAEO,IAAI,CAAC,GAA4B,EAAiC;QACzE,IAAI,IAAI,CAAC,MAAM;YAAE,OAAO,OAAO,CAAC,MAAM,CAAC,IAAI,KAAK,CAAC,4BAA4B,CAAC,CAAC,CAAC;QAChF,OAAO,IAAI,OAAO,CAAuB,CAAC,OAAO,EAAE,MAAM,EAAE,EAAE,CAAC;YAC7D,IAAI,CAAC,KAAK,CAAC,IAAI,CAAC,EAAE,OAAO,EAAE,MAAM,EAAE,CAAC,CAAC;YACrC,IAAI,CAAC,IAAI,CAAC,KAAK,CAAC,KAAK,CAAC,GAAG,IAAI,CAAC,SAAS,CAAC,GAAG,CAAC,IAAI,CAAC,CAAC;QAAA,CAClD,CAAC,CAAC;IAAA,CACH;IAED;;;;;;;;;;;;OAYG;IACH,KAAK,CAAC,KAAK,CAAC,IAAY,EAAE,CAAC,GAAG,EAAE,EAAE,SAAS,GAAoB,OAAO,EAA8B;QACnG,MAAM,GAAG,GAAG,MAAM,IAAI,CAAC,IAAI,CAC1B,SAAS,KAAK,OAAO,CAAC,CAAC,CAAC,EAAE,EAAE,EAAE,OAAO,EAAE,IAAI,EAAE,CAAC,EAAE,CAAC,CAAC,CAAC,EAAE,EAAE,EAAE,OAAO,EAAE,IAAI,EAAE,CAAC,EAAE,SAAS,EAAE,CACtF,CAAC;QACF,OAAO,GAAG,CAAC,OAAO,IAAI,EAAE,CAAC;IAAA,CACzB;IAED;;;;;;;;OAQG;IACH,KAAK,CAAC,MAAM,CAAC,KAAa,EAAE,QAAkC,EAAE,CAAS,EAAoC;QAC5G,MAAM,GAAG,GAAG,MAAM,IAAI,CAAC,IAAI,CAAC,EAAE,EAAE,EAAE,QAAQ,EAAE,KAAK,EAAE,QAAQ,EAAE,CAAC,EAAE,CAAC,CAAC;QAClE,OAAO,GAAG,CAAC,QAAQ,IAAI,EAAE,CAAC;IAAA,CAC1B;IAED;;;;OAIG;IACH,KAAK,CAAC,IAAI,CAAC,KAA0C,EAAgC;QACpF,MAAM,GAAG,GAAG,MAAM,IAAI,CAAC,IAAI,CAAC,EAAE,EAAE,EAAE,MAAM,EAAE,KAAK,EAAE,CAAC,CAAC;QACnD,OAAO,EAAE,QAAQ,EAAE,GAAG,CAAC,cAAc,IAAI,CAAC,EAAE,OAAO,EAAE,GAAG,CAAC,aAAa,IAAI,CAAC,EAAE,CAAC;IAAA,CAC9E;IAED,qEAAqE;IACrE,KAAK,CAAC,IAAI,GAAiC;QAC1C,MAAM,GAAG,GAAG,MAAM,IAAI,CAAC,IAAI,CAAC,EAAE,EAAE,EAAE,MAAM,EAAE,CAAC,CAAC;QAC5C,OAAO;YACN,OAAO,EAAE,GAAG,CAAC,QAAQ,IAAI,EAAE;YAC3B,GAAG,EAAE,GAAG,CAAC,GAAG,IAAI,CAAC;YACjB,KAAK,EAAE,GAAG,CAAC,KAAK,IAAI,CAAC;YACrB,MAAM,EAAE,GAAG,CAAC,MAAM;YAClB,MAAM,EAAE,GAAG,CAAC,MAAM;SAClB,CAAC;IAAA,CACF;IAED,yDAAyD;IACzD,KAAK,CAAC,MAAM,CAAC,EAAU,EAAoB;QAC1C,MAAM,GAAG,GAAG,MAAM,IAAI,CAAC,IAAI,CAAC,EAAE,EAAE,EAAE,QAAQ,EAAE,EAAE,EAAE,CAAC,CAAC;QAClD,OAAO,GAAG,CAAC,OAAO,KAAK,IAAI,CAAC;IAAA,CAC5B;IAED,uDAAuD;IACvD,KAAK,CAAC,OAAO,GAAkB;QAC9B,MAAM,IAAI,CAAC,IAAI,CAAC,EAAE,EAAE,EAAE,SAAS,EAAE,CAAC,CAAC;IAAA,CACnC;IAED,gDAAgD;IAChD,KAAK,CAAC,IAAI,GAAkB;QAC3B,MAAM,IAAI,CAAC,IAAI,CAAC,EAAE,EAAE,EAAE,MAAM,EAAE,CAAC,CAAC;IAAA,CAChC;IAED,+DAA+D;IAC/D,KAAK,CAAC,KAAK,GAAkB;QAC5B,IAAI,IAAI,CAAC,MAAM;YAAE,OAAO;QACxB,IAAI,CAAC,IAAI,CAAC,KAAK,CAAC,GAAG,EAAE,CAAC;QACtB,MAAM,IAAI,OAAO,CAAO,CAAC,OAAO,EAAE,EAAE,CAAC,IAAI,CAAC,IAAI,CAAC,EAAE,CAAC,MAAM,EAAE,GAAG,EAAE,CAAC,OAAO,EAAE,CAAC,CAAC,CAAC;IAAA,CAC5E;CACD","sourcesContent":["/**\n * Client for the `embsearch` stdio daemon (vendored from\n * github.com/kolisachint/embeddingsearchtools ts/client.ts).\n *\n * Spawns `embsearch serve` once and talks newline-delimited JSON over its\n * stdin/stdout. The process stays alive so the model and index load a single\n * time; every query after startup is hot.\n */\n\nimport { type ChildProcessWithoutNullStreams, spawn } from \"child_process\";\n\nexport interface EmbSearchResult {\n\tid: string;\n\tscore: number;\n}\n\nexport interface EmbSearchDaemonInfo {\n\tmodelId: string;\n\tdim: number;\n\tcount: number;\n\t/**\n\t * Whether the open store carries a BM25 index alongside its vectors.\n\t *\n\t * Fixed when the store is created — passing `--hybrid` at a non-hybrid\n\t * store only warns — so this is the only reliable way to find out, and the\n\t * signal that an existing store has to be rebuilt rather than reused.\n\t */\n\thybrid?: boolean;\n\t/**\n\t * Whether this daemon can actually serve `rerank`, or `undefined` from a\n\t * daemon too old to say.\n\t *\n\t * Reported separately from the version because the two came apart:\n\t * released binaries carry the `rerank` op but no longer bundle the ~23 MB\n\t * cross-encoder weights, so a version check alone would advertise a\n\t * reranker that fails on first call.\n\t */\n\trerank?: boolean;\n}\n\n/** One candidate sent for cross-encoder scoring. */\nexport interface EmbSearchRerankPassage {\n\tid: string;\n\ttext: string;\n}\n\n/** A cross-encoder relevance logit. Higher is more relevant, but the scale is\n *  unnormalized and comparable only within one call. */\nexport interface EmbSearchRerankResult {\n\tid: string;\n\tscore: number;\n}\n\nexport interface EmbSearchBulkResult {\n\tinserted: number;\n\tupdated: number;\n}\n\n/** Which daemon-side retriever answers a query (embsearch >= 0.2.0). */\nexport type DaemonRetriever = \"dense\" | \"lexical\" | \"hybrid\";\n\nexport interface EmbSearchClientOptions {\n\t/** Open/create the store with a BM25 lexical index alongside the vectors. */\n\thybrid?: boolean;\n\t/** Path to the `embsearch` binary. */\n\tbinaryPath: string;\n\t/** Store directory passed as `--path`. */\n\tstorePath: string;\n\t/**\n\t * Model directory passed as `--model` (onnx builds only): a dir holding\n\t * `model.onnx`, `tokenizer.json` and `model.json`.\n\t *\n\t * Omitted, the daemon uses the model bundled into the binary. Set, the\n\t * binary no longer determines which model produced a vector — which is why\n\t * the eval harness records `info.model_id` rather than the binary version.\n\t */\n\tmodelDir?: string;\n\t/** Metric for a freshly created store. Default: \"cosine\". */\n\tmetric?: \"cosine\" | \"dot\" | \"euclidean\";\n}\n\ninterface Pending {\n\tresolve: (value: EmbSearchRawResponse) => void;\n\treject: (err: Error) => void;\n}\n\ninterface EmbSearchRawResponse {\n\tok: boolean;\n\terror?: string;\n\tresults?: EmbSearchResult[];\n\tinserted?: boolean;\n\tremoved?: boolean;\n\tcount?: number;\n\tinserted_count?: number;\n\tupdated_count?: number;\n\tmodel_id?: string;\n\tdim?: number;\n\t/** Cross-encoder scores. Separate from `results` because the daemon's\n\t *  logits are not comparable to retrieval scores. */\n\treranked?: EmbSearchRerankResult[];\n\t/** `info` only: whether the open store has a BM25 index. */\n\thybrid?: boolean;\n\t/** `info` only: whether `rerank` will work. Absent on daemons too old to\n\t *  report it. */\n\trerank?: boolean;\n}\n\nexport class EmbSearchClient {\n\tprivate proc: ChildProcessWithoutNullStreams;\n\tprivate queue: Pending[] = [];\n\tprivate buffer = \"\";\n\tprivate closed = false;\n\tprivate readyPromise: Promise<void>;\n\n\tconstructor(opts: EmbSearchClientOptions) {\n\t\tconst args = [\"serve\", \"--path\", opts.storePath];\n\t\tif (opts.metric) args.push(\"--metric\", opts.metric);\n\t\tif (opts.modelDir) args.push(\"--model\", opts.modelDir);\n\t\t// Hybrid-ness is fixed when a store is created: passing --hybrid against\n\t\t// an existing non-hybrid store warns and is ignored daemon-side, so a\n\t\t// hybrid store needs its own directory.\n\t\tif (opts.hybrid) args.push(\"--hybrid\");\n\n\t\tthis.proc = spawn(opts.binaryPath, args, { stdio: [\"pipe\", \"pipe\", \"pipe\"] });\n\n\t\tthis.proc.stdout.setEncoding(\"utf8\");\n\t\tthis.proc.stdout.on(\"data\", (chunk: string) => this.onStdout(chunk));\n\t\t// Swallow the informational stderr banner; errors surface via responses/exit.\n\t\tthis.proc.stderr.resume();\n\n\t\t// Readiness = the daemon answering a ping (probes the actual request loop).\n\t\tthis.readyPromise = this.send({ op: \"ping\" }).then(() => undefined);\n\t\t// Spawn failures reject the pending ping via the exit handler; mark the\n\t\t// promise handled so a failed spawn doesn't raise an unhandled rejection\n\t\t// before ready() is awaited.\n\t\tthis.readyPromise.catch(() => {});\n\n\t\tthis.proc.on(\"error\", (err) => {\n\t\t\tthis.closed = true;\n\t\t\tfor (const p of this.queue.splice(0)) p.reject(err);\n\t\t});\n\t\tthis.proc.on(\"exit\", (code) => {\n\t\t\tthis.closed = true;\n\t\t\tconst err = new Error(`embsearch daemon exited (code ${code})`);\n\t\t\tfor (const p of this.queue.splice(0)) p.reject(err);\n\t\t});\n\t}\n\n\t/** Resolves once the daemon has loaded the model + index. */\n\tready(): Promise<void> {\n\t\treturn this.readyPromise;\n\t}\n\n\tget isClosed(): boolean {\n\t\treturn this.closed;\n\t}\n\n\tprivate onStdout(chunk: string): void {\n\t\tthis.buffer += chunk;\n\t\t// Each response is one line; dispatch FIFO against the pending queue.\n\t\tfor (let nl = this.buffer.indexOf(\"\\n\"); nl !== -1; nl = this.buffer.indexOf(\"\\n\")) {\n\t\t\tconst line = this.buffer.slice(0, nl).trim();\n\t\t\tthis.buffer = this.buffer.slice(nl + 1);\n\t\t\tif (!line) continue;\n\t\t\tconst pending = this.queue.shift();\n\t\t\tif (!pending) continue;\n\t\t\tlet msg: EmbSearchRawResponse;\n\t\t\ttry {\n\t\t\t\tmsg = JSON.parse(line) as EmbSearchRawResponse;\n\t\t\t} catch {\n\t\t\t\tpending.reject(new Error(`bad response: ${line}`));\n\t\t\t\tcontinue;\n\t\t\t}\n\t\t\tif (msg.ok) pending.resolve(msg);\n\t\t\telse pending.reject(new Error(msg.error ?? \"unknown error\"));\n\t\t}\n\t}\n\n\tprivate send(req: Record<string, unknown>): Promise<EmbSearchRawResponse> {\n\t\tif (this.closed) return Promise.reject(new Error(\"embsearch client is closed\"));\n\t\treturn new Promise<EmbSearchRawResponse>((resolve, reject) => {\n\t\t\tthis.queue.push({ resolve, reject });\n\t\t\tthis.proc.stdin.write(`${JSON.stringify(req)}\\n`);\n\t\t});\n\t}\n\n\t/**\n\t * Search for the top-`k` matches for `text`.\n\t *\n\t * `retriever` selects which leg answers:\n\t *  - `dense` (default) — vector search, the historical behaviour;\n\t *  - `lexical` — BM25 only, raw scores, no embedding computed;\n\t *  - `hybrid` — both, pre-fused by the daemon's own RRF constant.\n\t *\n\t * `lexical` and `hybrid` need a store created with `--hybrid`. Prefer\n\t * `lexical` over `hybrid` when fusing here: `hybrid` collapses both legs\n\t * into one RRF score, discarding the per-retriever ranks the trace records\n\t * and preventing n-way fusion with the grep leg.\n\t */\n\tasync query(text: string, k = 10, retriever: DaemonRetriever = \"dense\"): Promise<EmbSearchResult[]> {\n\t\tconst res = await this.send(\n\t\t\tretriever === \"dense\" ? { op: \"query\", text, k } : { op: \"query\", text, k, retriever },\n\t\t);\n\t\treturn res.results ?? [];\n\t}\n\n\t/**\n\t * Score `passages` against `query` with the daemon's cross-encoder and\n\t * return the best `k`, best first.\n\t *\n\t * Passages are sent inline rather than referenced by id: the caller has the\n\t * exact spans it intends to show the model, and a cross-encoder scores the\n\t * text it is given, so sending anything else would score the wrong thing.\n\t * Requires an onnx build with reranker weights (embsearch >= 0.3.0).\n\t */\n\tasync rerank(query: string, passages: EmbSearchRerankPassage[], k: number): Promise<EmbSearchRerankResult[]> {\n\t\tconst res = await this.send({ op: \"rerank\", query, passages, k });\n\t\treturn res.reranked ?? [];\n\t}\n\n\t/**\n\t * Batched insert-or-replace. One embedding inference for the whole batch —\n\t * the fast path for bulk indexing. Keep batches modest (e.g. 32–64) so a\n\t * concurrent query is not stuck behind a huge inference.\n\t */\n\tasync bulk(items: Array<{ id: string; text: string }>): Promise<EmbSearchBulkResult> {\n\t\tconst res = await this.send({ op: \"bulk\", items });\n\t\treturn { inserted: res.inserted_count ?? 0, updated: res.updated_count ?? 0 };\n\t}\n\n\t/** Model id, dimensionality, and live vector count of the daemon. */\n\tasync info(): Promise<EmbSearchDaemonInfo> {\n\t\tconst res = await this.send({ op: \"info\" });\n\t\treturn {\n\t\t\tmodelId: res.model_id ?? \"\",\n\t\t\tdim: res.dim ?? 0,\n\t\t\tcount: res.count ?? 0,\n\t\t\thybrid: res.hybrid,\n\t\t\trerank: res.rerank,\n\t\t};\n\t}\n\n\t/** Remove a record. Resolves to `true` if it existed. */\n\tasync remove(id: string): Promise<boolean> {\n\t\tconst res = await this.send({ op: \"remove\", id });\n\t\treturn res.removed === true;\n\t}\n\n\t/** Reclaim tombstoned rows left behind by `remove`. */\n\tasync compact(): Promise<void> {\n\t\tawait this.send({ op: \"compact\" });\n\t}\n\n\t/** Persist the index to the store directory. */\n\tasync save(): Promise<void> {\n\t\tawait this.send({ op: \"save\" });\n\t}\n\n\t/** Shut the daemon down, closing stdin so it exits cleanly. */\n\tasync close(): Promise<void> {\n\t\tif (this.closed) return;\n\t\tthis.proc.stdin.end();\n\t\tawait new Promise<void>((resolve) => this.proc.on(\"exit\", () => resolve()));\n\t}\n}\n"]}

package/dist/core/embsearch/embsearch-service.d.ts

@@ -38,6 +38,25 @@ export interface EmbsearchServiceOptions {
     cwd: string;
     /** Explicit binary path (settings override). Default: "embsearch" from PATH. */
     binaryPath?: string;
+    /**
+     * Model directory handed to the daemon as `--model`, overriding the model
+     * bundled in the binary.
+     *
+     * Only the eval harness sets this, to score two embedding models from one
+     * binary. Pair it with a distinct `storeDir`: vectors from different models
+     * are incompatible, and the daemon refuses to open a store built by another
+     * model rather than mixing them.
+     */
+    modelDir?: string;
+    /**
+     * Override the chunker's character cap.
+     *
+     * Only the eval harness sets this, to sweep the chunk window. It changes
+     * what every vector in the store *is*, and nothing stored records it, so it
+     * must be paired with a distinct `storeDir` exactly as `modelDir` is —
+     * otherwise a run silently scores an index built at another cap.
+     */
+    chunkMaxChars?: number;
     /** Minimum indexable bytes before indexing kicks in. */
     thresholdBytes: number;
     /**
@@ -83,6 +102,13 @@ export declare class EmbsearchService {
     constructor(options: EmbsearchServiceOptions);
     getState(): EmbsearchState;
     /** Semantic search is usable (index ready, or still building with partial data). */
+    /**
+     * Model id reported by the running daemon, once it is up.
+     *
+     * This — not the binary's version — identifies which model produced the
+     * vectors in the store, because `--model` decouples the two.
+     */
+    modelId(): string | undefined;
     isAvailable(): boolean;
     private setState;
     /**
@@ -122,6 +148,17 @@ export declare class EmbsearchService {
      */
     rerank(query: string, passages: EmbSearchRerankPassage[], k: number): Promise<EmbSearchRerankResult[]>;
     private probeBinaryVersion;
+    /**
+     * What the store on disk says about itself, read straight from its manifest.
+     *
+     * `store-info` exists precisely for the case where the daemon will not open
+     * the store: a `serve` pairs a store with an embedder and refuses the pair
+     * when their models disagree, so at that moment nothing else can tell us
+     * what built it. Returns undefined when there is no readable store — which
+     * includes a binary too old to have the subcommand, and so degrades to the
+     * previous behaviour rather than guessing.
+     */
+    private probeStore;
     private resolveBinary;
     /**
      * Scan, threshold-check, and (when needed) index in the background.

package/dist/core/embsearch/embsearch-service.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"embsearch-service.d.ts","sourceRoot":"","sources":["../../../src/core/embsearch/embsearch-service.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;GAcG;AAOH,OAAO,EACN,KAAK,eAAe,EAGpB,KAAK,sBAAsB,EAC3B,KAAK,qBAAqB,EAC1B,MAAM,aAAa,CAAC;AAuDrB,MAAM,MAAM,cAAc,GACvB;IAAE,KAAK,EAAE,MAAM,CAAA;CAAE,GACjB;IAAE,KAAK,EAAE,SAAS,CAAC;IAAC,MAAM,EAAE,MAAM,CAAA;CAAE,GACpC;IAAE,KAAK,EAAE,aAAa,CAAC;IAAC,aAAa,EAAE,MAAM,CAAC;IAAC,UAAU,EAAE,MAAM,GAAG,IAAI,CAAA;CAAE,GAC1E;IAAE,KAAK,EAAE,UAAU,CAAC;IAAC,IAAI,EAAE,MAAM,CAAC;IAAC,KAAK,EAAE,MAAM,CAAA;CAAE,GAClD;IAAE,KAAK,EAAE,OAAO,CAAC;IAAC,UAAU,EAAE,MAAM,CAAA;CAAE,GACtC;IAAE,KAAK,EAAE,aAAa,CAAC;IAAC,MAAM,EAAE,MAAM,CAAA;CAAE,CAAC;AAE5C,MAAM,WAAW,uBAAuB;IACvC,GAAG,EAAE,MAAM,CAAC;IACZ,gFAAgF;IAChF,UAAU,CAAC,EAAE,MAAM,CAAC;IACpB,wDAAwD;IACxD,cAAc,EAAE,MAAM,CAAC;IACvB;;;;OAIG;IACH,QAAQ,CAAC,EAAE,MAAM,CAAC;IAClB;;;;;;;OAOG;IACH,WAAW,CAAC,EAAE,OAAO,CAAC;IACtB,wDAAwD;IACxD,UAAU,CAAC,EAAE,CAAC,KAAK,EAAE,cAAc,KAAK,IAAI,CAAC;CAC7C;AAED,MAAM,WAAW,WAAW;IAC3B,IAAI,EAAE,MAAM,CAAC;IACb,SAAS,EAAE,MAAM,CAAC;IAClB,OAAO,EAAE,MAAM,CAAC;IAChB,KAAK,EAAE,MAAM,CAAC;CACd;AAED,MAAM,WAAW,gBAAiB,SAAQ,WAAW;IACpD,sFAAoF;IACpF,EAAE,EAAE,MAAM,CAAC;CACX;AAED,qBAAa,gBAAgB;IAC5B,OAAO,CAAC,QAAQ,CAAC,OAAO,CAA0B;IAClD,OAAO,CAAC,MAAM,CAA8B;IAC5C,OAAO,CAAC,IAAI,CAAwB;IACpC,OAAO,CAAC,KAAK,CAAqC;IAClD,OAAO,CAAC,QAAQ,CAAS;IACzB,iEAAiE;IACjE,OAAO,CAAC,gBAAgB,CAAS;IACjC,8DAA8D;IAC9D,OAAO,CAAC,WAAW,CAAS;IAC5B,wEAAwE;IACxE,OAAO,CAAC,YAAY,CAAS;IAE7B,YAAY,OAAO,EAAE,uBAAuB,EAE3C;IAED,QAAQ,IAAI,cAAc,CAEzB;IAED,oFAAoF;IACpF,WAAW,IAAI,OAAO,CAErB;IAED,OAAO,CAAC,QAAQ;IAKhB;;;OAGG;IACH,wBAAwB,IAAI,OAAO,CAElC;IAED;;;;;;;;;;;;;;;;;;;OAmBG;IACH,UAAU,CAAC,MAAM,CAAC,EAAE,WAAW,GAAG,MAAM,EAAE,CAezC;IAED,iEAAiE;IACjE,oBAAoB,IAAI,OAAO,CAE9B;IAED;;;;;;OAMG;IACG,MAAM,CAAC,KAAK,EAAE,MAAM,EAAE,QAAQ,EAAE,sBAAsB,EAAE,EAAE,CAAC,EAAE,MAAM,GAAG,OAAO,CAAC,qBAAqB,EAAE,CAAC,CAY3G;IAED,OAAO,CAAC,kBAAkB;YAQZ,aAAa;IAY3B;;;OAGG;IACG,KAAK,CAAC,MAAM,CAAC,EAAE,WAAW,GAAG,OAAO,CAAC,IAAI,CAAC,CAQ/C;YAEa,GAAG;YAsEH,iBAAiB;IAqF/B,OAAO,CAAC,WAAW;IAMnB,4DAA4D;IACtD,MAAM,CAAC,KAAK,EAAE,MAAM,EAAE,CAAC,SAAK,GAAG,OAAO,CAAC,WAAW,EAAE,CAAC,CAE1D;IAED,wEAAwE;IAClE,YAAY,CACjB,KAAK,EAAE,MAAM,EACb,CAAC,SAAK,EACN,IAAI,CAAC,EAAE,MAAM,EACb,SAAS,GAAE,eAAyB,GAClC,OAAO,CAAC,gBAAgB,EAAE,CAAC,CA2B7B;IAED;;;;;OAKG;IACH,kBAAkB,CACjB,GAAG,EAAE,MAAM,EACX,IAAI,EAAE,MAAM,GACV;QAAE,EAAE,EAAE,MAAM,CAAC;QAAC,IAAI,EAAE,MAAM,CAAC;QAAC,SAAS,EAAE,MAAM,CAAC;QAAC,OAAO,EAAE,MAAM,CAAA;KAAE,GAAG,SAAS,CAU9E;YAEa,WAAW;IAYzB,kEAAkE;IAC5D,OAAO,IAAI,OAAO,CAAC,IAAI,CAAC,CAW7B;CACD;AAWD,wBAAgB,wBAAwB,CAAC,GAAG,EAAE,MAAM,EAAE,OAAO,EAAE,gBAAgB,GAAG,IAAI,CAMrF;AAED,wBAAgB,mBAAmB,CAAC,GAAG,EAAE,MAAM,GAAG,gBAAgB,GAAG,SAAS,CAE7E;AAED,wBAAgB,0BAA0B,CAAC,GAAG,EAAE,MAAM,GAAG,IAAI,CAE5D","sourcesContent":["/**\n * Orchestrates semantic indexing and search for a repository.\n *\n * Lifecycle (all behind --enable-semantic-index):\n *  1. `start()` — resolve the embsearch binary, scan the repo (ignore-aware),\n *     apply the byte threshold. Under threshold → dormant. Over → spawn the\n *     daemon, verify the backend is not the mock embedder, then index changed\n *     files in the background in small batches, reporting progress.\n *  2. `search()` — top-k semantic query, mapping chunk ids back to\n *     `path:start-end` via the sidecar metadata.\n *  3. `dispose()` — save + close the daemon.\n *\n * Every failure degrades to `unavailable` with a reason; nothing here ever\n * blocks session startup or affects grep/find.\n */\n\nimport { execFileSync } from \"child_process\";\nimport { readFileSync, rmSync } from \"fs\";\nimport { minimatch } from \"minimatch\";\nimport { ensureTool } from \"../../utils/tools-manager.js\";\nimport { chunkFile } from \"./chunker.js\";\nimport {\n\ttype DaemonRetriever,\n\tEmbSearchClient,\n\ttype EmbSearchDaemonInfo,\n\ttype EmbSearchRerankPassage,\n\ttype EmbSearchRerankResult,\n} from \"./client.js\";\nimport {\n\temptyIndexMeta,\n\ttype FileMeta,\n\tgetEmbsearchStoreDir,\n\tgetVectorStoreDir,\n\thashContent,\n\ttype IndexMeta,\n\tloadIndexMeta,\n\tsaveIndexMeta,\n} from \"./index-meta.js\";\nimport { type RepoScanFile, scanRepo } from \"./repo-scan.js\";\n\n/** Chunks per bulk request. Small enough that a concurrent query is never\n *  stuck long behind one padded batch inference. */\nconst BULK_BATCH_SIZE = 48;\n/** Yield between batches so background indexing doesn't starve the session. */\nconst BATCH_YIELD_MS = 15;\n/** The Rust mock backend's model id — semantically meaningless, never index with it. */\nconst MOCK_MODEL_ID = \"mock-hash-v1\";\n/**\n * First embsearch release serving `retriever: \"lexical\"`.\n *\n * The guard matters because the daemon does not reject unknown request fields:\n * an older binary silently ignores `retriever` and answers with dense results.\n * Fusing that list a second time as a \"bm25\" leg would double-count it and\n * corrupt the ranking with no error anywhere — so refuse instead of degrading.\n */\nconst MIN_LEXICAL_RETRIEVER_VERSION = [0, 2, 0] as const;\n/**\n * First embsearch release serving the `rerank` op.\n *\n * Necessary but no longer sufficient. Releases from 0.3.1 carry the op without\n * the ~23 MB cross-encoder weights — they measured worse than the\n * deterministic reranker on five of six query classes, so they are no longer\n * bundled — and such a daemon answers `rerank` with an error. The version is\n * therefore only the floor for *asking*; `info.rerank` is the answer, and\n * {@link EmbsearchService.supportsCrossEncoder} needs both.\n */\nconst MIN_RERANK_VERSION = [0, 3, 0] as const;\n\n/** `embsearch 0.2.0` -> [0, 2, 0]; undefined when it cannot be parsed. */\nfunction parseBinaryVersion(output: string): number[] | undefined {\n\tconst match = output.match(/(\\d+)\\.(\\d+)\\.(\\d+)/);\n\treturn match ? [Number(match[1]), Number(match[2]), Number(match[3])] : undefined;\n}\n\nfunction atLeast(version: readonly number[], minimum: readonly number[]): boolean {\n\tfor (let i = 0; i < minimum.length; i++) {\n\t\tconst part = version[i] ?? 0;\n\t\tif (part !== minimum[i]) return part > minimum[i];\n\t}\n\treturn true;\n}\n\nexport type EmbsearchState =\n\t| { phase: \"idle\" }\n\t| { phase: \"skipped\"; reason: string }\n\t| { phase: \"downloading\"; receivedBytes: number; totalBytes: number | null }\n\t| { phase: \"indexing\"; done: number; total: number }\n\t| { phase: \"ready\"; chunkCount: number }\n\t| { phase: \"unavailable\"; reason: string };\n\nexport interface EmbsearchServiceOptions {\n\tcwd: string;\n\t/** Explicit binary path (settings override). Default: \"embsearch\" from PATH. */\n\tbinaryPath?: string;\n\t/** Minimum indexable bytes before indexing kicks in. */\n\tthresholdBytes: number;\n\t/**\n\t * Override the store location. Only the eval harness sets this, so a\n\t * second index (e.g. a BM25-hybrid store) can exist for the same repo\n\t * without colliding with the primary one.\n\t */\n\tstoreDir?: string;\n\t/**\n\t * Create the store with the daemon's BM25 lexical index.\n\t *\n\t * Defaults to whatever the daemon can serve. Fixed at store creation, so an\n\t * existing store that disagrees is rebuilt once; when overriding this to\n\t * hold two different stores for one repo, pair it with a distinct\n\t * `storeDir` so they do not fight over the same directory.\n\t */\n\thybridStore?: boolean;\n\t/** Progress callback for UI (footer / stderr lines). */\n\tonProgress?: (state: EmbsearchState) => void;\n}\n\nexport interface SemanticHit {\n\tpath: string;\n\tstartLine: number;\n\tendLine: number;\n\tscore: number;\n}\n\nexport interface SemanticChunkHit extends SemanticHit {\n\t/** Per-build chunk id (`relpath#index`) — the fusion identity for hybrid search. */\n\tid: string;\n}\n\nexport class EmbsearchService {\n\tprivate readonly options: EmbsearchServiceOptions;\n\tprivate client: EmbSearchClient | undefined;\n\tprivate meta: IndexMeta | undefined;\n\tprivate state: EmbsearchState = { phase: \"idle\" };\n\tprivate disposed = false;\n\t/** Whether the resolved binary serves `retriever: \"lexical\"`. */\n\tprivate lexicalRetriever = false;\n\t/** Whether the store actually opened carries a BM25 index. */\n\tprivate hybridStore = false;\n\t/** Whether the resolved binary serves the cross-encoder `rerank` op. */\n\tprivate crossEncoder = false;\n\n\tconstructor(options: EmbsearchServiceOptions) {\n\t\tthis.options = options;\n\t}\n\n\tgetState(): EmbsearchState {\n\t\treturn this.state;\n\t}\n\n\t/** Semantic search is usable (index ready, or still building with partial data). */\n\tisAvailable(): boolean {\n\t\treturn this.state.phase === \"ready\" || this.state.phase === \"indexing\";\n\t}\n\n\tprivate setState(state: EmbsearchState): void {\n\t\tthis.state = state;\n\t\tthis.options.onProgress?.(state);\n\t}\n\n\t/**\n\t * Whether a BM25-only query will work: the daemon has to understand the\n\t * `lexical` retriever *and* the open store has to carry a BM25 index.\n\t */\n\tsupportsLexicalRetriever(): boolean {\n\t\treturn this.lexicalRetriever && this.hybridStore;\n\t}\n\n\t/**\n\t * Repo files whose on-disk content the index does not have — unknown to it,\n\t * or changed since it last read them.\n\t *\n\t * This is the set BM25 is structurally blind to, and the only place the\n\t * grep leg still earns its keep once BM25 is available. An agent that edits\n\t * a file and immediately searches for what it wrote is asking about exactly\n\t * these files; the index cannot answer until the next pass.\n\t *\n\t * Compares mtime and size only, never hashing: the check runs per query, and\n\t * a false positive merely lets grep cover a file BM25 already covers, while\n\t * a false negative would lose the edit.\n\t *\n\t * Deliberately uncached. A cache here caches the *absence* of an edit, which\n\t * is the one thing this must never do — an agent writes a file and searches\n\t * for it in the same breath. A 1s TTL was tried and cost the live-edit set\n\t * 75% to 100% of its score depending on how the timing fell, which is worse\n\t * than wrong: it was non-deterministic. One scan is ~25ms over ~1k files and\n\t * happens once per search, against retrieval that already costs more.\n\t */\n\tstaleFiles(signal?: AbortSignal): string[] {\n\t\tif (!this.meta) return [];\n\t\tconst meta = this.meta;\n\t\tconst files: string[] = [];\n\t\ttry {\n\t\t\tfor (const file of scanRepo(this.options.cwd, signal).files) {\n\t\t\t\tconst known = meta.files[file.rel];\n\t\t\t\tif (!known || known.mtimeMs !== file.mtimeMs || known.size !== file.size) files.push(file.rel);\n\t\t\t}\n\t\t} catch {\n\t\t\t// A failed scan must not silently narrow the grep leg to nothing;\n\t\t\t// report no staleness and let the indexed legs answer.\n\t\t\treturn [];\n\t\t}\n\t\treturn files;\n\t}\n\n\t/** Whether the running daemon can score with a cross-encoder. */\n\tsupportsCrossEncoder(): boolean {\n\t\treturn this.crossEncoder;\n\t}\n\n\t/**\n\t * Cross-encoder rerank of caller-supplied passages.\n\t *\n\t * Unlike the retrievers this does not consult the index at all — it scores\n\t * exactly the text passed in, which is why the caller sends its expanded\n\t * windows rather than chunk ids.\n\t */\n\tasync rerank(query: string, passages: EmbSearchRerankPassage[], k: number): Promise<EmbSearchRerankResult[]> {\n\t\tif (!this.client || this.client.isClosed) {\n\t\t\tthrow new Error(\"semantic index is not available\");\n\t\t}\n\t\tif (!this.crossEncoder) {\n\t\t\tthrow new Error(\n\t\t\t\t`this embsearch daemon cannot rerank (needs >= ${MIN_RERANK_VERSION.join(\".\")} reporting ` +\n\t\t\t\t\t\"`rerank: true`); released binaries ship without cross-encoder weights — start the daemon \" +\n\t\t\t\t\t\"with --reranker-model <dir>\",\n\t\t\t);\n\t\t}\n\t\treturn await this.client.rerank(query, passages, k);\n\t}\n\n\tprivate probeBinaryVersion(binary: string): number[] | undefined {\n\t\ttry {\n\t\t\treturn parseBinaryVersion(execFileSync(binary, [\"--version\"], { encoding: \"utf-8\", timeout: 10_000 }));\n\t\t} catch {\n\t\t\treturn undefined;\n\t\t}\n\t}\n\n\tprivate async resolveBinary(): Promise<string | undefined> {\n\t\tif (this.options.binaryPath) {\n\t\t\treturn this.options.binaryPath;\n\t\t}\n\t\t// Surface the on-demand binary download through the same progress channel as\n\t\t// indexing, so the first-run fetch renders a progress bar instead of a stall.\n\t\t// A cached binary resolves without ever invoking this callback.\n\t\treturn await ensureTool(\"embsearch\", true, (receivedBytes, totalBytes) => {\n\t\t\tthis.setState({ phase: \"downloading\", receivedBytes, totalBytes });\n\t\t});\n\t}\n\n\t/**\n\t * Scan, threshold-check, and (when needed) index in the background.\n\t * Resolves when indexing completes or the feature settles dormant.\n\t */\n\tasync start(signal?: AbortSignal): Promise<void> {\n\t\ttry {\n\t\t\tawait this.run(signal);\n\t\t} catch (e) {\n\t\t\tconst reason = e instanceof Error ? e.message : String(e);\n\t\t\tthis.setState({ phase: \"unavailable\", reason });\n\t\t\tawait this.closeClient();\n\t\t}\n\t}\n\n\tprivate async run(signal?: AbortSignal): Promise<void> {\n\t\tconst binary = await this.resolveBinary();\n\t\tif (!binary) {\n\t\t\tthis.setState({\n\t\t\t\tphase: \"unavailable\",\n\t\t\t\treason: \"embsearch binary not found (PATH or embsearchBinaryPath setting)\",\n\t\t\t});\n\t\t\treturn;\n\t\t}\n\n\t\tconst scan = scanRepo(this.options.cwd, signal);\n\t\tif (scan.totalBytes < this.options.thresholdBytes) {\n\t\t\tthis.setState({\n\t\t\t\tphase: \"skipped\",\n\t\t\t\treason: `repo under threshold (${scan.totalBytes} < ${this.options.thresholdBytes} bytes)`,\n\t\t\t});\n\t\t\treturn;\n\t\t}\n\n\t\tconst binaryVersion = this.probeBinaryVersion(binary);\n\t\tthis.lexicalRetriever = binaryVersion !== undefined && atLeast(binaryVersion, MIN_LEXICAL_RETRIEVER_VERSION);\n\t\t// Provisional: the version says the daemon understands `rerank`. Whether\n\t\t// it can serve one is answered by `info` below, once the client is up.\n\t\tthis.crossEncoder = binaryVersion !== undefined && atLeast(binaryVersion, MIN_RERANK_VERSION);\n\n\t\tconst storeDir = this.options.storeDir ?? getEmbsearchStoreDir(this.options.cwd);\n\t\t// A hybrid store carries a BM25 index next to its vectors, which is what\n\t\t// lets search use BM25 as its lexical leg instead of ripgrep. Callers may\n\t\t// force it either way; by default it follows what the daemon can serve.\n\t\tconst wantHybrid = this.options.hybridStore ?? this.lexicalRetriever;\n\n\t\tconst openClient = async (): Promise<EmbSearchDaemonInfo> => {\n\t\t\tthis.client = new EmbSearchClient({\n\t\t\t\tbinaryPath: binary,\n\t\t\t\tstorePath: getVectorStoreDir(storeDir),\n\t\t\t\thybrid: wantHybrid,\n\t\t\t});\n\t\t\tawait this.client.ready();\n\t\t\treturn await this.client.info();\n\t\t};\n\n\t\tlet info = await openClient();\n\n\t\t// Hybrid-ness is fixed when a store is created and `--hybrid` against an\n\t\t// existing plain store only warns, so an index built before this was the\n\t\t// default would silently stay dense-only and every BM25 query against it\n\t\t// would fail. Ask the store itself rather than trusting the sidecar, and\n\t\t// rebuild once when it disagrees. `info.hybrid` is undefined on daemons\n\t\t// too old to report it — those cannot serve BM25 anyway, so leave them be.\n\t\tif (wantHybrid && info.hybrid === false) {\n\t\t\tthis.setState({ phase: \"indexing\", done: 0, total: 0 });\n\t\t\tawait this.closeClient();\n\t\t\trmSync(storeDir, { recursive: true, force: true });\n\t\t\tinfo = await openClient();\n\t\t}\n\t\tthis.hybridStore = info.hybrid === true;\n\t\t// A daemon old enough to omit the field is left on the version verdict —\n\t\t// back then the weights were bundled, so version did imply capability.\n\t\tif (info.rerank === false) this.crossEncoder = false;\n\t\tif (info.modelId === MOCK_MODEL_ID) {\n\t\t\tthrow new Error(\"embsearch binary uses the mock embedder (not semantic); install an onnx build\");\n\t\t}\n\n\t\t// Missing/stale sidecar (format, chunker, or model changed) → clean rebuild.\n\t\tthis.meta = loadIndexMeta(storeDir, info.modelId) ?? emptyIndexMeta(this.options.cwd, info.modelId);\n\t\tthis.meta.lastUsedMs = Date.now();\n\n\t\tawait this.indexChangedFiles(scan.files, storeDir, signal);\n\t}\n\n\tprivate async indexChangedFiles(files: RepoScanFile[], storeDir: string, signal?: AbortSignal): Promise<void> {\n\t\tconst meta = this.meta!;\n\t\tconst client = this.client!;\n\n\t\t// Diff scan vs sidecar: cheap mtime+size check first, hash only on delta.\n\t\tconst toIndex: Array<{ file: RepoScanFile; content: string; hash: string }> = [];\n\t\tconst seen = new Set<string>();\n\t\tfor (const file of files) {\n\t\t\tseen.add(file.rel);\n\t\t\tconst known = meta.files[file.rel];\n\t\t\tif (known && known.mtimeMs === file.mtimeMs && known.size === file.size) continue;\n\t\t\tlet content: string;\n\t\t\ttry {\n\t\t\t\tcontent = readFileSync(file.abs, \"utf-8\");\n\t\t\t} catch {\n\t\t\t\tcontinue;\n\t\t\t}\n\t\t\tconst hash = hashContent(content);\n\t\t\tif (known && known.hash === hash) {\n\t\t\t\t// Touched but unchanged — refresh stat info only.\n\t\t\t\tknown.mtimeMs = file.mtimeMs;\n\t\t\t\tknown.size = file.size;\n\t\t\t\tcontinue;\n\t\t\t}\n\t\t\ttoIndex.push({ file, content, hash });\n\t\t}\n\t\tconst toRemove = Object.keys(meta.files).filter((rel) => !seen.has(rel));\n\n\t\tif (toIndex.length === 0 && toRemove.length === 0) {\n\t\t\tsaveIndexMeta(storeDir, meta);\n\t\t\tthis.setState({ phase: \"ready\", chunkCount: this.countChunks(meta) });\n\t\t\treturn;\n\t\t}\n\n\t\t// Chunk changed files; count total upserts for exact progress.\n\t\tconst work: Array<{ rel: string; fileMeta: FileMeta; chunks: Array<{ id: string; text: string }> }> = [];\n\t\tlet totalChunks = 0;\n\t\tfor (const { file, content, hash } of toIndex) {\n\t\t\tconst chunks = chunkFile(file.rel, content);\n\t\t\twork.push({\n\t\t\t\trel: file.rel,\n\t\t\t\tfileMeta: {\n\t\t\t\t\tmtimeMs: file.mtimeMs,\n\t\t\t\t\tsize: file.size,\n\t\t\t\t\thash,\n\t\t\t\t\tchunks: chunks.map((c) => [c.startLine, c.endLine]),\n\t\t\t\t},\n\t\t\t\tchunks: chunks.map((c) => ({ id: c.id, text: c.text })),\n\t\t\t});\n\t\t\ttotalChunks += chunks.length;\n\t\t}\n\n\t\tthis.setState({ phase: \"indexing\", done: 0, total: totalChunks });\n\n\t\t// Drop vectors of deleted files and superseded chunk tails.\n\t\tfor (const rel of toRemove) {\n\t\t\tfor (let i = 0; i < meta.files[rel].chunks.length; i++) await client.remove(`${rel}#${i}`);\n\t\t\tdelete meta.files[rel];\n\t\t}\n\n\t\tlet done = 0;\n\t\tfor (const item of work) {\n\t\t\tif (signal?.aborted || this.disposed) return;\n\t\t\tconst oldChunkCount = meta.files[item.rel]?.chunks.length ?? 0;\n\t\t\t// Remove old chunks beyond the new count (upsert covers the rest).\n\t\t\tfor (let i = item.chunks.length; i < oldChunkCount; i++) await client.remove(`${item.rel}#${i}`);\n\n\t\t\tfor (let offset = 0; offset < item.chunks.length; offset += BULK_BATCH_SIZE) {\n\t\t\t\tif (signal?.aborted || this.disposed) return;\n\t\t\t\tconst batch = item.chunks.slice(offset, offset + BULK_BATCH_SIZE);\n\t\t\t\tawait client.bulk(batch);\n\t\t\t\tdone += batch.length;\n\t\t\t\tthis.setState({ phase: \"indexing\", done, total: totalChunks });\n\t\t\t\t// Yield so queries and the event loop stay responsive.\n\t\t\t\tawait new Promise((resolve) => setTimeout(resolve, BATCH_YIELD_MS));\n\t\t\t}\n\t\t\tmeta.files[item.rel] = item.fileMeta;\n\t\t}\n\n\t\tawait client.compact();\n\t\tawait client.save();\n\t\tsaveIndexMeta(storeDir, meta);\n\t\tthis.setState({ phase: \"ready\", chunkCount: this.countChunks(meta) });\n\t}\n\n\tprivate countChunks(meta: IndexMeta): number {\n\t\tlet n = 0;\n\t\tfor (const rel of Object.keys(meta.files)) n += meta.files[rel].chunks.length;\n\t\treturn n;\n\t}\n\n\t/** Top-`k` semantic hits as `path` + line range + score. */\n\tasync search(query: string, k = 10): Promise<SemanticHit[]> {\n\t\treturn await this.searchChunks(query, k);\n\t}\n\n\t/** Top-`k` semantic hits including their chunk ids, for rank fusion. */\n\tasync searchChunks(\n\t\tquery: string,\n\t\tk = 10,\n\t\tglob?: string,\n\t\tretriever: DaemonRetriever = \"dense\",\n\t): Promise<SemanticChunkHit[]> {\n\t\tif (!this.client || this.client.isClosed || !this.meta) {\n\t\t\tthrow new Error(\"semantic index is not available\");\n\t\t}\n\t\tif (retriever === \"lexical\" && !this.lexicalRetriever) {\n\t\t\tthrow new Error(\n\t\t\t\t`embsearch is too old for retriever \"lexical\" (needs >= ${MIN_LEXICAL_RETRIEVER_VERSION.join(\".\")}); ` +\n\t\t\t\t\t\"an older daemon ignores the field and answers with dense results\",\n\t\t\t);\n\t\t}\n\t\tconst results = await this.client.query(query, k, retriever);\n\t\tconst hits: SemanticChunkHit[] = [];\n\t\tconst matchGlob = (rel: string): boolean => {\n\t\t\tif (!glob) return true;\n\t\t\treturn minimatch(rel, glob, { dot: true, matchBase: !glob.includes(\"/\") });\n\t\t};\n\t\tfor (const result of results) {\n\t\t\tconst sep = result.id.lastIndexOf(\"#\");\n\t\t\tif (sep === -1) continue;\n\t\t\tconst rel = result.id.slice(0, sep);\n\t\t\tif (!matchGlob(rel)) continue;\n\t\t\tconst chunkIndex = Number.parseInt(result.id.slice(sep + 1), 10);\n\t\t\tconst range = this.meta.files[rel]?.chunks[chunkIndex];\n\t\t\tif (!range) continue;\n\t\t\thits.push({ id: result.id, path: rel, startLine: range[0], endLine: range[1], score: result.score });\n\t\t}\n\t\treturn hits;\n\t}\n\n\t/**\n\t * Resolve a repo-relative path + line to its enclosing indexed chunk, or\n\t * undefined when the file/line is not covered by the index. Chunks overlap\n\t * by a few lines; the first (lowest-index) containing chunk wins so the\n\t * mapping is deterministic.\n\t */\n\tfindEnclosingChunk(\n\t\trel: string,\n\t\tline: number,\n\t): { id: string; path: string; startLine: number; endLine: number } | undefined {\n\t\tconst file = this.meta?.files[rel];\n\t\tif (!file) return undefined;\n\t\tfor (let i = 0; i < file.chunks.length; i++) {\n\t\t\tconst [startLine, endLine] = file.chunks[i];\n\t\t\tif (line >= startLine && line <= endLine) {\n\t\t\t\treturn { id: `${rel}#${i}`, path: rel, startLine, endLine };\n\t\t\t}\n\t\t}\n\t\treturn undefined;\n\t}\n\n\tprivate async closeClient(): Promise<void> {\n\t\tconst client = this.client;\n\t\tthis.client = undefined;\n\t\tif (client && !client.isClosed) {\n\t\t\ttry {\n\t\t\t\tawait client.close();\n\t\t\t} catch {\n\t\t\t\t// already dead\n\t\t\t}\n\t\t}\n\t}\n\n\t/** Persist state and shut the daemon down. Safe to call twice. */\n\tasync dispose(): Promise<void> {\n\t\tif (this.disposed) return;\n\t\tthis.disposed = true;\n\t\tif (this.client && !this.client.isClosed) {\n\t\t\ttry {\n\t\t\t\tawait this.client.save();\n\t\t\t} catch {\n\t\t\t\t// daemon may have exited; nothing to save\n\t\t\t}\n\t\t}\n\t\tawait this.closeClient();\n\t}\n}\n\n// --- Per-cwd service registry ---\n//\n// The search tool is constructed by the generic tool factory table and\n// only receives `cwd`; the service is created later during session init (flag\n// gated). This registry connects the two without threading a service instance\n// through every layer between main.ts and the tool factories.\n\nconst services = new Map<string, EmbsearchService>();\n\nexport function registerEmbsearchService(cwd: string, service: EmbsearchService): void {\n\tconst old = services.get(cwd);\n\tif (old && old !== service) {\n\t\told.dispose().catch(() => {});\n\t}\n\tservices.set(cwd, service);\n}\n\nexport function getEmbsearchService(cwd: string): EmbsearchService | undefined {\n\treturn services.get(cwd);\n}\n\nexport function unregisterEmbsearchService(cwd: string): void {\n\tservices.delete(cwd);\n}\n"]}
+{"version":3,"file":"embsearch-service.d.ts","sourceRoot":"","sources":["../../../src/core/embsearch/embsearch-service.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;GAcG;AAOH,OAAO,EACN,KAAK,eAAe,EAGpB,KAAK,sBAAsB,EAC3B,KAAK,qBAAqB,EAC1B,MAAM,aAAa,CAAC;AAuDrB,MAAM,MAAM,cAAc,GACvB;IAAE,KAAK,EAAE,MAAM,CAAA;CAAE,GACjB;IAAE,KAAK,EAAE,SAAS,CAAC;IAAC,MAAM,EAAE,MAAM,CAAA;CAAE,GACpC;IAAE,KAAK,EAAE,aAAa,CAAC;IAAC,aAAa,EAAE,MAAM,CAAC;IAAC,UAAU,EAAE,MAAM,GAAG,IAAI,CAAA;CAAE,GAC1E;IAAE,KAAK,EAAE,UAAU,CAAC;IAAC,IAAI,EAAE,MAAM,CAAC;IAAC,KAAK,EAAE,MAAM,CAAA;CAAE,GAClD;IAAE,KAAK,EAAE,OAAO,CAAC;IAAC,UAAU,EAAE,MAAM,CAAA;CAAE,GACtC;IAAE,KAAK,EAAE,aAAa,CAAC;IAAC,MAAM,EAAE,MAAM,CAAA;CAAE,CAAC;AAE5C,MAAM,WAAW,uBAAuB;IACvC,GAAG,EAAE,MAAM,CAAC;IACZ,gFAAgF;IAChF,UAAU,CAAC,EAAE,MAAM,CAAC;IACpB;;;;;;;;OAQG;IACH,QAAQ,CAAC,EAAE,MAAM,CAAC;IAClB;;;;;;;OAOG;IACH,aAAa,CAAC,EAAE,MAAM,CAAC;IACvB,wDAAwD;IACxD,cAAc,EAAE,MAAM,CAAC;IACvB;;;;OAIG;IACH,QAAQ,CAAC,EAAE,MAAM,CAAC;IAClB;;;;;;;OAOG;IACH,WAAW,CAAC,EAAE,OAAO,CAAC;IACtB,wDAAwD;IACxD,UAAU,CAAC,EAAE,CAAC,KAAK,EAAE,cAAc,KAAK,IAAI,CAAC;CAC7C;AAED,MAAM,WAAW,WAAW;IAC3B,IAAI,EAAE,MAAM,CAAC;IACb,SAAS,EAAE,MAAM,CAAC;IAClB,OAAO,EAAE,MAAM,CAAC;IAChB,KAAK,EAAE,MAAM,CAAC;CACd;AAED,MAAM,WAAW,gBAAiB,SAAQ,WAAW;IACpD,sFAAoF;IACpF,EAAE,EAAE,MAAM,CAAC;CACX;AAED,qBAAa,gBAAgB;IAC5B,OAAO,CAAC,QAAQ,CAAC,OAAO,CAA0B;IAClD,OAAO,CAAC,MAAM,CAA8B;IAC5C,OAAO,CAAC,IAAI,CAAwB;IACpC,OAAO,CAAC,KAAK,CAAqC;IAClD,OAAO,CAAC,QAAQ,CAAS;IACzB,iEAAiE;IACjE,OAAO,CAAC,gBAAgB,CAAS;IACjC,8DAA8D;IAC9D,OAAO,CAAC,WAAW,CAAS;IAC5B,wEAAwE;IACxE,OAAO,CAAC,YAAY,CAAS;IAE7B,YAAY,OAAO,EAAE,uBAAuB,EAE3C;IAED,QAAQ,IAAI,cAAc,CAEzB;IAED,oFAAoF;IACpF;;;;;OAKG;IACH,OAAO,IAAI,MAAM,GAAG,SAAS,CAE5B;IAED,WAAW,IAAI,OAAO,CAErB;IAED,OAAO,CAAC,QAAQ;IAKhB;;;OAGG;IACH,wBAAwB,IAAI,OAAO,CAElC;IAED;;;;;;;;;;;;;;;;;;;OAmBG;IACH,UAAU,CAAC,MAAM,CAAC,EAAE,WAAW,GAAG,MAAM,EAAE,CAezC;IAED,iEAAiE;IACjE,oBAAoB,IAAI,OAAO,CAE9B;IAED;;;;;;OAMG;IACG,MAAM,CAAC,KAAK,EAAE,MAAM,EAAE,QAAQ,EAAE,sBAAsB,EAAE,EAAE,CAAC,EAAE,MAAM,GAAG,OAAO,CAAC,qBAAqB,EAAE,CAAC,CAY3G;IAED,OAAO,CAAC,kBAAkB;IAQ1B;;;;;;;;;OASG;IACH,OAAO,CAAC,UAAU;YAeJ,aAAa;IAY3B;;;OAGG;IACG,KAAK,CAAC,MAAM,CAAC,EAAE,WAAW,GAAG,OAAO,CAAC,IAAI,CAAC,CAQ/C;YAEa,GAAG;YAiHH,iBAAiB;IAqF/B,OAAO,CAAC,WAAW;IAMnB,4DAA4D;IACtD,MAAM,CAAC,KAAK,EAAE,MAAM,EAAE,CAAC,SAAK,GAAG,OAAO,CAAC,WAAW,EAAE,CAAC,CAE1D;IAED,wEAAwE;IAClE,YAAY,CACjB,KAAK,EAAE,MAAM,EACb,CAAC,SAAK,EACN,IAAI,CAAC,EAAE,MAAM,EACb,SAAS,GAAE,eAAyB,GAClC,OAAO,CAAC,gBAAgB,EAAE,CAAC,CA2B7B;IAED;;;;;OAKG;IACH,kBAAkB,CACjB,GAAG,EAAE,MAAM,EACX,IAAI,EAAE,MAAM,GACV;QAAE,EAAE,EAAE,MAAM,CAAC;QAAC,IAAI,EAAE,MAAM,CAAC;QAAC,SAAS,EAAE,MAAM,CAAC;QAAC,OAAO,EAAE,MAAM,CAAA;KAAE,GAAG,SAAS,CAU9E;YAEa,WAAW;IAYzB,kEAAkE;IAC5D,OAAO,IAAI,OAAO,CAAC,IAAI,CAAC,CAW7B;CACD;AAWD,wBAAgB,wBAAwB,CAAC,GAAG,EAAE,MAAM,EAAE,OAAO,EAAE,gBAAgB,GAAG,IAAI,CAMrF;AAED,wBAAgB,mBAAmB,CAAC,GAAG,EAAE,MAAM,GAAG,gBAAgB,GAAG,SAAS,CAE7E;AAED,wBAAgB,0BAA0B,CAAC,GAAG,EAAE,MAAM,GAAG,IAAI,CAE5D","sourcesContent":["/**\n * Orchestrates semantic indexing and search for a repository.\n *\n * Lifecycle (all behind --enable-semantic-index):\n *  1. `start()` — resolve the embsearch binary, scan the repo (ignore-aware),\n *     apply the byte threshold. Under threshold → dormant. Over → spawn the\n *     daemon, verify the backend is not the mock embedder, then index changed\n *     files in the background in small batches, reporting progress.\n *  2. `search()` — top-k semantic query, mapping chunk ids back to\n *     `path:start-end` via the sidecar metadata.\n *  3. `dispose()` — save + close the daemon.\n *\n * Every failure degrades to `unavailable` with a reason; nothing here ever\n * blocks session startup or affects grep/find.\n */\n\nimport { execFileSync } from \"child_process\";\nimport { readFileSync, rmSync } from \"fs\";\nimport { minimatch } from \"minimatch\";\nimport { ensureTool } from \"../../utils/tools-manager.js\";\nimport { chunkFile } from \"./chunker.js\";\nimport {\n\ttype DaemonRetriever,\n\tEmbSearchClient,\n\ttype EmbSearchDaemonInfo,\n\ttype EmbSearchRerankPassage,\n\ttype EmbSearchRerankResult,\n} from \"./client.js\";\nimport {\n\temptyIndexMeta,\n\ttype FileMeta,\n\tgetEmbsearchStoreDir,\n\tgetVectorStoreDir,\n\thashContent,\n\ttype IndexMeta,\n\tloadIndexMeta,\n\tsaveIndexMeta,\n} from \"./index-meta.js\";\nimport { type RepoScanFile, scanRepo } from \"./repo-scan.js\";\n\n/** Chunks per bulk request. Small enough that a concurrent query is never\n *  stuck long behind one padded batch inference. */\nconst BULK_BATCH_SIZE = 48;\n/** Yield between batches so background indexing doesn't starve the session. */\nconst BATCH_YIELD_MS = 15;\n/** The Rust mock backend's model id — semantically meaningless, never index with it. */\nconst MOCK_MODEL_ID = \"mock-hash-v1\";\n/**\n * First embsearch release serving `retriever: \"lexical\"`.\n *\n * The guard matters because the daemon does not reject unknown request fields:\n * an older binary silently ignores `retriever` and answers with dense results.\n * Fusing that list a second time as a \"bm25\" leg would double-count it and\n * corrupt the ranking with no error anywhere — so refuse instead of degrading.\n */\nconst MIN_LEXICAL_RETRIEVER_VERSION = [0, 2, 0] as const;\n/**\n * First embsearch release serving the `rerank` op.\n *\n * Necessary but no longer sufficient. Releases from 0.3.1 carry the op without\n * the ~23 MB cross-encoder weights — they measured worse than the\n * deterministic reranker on five of six query classes, so they are no longer\n * bundled — and such a daemon answers `rerank` with an error. The version is\n * therefore only the floor for *asking*; `info.rerank` is the answer, and\n * {@link EmbsearchService.supportsCrossEncoder} needs both.\n */\nconst MIN_RERANK_VERSION = [0, 3, 0] as const;\n\n/** `embsearch 0.2.0` -> [0, 2, 0]; undefined when it cannot be parsed. */\nfunction parseBinaryVersion(output: string): number[] | undefined {\n\tconst match = output.match(/(\\d+)\\.(\\d+)\\.(\\d+)/);\n\treturn match ? [Number(match[1]), Number(match[2]), Number(match[3])] : undefined;\n}\n\nfunction atLeast(version: readonly number[], minimum: readonly number[]): boolean {\n\tfor (let i = 0; i < minimum.length; i++) {\n\t\tconst part = version[i] ?? 0;\n\t\tif (part !== minimum[i]) return part > minimum[i];\n\t}\n\treturn true;\n}\n\nexport type EmbsearchState =\n\t| { phase: \"idle\" }\n\t| { phase: \"skipped\"; reason: string }\n\t| { phase: \"downloading\"; receivedBytes: number; totalBytes: number | null }\n\t| { phase: \"indexing\"; done: number; total: number }\n\t| { phase: \"ready\"; chunkCount: number }\n\t| { phase: \"unavailable\"; reason: string };\n\nexport interface EmbsearchServiceOptions {\n\tcwd: string;\n\t/** Explicit binary path (settings override). Default: \"embsearch\" from PATH. */\n\tbinaryPath?: string;\n\t/**\n\t * Model directory handed to the daemon as `--model`, overriding the model\n\t * bundled in the binary.\n\t *\n\t * Only the eval harness sets this, to score two embedding models from one\n\t * binary. Pair it with a distinct `storeDir`: vectors from different models\n\t * are incompatible, and the daemon refuses to open a store built by another\n\t * model rather than mixing them.\n\t */\n\tmodelDir?: string;\n\t/**\n\t * Override the chunker's character cap.\n\t *\n\t * Only the eval harness sets this, to sweep the chunk window. It changes\n\t * what every vector in the store *is*, and nothing stored records it, so it\n\t * must be paired with a distinct `storeDir` exactly as `modelDir` is —\n\t * otherwise a run silently scores an index built at another cap.\n\t */\n\tchunkMaxChars?: number;\n\t/** Minimum indexable bytes before indexing kicks in. */\n\tthresholdBytes: number;\n\t/**\n\t * Override the store location. Only the eval harness sets this, so a\n\t * second index (e.g. a BM25-hybrid store) can exist for the same repo\n\t * without colliding with the primary one.\n\t */\n\tstoreDir?: string;\n\t/**\n\t * Create the store with the daemon's BM25 lexical index.\n\t *\n\t * Defaults to whatever the daemon can serve. Fixed at store creation, so an\n\t * existing store that disagrees is rebuilt once; when overriding this to\n\t * hold two different stores for one repo, pair it with a distinct\n\t * `storeDir` so they do not fight over the same directory.\n\t */\n\thybridStore?: boolean;\n\t/** Progress callback for UI (footer / stderr lines). */\n\tonProgress?: (state: EmbsearchState) => void;\n}\n\nexport interface SemanticHit {\n\tpath: string;\n\tstartLine: number;\n\tendLine: number;\n\tscore: number;\n}\n\nexport interface SemanticChunkHit extends SemanticHit {\n\t/** Per-build chunk id (`relpath#index`) — the fusion identity for hybrid search. */\n\tid: string;\n}\n\nexport class EmbsearchService {\n\tprivate readonly options: EmbsearchServiceOptions;\n\tprivate client: EmbSearchClient | undefined;\n\tprivate meta: IndexMeta | undefined;\n\tprivate state: EmbsearchState = { phase: \"idle\" };\n\tprivate disposed = false;\n\t/** Whether the resolved binary serves `retriever: \"lexical\"`. */\n\tprivate lexicalRetriever = false;\n\t/** Whether the store actually opened carries a BM25 index. */\n\tprivate hybridStore = false;\n\t/** Whether the resolved binary serves the cross-encoder `rerank` op. */\n\tprivate crossEncoder = false;\n\n\tconstructor(options: EmbsearchServiceOptions) {\n\t\tthis.options = options;\n\t}\n\n\tgetState(): EmbsearchState {\n\t\treturn this.state;\n\t}\n\n\t/** Semantic search is usable (index ready, or still building with partial data). */\n\t/**\n\t * Model id reported by the running daemon, once it is up.\n\t *\n\t * This — not the binary's version — identifies which model produced the\n\t * vectors in the store, because `--model` decouples the two.\n\t */\n\tmodelId(): string | undefined {\n\t\treturn this.meta?.modelId;\n\t}\n\n\tisAvailable(): boolean {\n\t\treturn this.state.phase === \"ready\" || this.state.phase === \"indexing\";\n\t}\n\n\tprivate setState(state: EmbsearchState): void {\n\t\tthis.state = state;\n\t\tthis.options.onProgress?.(state);\n\t}\n\n\t/**\n\t * Whether a BM25-only query will work: the daemon has to understand the\n\t * `lexical` retriever *and* the open store has to carry a BM25 index.\n\t */\n\tsupportsLexicalRetriever(): boolean {\n\t\treturn this.lexicalRetriever && this.hybridStore;\n\t}\n\n\t/**\n\t * Repo files whose on-disk content the index does not have — unknown to it,\n\t * or changed since it last read them.\n\t *\n\t * This is the set BM25 is structurally blind to, and the only place the\n\t * grep leg still earns its keep once BM25 is available. An agent that edits\n\t * a file and immediately searches for what it wrote is asking about exactly\n\t * these files; the index cannot answer until the next pass.\n\t *\n\t * Compares mtime and size only, never hashing: the check runs per query, and\n\t * a false positive merely lets grep cover a file BM25 already covers, while\n\t * a false negative would lose the edit.\n\t *\n\t * Deliberately uncached. A cache here caches the *absence* of an edit, which\n\t * is the one thing this must never do — an agent writes a file and searches\n\t * for it in the same breath. A 1s TTL was tried and cost the live-edit set\n\t * 75% to 100% of its score depending on how the timing fell, which is worse\n\t * than wrong: it was non-deterministic. One scan is ~25ms over ~1k files and\n\t * happens once per search, against retrieval that already costs more.\n\t */\n\tstaleFiles(signal?: AbortSignal): string[] {\n\t\tif (!this.meta) return [];\n\t\tconst meta = this.meta;\n\t\tconst files: string[] = [];\n\t\ttry {\n\t\t\tfor (const file of scanRepo(this.options.cwd, signal).files) {\n\t\t\t\tconst known = meta.files[file.rel];\n\t\t\t\tif (!known || known.mtimeMs !== file.mtimeMs || known.size !== file.size) files.push(file.rel);\n\t\t\t}\n\t\t} catch {\n\t\t\t// A failed scan must not silently narrow the grep leg to nothing;\n\t\t\t// report no staleness and let the indexed legs answer.\n\t\t\treturn [];\n\t\t}\n\t\treturn files;\n\t}\n\n\t/** Whether the running daemon can score with a cross-encoder. */\n\tsupportsCrossEncoder(): boolean {\n\t\treturn this.crossEncoder;\n\t}\n\n\t/**\n\t * Cross-encoder rerank of caller-supplied passages.\n\t *\n\t * Unlike the retrievers this does not consult the index at all — it scores\n\t * exactly the text passed in, which is why the caller sends its expanded\n\t * windows rather than chunk ids.\n\t */\n\tasync rerank(query: string, passages: EmbSearchRerankPassage[], k: number): Promise<EmbSearchRerankResult[]> {\n\t\tif (!this.client || this.client.isClosed) {\n\t\t\tthrow new Error(\"semantic index is not available\");\n\t\t}\n\t\tif (!this.crossEncoder) {\n\t\t\tthrow new Error(\n\t\t\t\t`this embsearch daemon cannot rerank (needs >= ${MIN_RERANK_VERSION.join(\".\")} reporting ` +\n\t\t\t\t\t\"`rerank: true`); released binaries ship without cross-encoder weights — start the daemon \" +\n\t\t\t\t\t\"with --reranker-model <dir>\",\n\t\t\t);\n\t\t}\n\t\treturn await this.client.rerank(query, passages, k);\n\t}\n\n\tprivate probeBinaryVersion(binary: string): number[] | undefined {\n\t\ttry {\n\t\t\treturn parseBinaryVersion(execFileSync(binary, [\"--version\"], { encoding: \"utf-8\", timeout: 10_000 }));\n\t\t} catch {\n\t\t\treturn undefined;\n\t\t}\n\t}\n\n\t/**\n\t * What the store on disk says about itself, read straight from its manifest.\n\t *\n\t * `store-info` exists precisely for the case where the daemon will not open\n\t * the store: a `serve` pairs a store with an embedder and refuses the pair\n\t * when their models disagree, so at that moment nothing else can tell us\n\t * what built it. Returns undefined when there is no readable store — which\n\t * includes a binary too old to have the subcommand, and so degrades to the\n\t * previous behaviour rather than guessing.\n\t */\n\tprivate probeStore(binary: string, storeDir: string): { modelId: string; live: number } | undefined {\n\t\ttry {\n\t\t\tconst out = execFileSync(binary, [\"store-info\", \"--path\", getVectorStoreDir(storeDir), \"--json\"], {\n\t\t\t\tencoding: \"utf-8\",\n\t\t\t\ttimeout: 10_000,\n\t\t\t\tstdio: [\"ignore\", \"pipe\", \"ignore\"],\n\t\t\t});\n\t\t\tconst parsed = JSON.parse(out) as { model_id?: string; live?: number };\n\t\t\tif (typeof parsed.model_id !== \"string\") return undefined;\n\t\t\treturn { modelId: parsed.model_id, live: parsed.live ?? 0 };\n\t\t} catch {\n\t\t\treturn undefined;\n\t\t}\n\t}\n\n\tprivate async resolveBinary(): Promise<string | undefined> {\n\t\tif (this.options.binaryPath) {\n\t\t\treturn this.options.binaryPath;\n\t\t}\n\t\t// Surface the on-demand binary download through the same progress channel as\n\t\t// indexing, so the first-run fetch renders a progress bar instead of a stall.\n\t\t// A cached binary resolves without ever invoking this callback.\n\t\treturn await ensureTool(\"embsearch\", true, (receivedBytes, totalBytes) => {\n\t\t\tthis.setState({ phase: \"downloading\", receivedBytes, totalBytes });\n\t\t});\n\t}\n\n\t/**\n\t * Scan, threshold-check, and (when needed) index in the background.\n\t * Resolves when indexing completes or the feature settles dormant.\n\t */\n\tasync start(signal?: AbortSignal): Promise<void> {\n\t\ttry {\n\t\t\tawait this.run(signal);\n\t\t} catch (e) {\n\t\t\tconst reason = e instanceof Error ? e.message : String(e);\n\t\t\tthis.setState({ phase: \"unavailable\", reason });\n\t\t\tawait this.closeClient();\n\t\t}\n\t}\n\n\tprivate async run(signal?: AbortSignal): Promise<void> {\n\t\tconst binary = await this.resolveBinary();\n\t\tif (!binary) {\n\t\t\tthis.setState({\n\t\t\t\tphase: \"unavailable\",\n\t\t\t\treason: \"embsearch binary not found (PATH or embsearchBinaryPath setting)\",\n\t\t\t});\n\t\t\treturn;\n\t\t}\n\n\t\tconst scan = scanRepo(this.options.cwd, signal);\n\t\tif (scan.totalBytes < this.options.thresholdBytes) {\n\t\t\tthis.setState({\n\t\t\t\tphase: \"skipped\",\n\t\t\t\treason: `repo under threshold (${scan.totalBytes} < ${this.options.thresholdBytes} bytes)`,\n\t\t\t});\n\t\t\treturn;\n\t\t}\n\n\t\tconst binaryVersion = this.probeBinaryVersion(binary);\n\t\tthis.lexicalRetriever = binaryVersion !== undefined && atLeast(binaryVersion, MIN_LEXICAL_RETRIEVER_VERSION);\n\t\t// Provisional: the version says the daemon understands `rerank`. Whether\n\t\t// it can serve one is answered by `info` below, once the client is up.\n\t\tthis.crossEncoder = binaryVersion !== undefined && atLeast(binaryVersion, MIN_RERANK_VERSION);\n\n\t\tconst storeDir = this.options.storeDir ?? getEmbsearchStoreDir(this.options.cwd);\n\t\t// A hybrid store carries a BM25 index next to its vectors, which is what\n\t\t// lets search use BM25 as its lexical leg instead of ripgrep. Callers may\n\t\t// force it either way; by default it follows what the daemon can serve.\n\t\tconst wantHybrid = this.options.hybridStore ?? this.lexicalRetriever;\n\n\t\tconst openClient = async (): Promise<EmbSearchDaemonInfo> => {\n\t\t\tthis.client = new EmbSearchClient({\n\t\t\t\tbinaryPath: binary,\n\t\t\t\tstorePath: getVectorStoreDir(storeDir),\n\t\t\t\thybrid: wantHybrid,\n\t\t\t\tmodelDir: this.options.modelDir,\n\t\t\t});\n\t\t\tawait this.client.ready();\n\t\t\treturn await this.client.info();\n\t\t};\n\n\t\t/** Discard the store and start clean. The only recovery from a store the\n\t\t *  current binary cannot use — and the only way to be rid of vectors that\n\t\t *  outlived the metadata describing them. */\n\t\tconst rebuildFrom = async (why: string): Promise<EmbSearchDaemonInfo> => {\n\t\t\tconsole.error(`embsearch: rebuilding the index (${why})`);\n\t\t\tthis.setState({ phase: \"indexing\", done: 0, total: 0 });\n\t\t\tawait this.closeClient();\n\t\t\trmSync(storeDir, { recursive: true, force: true });\n\t\t\treturn await openClient();\n\t\t};\n\n\t\tlet info: EmbSearchDaemonInfo;\n\t\ttry {\n\t\t\tinfo = await openClient();\n\t\t} catch (err) {\n\t\t\t// The daemon refuses to open a store whose recorded model disagrees\n\t\t\t// with its own — correctly, since vectors from different models are\n\t\t\t// not comparable. But refusing is where it stopped: the store stayed\n\t\t\t// on disk, the daemon never came up, and this service went\n\t\t\t// permanently unavailable with a rebuild one directory-removal away.\n\t\t\t//\n\t\t\t// Only a store that is present and *readable* is treated this way. If\n\t\t\t// `store-info` cannot read it either, the problem is not a model\n\t\t\t// mismatch and destroying an index would be the wrong response, so\n\t\t\t// the original failure stands.\n\t\t\tconst store = this.probeStore(binary, storeDir);\n\t\t\tif (!store) throw err;\n\t\t\tinfo = await rebuildFrom(`built by model '${store.modelId}', which this binary cannot read`);\n\t\t}\n\n\t\t// Hybrid-ness is fixed when a store is created and `--hybrid` against an\n\t\t// existing plain store only warns, so an index built before this was the\n\t\t// default would silently stay dense-only and every BM25 query against it\n\t\t// would fail. Ask the store itself rather than trusting the sidecar, and\n\t\t// rebuild once when it disagrees. `info.hybrid` is undefined on daemons\n\t\t// too old to report it — those cannot serve BM25 anyway, so leave them be.\n\t\tif (wantHybrid && info.hybrid === false) {\n\t\t\tinfo = await rebuildFrom(\"the existing store carries no BM25 index\");\n\t\t}\n\t\tthis.hybridStore = info.hybrid === true;\n\t\t// A daemon old enough to omit the field is left on the version verdict —\n\t\t// back then the weights were bundled, so version did imply capability.\n\t\tif (info.rerank === false) this.crossEncoder = false;\n\t\tif (info.modelId === MOCK_MODEL_ID) {\n\t\t\tthrow new Error(\"embsearch binary uses the mock embedder (not semantic); install an onnx build\");\n\t\t}\n\n\t\t// Missing/stale sidecar (format, chunker, or model changed) → clean rebuild.\n\t\t//\n\t\t// Resetting the sidecar alone is not enough, and used to be all this did.\n\t\t// The sidecar is the only record of how many chunks each file produced,\n\t\t// so an empty one reports zero for every file — and `indexChangedFiles`\n\t\t// removes stale chunks by counting down from that number. Re-chunking a\n\t\t// file into *fewer* pieces then leaves its tail vectors (`path#N`,\n\t\t// `path#N+1`, …) in the store, holding text that no longer exists\n\t\t// anywhere, retrievable forever. Upserts hide it: chunk counts look\n\t\t// right, the sidecar looks right, and only search results are wrong.\n\t\t//\n\t\t// So when the sidecar cannot be trusted and the store is not already\n\t\t// empty, the store goes too.\n\t\tlet meta = loadIndexMeta(storeDir, info.modelId);\n\t\tif (!meta && info.count > 0) {\n\t\t\tinfo = await rebuildFrom(\"index metadata is missing or was written by a different chunker or model\");\n\t\t\tmeta = loadIndexMeta(storeDir, info.modelId);\n\t\t}\n\t\tthis.meta = meta ?? emptyIndexMeta(this.options.cwd, info.modelId);\n\t\tthis.meta.lastUsedMs = Date.now();\n\n\t\tawait this.indexChangedFiles(scan.files, storeDir, signal);\n\t}\n\n\tprivate async indexChangedFiles(files: RepoScanFile[], storeDir: string, signal?: AbortSignal): Promise<void> {\n\t\tconst meta = this.meta!;\n\t\tconst client = this.client!;\n\n\t\t// Diff scan vs sidecar: cheap mtime+size check first, hash only on delta.\n\t\tconst toIndex: Array<{ file: RepoScanFile; content: string; hash: string }> = [];\n\t\tconst seen = new Set<string>();\n\t\tfor (const file of files) {\n\t\t\tseen.add(file.rel);\n\t\t\tconst known = meta.files[file.rel];\n\t\t\tif (known && known.mtimeMs === file.mtimeMs && known.size === file.size) continue;\n\t\t\tlet content: string;\n\t\t\ttry {\n\t\t\t\tcontent = readFileSync(file.abs, \"utf-8\");\n\t\t\t} catch {\n\t\t\t\tcontinue;\n\t\t\t}\n\t\t\tconst hash = hashContent(content);\n\t\t\tif (known && known.hash === hash) {\n\t\t\t\t// Touched but unchanged — refresh stat info only.\n\t\t\t\tknown.mtimeMs = file.mtimeMs;\n\t\t\t\tknown.size = file.size;\n\t\t\t\tcontinue;\n\t\t\t}\n\t\t\ttoIndex.push({ file, content, hash });\n\t\t}\n\t\tconst toRemove = Object.keys(meta.files).filter((rel) => !seen.has(rel));\n\n\t\tif (toIndex.length === 0 && toRemove.length === 0) {\n\t\t\tsaveIndexMeta(storeDir, meta);\n\t\t\tthis.setState({ phase: \"ready\", chunkCount: this.countChunks(meta) });\n\t\t\treturn;\n\t\t}\n\n\t\t// Chunk changed files; count total upserts for exact progress.\n\t\tconst work: Array<{ rel: string; fileMeta: FileMeta; chunks: Array<{ id: string; text: string }> }> = [];\n\t\tlet totalChunks = 0;\n\t\tfor (const { file, content, hash } of toIndex) {\n\t\t\tconst chunks = chunkFile(file.rel, content, this.options.chunkMaxChars);\n\t\t\twork.push({\n\t\t\t\trel: file.rel,\n\t\t\t\tfileMeta: {\n\t\t\t\t\tmtimeMs: file.mtimeMs,\n\t\t\t\t\tsize: file.size,\n\t\t\t\t\thash,\n\t\t\t\t\tchunks: chunks.map((c) => [c.startLine, c.endLine]),\n\t\t\t\t},\n\t\t\t\tchunks: chunks.map((c) => ({ id: c.id, text: c.text })),\n\t\t\t});\n\t\t\ttotalChunks += chunks.length;\n\t\t}\n\n\t\tthis.setState({ phase: \"indexing\", done: 0, total: totalChunks });\n\n\t\t// Drop vectors of deleted files and superseded chunk tails.\n\t\tfor (const rel of toRemove) {\n\t\t\tfor (let i = 0; i < meta.files[rel].chunks.length; i++) await client.remove(`${rel}#${i}`);\n\t\t\tdelete meta.files[rel];\n\t\t}\n\n\t\tlet done = 0;\n\t\tfor (const item of work) {\n\t\t\tif (signal?.aborted || this.disposed) return;\n\t\t\tconst oldChunkCount = meta.files[item.rel]?.chunks.length ?? 0;\n\t\t\t// Remove old chunks beyond the new count (upsert covers the rest).\n\t\t\tfor (let i = item.chunks.length; i < oldChunkCount; i++) await client.remove(`${item.rel}#${i}`);\n\n\t\t\tfor (let offset = 0; offset < item.chunks.length; offset += BULK_BATCH_SIZE) {\n\t\t\t\tif (signal?.aborted || this.disposed) return;\n\t\t\t\tconst batch = item.chunks.slice(offset, offset + BULK_BATCH_SIZE);\n\t\t\t\tawait client.bulk(batch);\n\t\t\t\tdone += batch.length;\n\t\t\t\tthis.setState({ phase: \"indexing\", done, total: totalChunks });\n\t\t\t\t// Yield so queries and the event loop stay responsive.\n\t\t\t\tawait new Promise((resolve) => setTimeout(resolve, BATCH_YIELD_MS));\n\t\t\t}\n\t\t\tmeta.files[item.rel] = item.fileMeta;\n\t\t}\n\n\t\tawait client.compact();\n\t\tawait client.save();\n\t\tsaveIndexMeta(storeDir, meta);\n\t\tthis.setState({ phase: \"ready\", chunkCount: this.countChunks(meta) });\n\t}\n\n\tprivate countChunks(meta: IndexMeta): number {\n\t\tlet n = 0;\n\t\tfor (const rel of Object.keys(meta.files)) n += meta.files[rel].chunks.length;\n\t\treturn n;\n\t}\n\n\t/** Top-`k` semantic hits as `path` + line range + score. */\n\tasync search(query: string, k = 10): Promise<SemanticHit[]> {\n\t\treturn await this.searchChunks(query, k);\n\t}\n\n\t/** Top-`k` semantic hits including their chunk ids, for rank fusion. */\n\tasync searchChunks(\n\t\tquery: string,\n\t\tk = 10,\n\t\tglob?: string,\n\t\tretriever: DaemonRetriever = \"dense\",\n\t): Promise<SemanticChunkHit[]> {\n\t\tif (!this.client || this.client.isClosed || !this.meta) {\n\t\t\tthrow new Error(\"semantic index is not available\");\n\t\t}\n\t\tif (retriever === \"lexical\" && !this.lexicalRetriever) {\n\t\t\tthrow new Error(\n\t\t\t\t`embsearch is too old for retriever \"lexical\" (needs >= ${MIN_LEXICAL_RETRIEVER_VERSION.join(\".\")}); ` +\n\t\t\t\t\t\"an older daemon ignores the field and answers with dense results\",\n\t\t\t);\n\t\t}\n\t\tconst results = await this.client.query(query, k, retriever);\n\t\tconst hits: SemanticChunkHit[] = [];\n\t\tconst matchGlob = (rel: string): boolean => {\n\t\t\tif (!glob) return true;\n\t\t\treturn minimatch(rel, glob, { dot: true, matchBase: !glob.includes(\"/\") });\n\t\t};\n\t\tfor (const result of results) {\n\t\t\tconst sep = result.id.lastIndexOf(\"#\");\n\t\t\tif (sep === -1) continue;\n\t\t\tconst rel = result.id.slice(0, sep);\n\t\t\tif (!matchGlob(rel)) continue;\n\t\t\tconst chunkIndex = Number.parseInt(result.id.slice(sep + 1), 10);\n\t\t\tconst range = this.meta.files[rel]?.chunks[chunkIndex];\n\t\t\tif (!range) continue;\n\t\t\thits.push({ id: result.id, path: rel, startLine: range[0], endLine: range[1], score: result.score });\n\t\t}\n\t\treturn hits;\n\t}\n\n\t/**\n\t * Resolve a repo-relative path + line to its enclosing indexed chunk, or\n\t * undefined when the file/line is not covered by the index. Chunks overlap\n\t * by a few lines; the first (lowest-index) containing chunk wins so the\n\t * mapping is deterministic.\n\t */\n\tfindEnclosingChunk(\n\t\trel: string,\n\t\tline: number,\n\t): { id: string; path: string; startLine: number; endLine: number } | undefined {\n\t\tconst file = this.meta?.files[rel];\n\t\tif (!file) return undefined;\n\t\tfor (let i = 0; i < file.chunks.length; i++) {\n\t\t\tconst [startLine, endLine] = file.chunks[i];\n\t\t\tif (line >= startLine && line <= endLine) {\n\t\t\t\treturn { id: `${rel}#${i}`, path: rel, startLine, endLine };\n\t\t\t}\n\t\t}\n\t\treturn undefined;\n\t}\n\n\tprivate async closeClient(): Promise<void> {\n\t\tconst client = this.client;\n\t\tthis.client = undefined;\n\t\tif (client && !client.isClosed) {\n\t\t\ttry {\n\t\t\t\tawait client.close();\n\t\t\t} catch {\n\t\t\t\t// already dead\n\t\t\t}\n\t\t}\n\t}\n\n\t/** Persist state and shut the daemon down. Safe to call twice. */\n\tasync dispose(): Promise<void> {\n\t\tif (this.disposed) return;\n\t\tthis.disposed = true;\n\t\tif (this.client && !this.client.isClosed) {\n\t\t\ttry {\n\t\t\t\tawait this.client.save();\n\t\t\t} catch {\n\t\t\t\t// daemon may have exited; nothing to save\n\t\t\t}\n\t\t}\n\t\tawait this.closeClient();\n\t}\n}\n\n// --- Per-cwd service registry ---\n//\n// The search tool is constructed by the generic tool factory table and\n// only receives `cwd`; the service is created later during session init (flag\n// gated). This registry connects the two without threading a service instance\n// through every layer between main.ts and the tool factories.\n\nconst services = new Map<string, EmbsearchService>();\n\nexport function registerEmbsearchService(cwd: string, service: EmbsearchService): void {\n\tconst old = services.get(cwd);\n\tif (old && old !== service) {\n\t\told.dispose().catch(() => {});\n\t}\n\tservices.set(cwd, service);\n}\n\nexport function getEmbsearchService(cwd: string): EmbsearchService | undefined {\n\treturn services.get(cwd);\n}\n\nexport function unregisterEmbsearchService(cwd: string): void {\n\tservices.delete(cwd);\n}\n"]}