@cvfile/embed 0.1.0 → 0.3.1

package/dist/index.cjs

@@ -3,25 +3,31 @@
 var transformers = require('@huggingface/transformers');
 
 // src/chunk.ts
+var encoder = new TextEncoder();
+var decoder = new TextDecoder();
 var HEADING = /^(#{1,6})\s+(.+?)\s*$/;
 function chunkMarkdown(markdown, opts = {}) {
   const mode = opts.mode ?? "section";
+  const bytes = encoder.encode(markdown);
   if (mode === "document") {
-    return [{ id: "document", textOffset: 0, textLength: markdown.length, text: markdown }];
+    return [documentChunk(bytes)];
   }
   if (mode === "paragraph") {
-    return paragraphChunks(markdown);
+    return paragraphChunks(bytes);
   }
-  return sectionChunks(markdown);
+  return sectionChunks(bytes);
 }
-function sectionChunks(markdown) {
-  const lines = splitWithOffsets(markdown);
+function documentChunk(bytes) {
+  return { id: "document", textOffset: 0, textLength: bytes.byteLength, text: sliceText(bytes, 0, bytes.byteLength) };
+}
+function sectionChunks(bytes) {
+  const lines = splitWithByteOffsets(bytes);
   const sections = [];
   let current = null;
   const ids = /* @__PURE__ */ new Set();
   function flush(end) {
     if (!current) return;
-    const text = markdown.slice(current.start, end);
+    const text = sliceText(bytes, current.start, end);
     if (text.trim().length === 0) {
       current = null;
       return;
@@ -31,59 +37,84 @@ function sectionChunks(markdown) {
   }
   for (const line of lines) {
     const match = HEADING.exec(line.text);
+    const lineEnd = line.offset + line.byteLength;
     if (match) {
       flush(line.offset);
       const id = uniqueId(slugify(match[2] ?? `section-${sections.length + 1}`), ids);
       ids.add(id);
-      current = { id, start: line.offset, end: line.offset + line.text.length };
+      current = { id, start: line.offset, end: lineEnd };
       continue;
     }
     if (current === null) {
       const id = uniqueId("preamble", ids);
       ids.add(id);
-      current = { id, start: line.offset, end: line.offset + line.text.length };
+      current = { id, start: line.offset, end: lineEnd };
     } else {
-      current.end = line.offset + line.text.length;
+      current.end = lineEnd;
     }
   }
-  flush(markdown.length);
+  flush(bytes.byteLength);
   if (sections.length === 0) {
-    return [{ id: "document", textOffset: 0, textLength: markdown.length, text: markdown }];
+    return [documentChunk(bytes)];
   }
   return sections;
 }
-function paragraphChunks(markdown) {
+function paragraphChunks(bytes) {
   const out = [];
   const ids = /* @__PURE__ */ new Set();
+  const separator = encoder.encode("\n\n");
   let cursor = 0;
   let i = 0;
-  while (cursor < markdown.length) {
-    let end = markdown.indexOf("\n\n", cursor);
-    if (end === -1) end = markdown.length;
-    const text = markdown.slice(cursor, end);
+  while (cursor < bytes.byteLength) {
+    let end = indexOfBytes(bytes, separator, cursor);
+    if (end === -1) end = bytes.byteLength;
+    const text = sliceText(bytes, cursor, end);
     if (text.trim().length > 0) {
       const id = uniqueId(slugify(text.split("\n")[0] ?? `p-${i}`), ids);
       ids.add(id);
-      out.push({ id, textOffset: cursor, textLength: text.length, text });
+      out.push({ id, textOffset: cursor, textLength: end - cursor, text });
       i += 1;
     }
-    cursor = end + 2;
+    cursor = end + separator.byteLength;
   }
   if (out.length === 0) {
-    return [{ id: "document", textOffset: 0, textLength: markdown.length, text: markdown }];
+    return [documentChunk(bytes)];
   }
   return out;
 }
-function splitWithOffsets(s) {
+function sliceText(bytes, start, end) {
+  return decoder.decode(bytes.subarray(start, end));
+}
+function splitWithByteOffsets(bytes) {
+  const newline = 10;
   const lines = [];
-  let offset = 0;
-  for (const line of s.split("\n")) {
-    const withNl = line + (offset + line.length < s.length ? "\n" : "");
-    lines.push({ text: withNl, offset });
-    offset += withNl.length;
+  let start = 0;
+  for (let i = 0; i < bytes.byteLength; i += 1) {
+    if (bytes[i] === newline) {
+      const byteLength = i - start + 1;
+      lines.push({ text: sliceText(bytes, start, i + 1), offset: start, byteLength });
+      start = i + 1;
+    }
+  }
+  if (start < bytes.byteLength) {
+    lines.push({ text: sliceText(bytes, start, bytes.byteLength), offset: start, byteLength: bytes.byteLength - start });
   }
   return lines;
 }
+function indexOfBytes(haystack, needle, from) {
+  const last = haystack.byteLength - needle.byteLength;
+  for (let i = from; i <= last; i += 1) {
+    let matched = true;
+    for (let j = 0; j < needle.byteLength; j += 1) {
+      if (haystack[i + j] !== needle[j]) {
+        matched = false;
+        break;
+      }
+    }
+    if (matched) return i;
+  }
+  return -1;
+}
 function slugify(s) {
   return s.toLowerCase().replace(/[^a-z0-9]+/g, "-").replace(/^-+|-+$/g, "").slice(0, 64) || "section";
 }

package/dist/index.cjs.map

@@ -1 +1 @@
-{"version":3,"sources":["../src/chunk.ts","../src/transformers-backend.ts","../src/types.ts","../src/embed.ts","../src/search.ts","../src/huggingface-backend.ts"],"names":["env","pipeline"],"mappings":";;;;;AAsBA,IAAM,OAAA,GAAU,uBAAA;AAET,SAAS,aAAA,CAAc,QAAA,EAAkB,IAAA,GAAqB,EAAC,EAAoB;AACxF,EAAA,MAAM,IAAA,GAAO,KAAK,IAAA,IAAQ,SAAA;AAC1B,EAAA,IAAI,SAAS,UAAA,EAAY;AACvB,IAAA,OAAO,CAAC,EAAE,EAAA,EAAI,UAAA,EAAY,UAAA,EAAY,CAAA,EAAG,UAAA,EAAY,QAAA,CAAS,MAAA,EAAQ,IAAA,EAAM,QAAA,EAAU,CAAA;AAAA,EACxF;AACA,EAAA,IAAI,SAAS,WAAA,EAAa;AACxB,IAAA,OAAO,gBAAgB,QAAQ,CAAA;AAAA,EACjC;AACA,EAAA,OAAO,cAAc,QAAQ,CAAA;AAC/B;AAEA,SAAS,cAAc,QAAA,EAAmC;AACxD,EAAA,MAAM,KAAA,GAAQ,iBAAiB,QAAQ,CAAA;AACvC,EAAA,MAAM,WAA4B,EAAC;AACnC,EAAA,IAAI,OAAA,GAA6D,IAAA;AACjE,EAAA,MAAM,GAAA,uBAAU,GAAA,EAAY;AAE5B,EAAA,SAAS,MAAM,GAAA,EAAmB;AAChC,IAAA,IAAI,CAAC,OAAA,EAAS;AACd,IAAA,MAAM,IAAA,GAAO,QAAA,CAAS,KAAA,CAAM,OAAA,CAAQ,OAAO,GAAG,CAAA;AAC9C,IAAA,IAAI,IAAA,CAAK,IAAA,EAAK,CAAE,MAAA,KAAW,CAAA,EAAG;AAC5B,MAAA,OAAA,GAAU,IAAA;AACV,MAAA;AAAA,IACF;AACA,IAAA,QAAA,CAAS,IAAA,CAAK,EAAE,EAAA,EAAI,OAAA,CAAQ,EAAA,EAAI,UAAA,EAAY,OAAA,CAAQ,KAAA,EAAO,UAAA,EAAY,GAAA,GAAM,OAAA,CAAQ,KAAA,EAAO,MAAM,CAAA;AAClG,IAAA,OAAA,GAAU,IAAA;AAAA,EACZ;AAEA,EAAA,KAAA,MAAW,QAAQ,KAAA,EAAO;AACxB,IAAA,MAAM,KAAA,GAAQ,OAAA,CAAQ,IAAA,CAAK,IAAA,CAAK,IAAI,CAAA;AACpC,IAAA,IAAI,KAAA,EAAO;AACT,MAAA,KAAA,CAAM,KAAK,MAAM,CAAA;AACjB,MAAA,MAAM,EAAA,GAAK,QAAA,CAAS,OAAA,CAAQ,KAAA,CAAM,CAAC,CAAA,IAAK,CAAA,QAAA,EAAW,QAAA,CAAS,MAAA,GAAS,CAAC,CAAA,CAAE,CAAA,EAAG,GAAG,CAAA;AAC9E,MAAA,GAAA,CAAI,IAAI,EAAE,CAAA;AACV,MAAA,OAAA,GAAU,EAAE,EAAA,EAAI,KAAA,EAAO,IAAA,CAAK,MAAA,EAAQ,KAAK,IAAA,CAAK,MAAA,GAAS,IAAA,CAAK,IAAA,CAAK,MAAA,EAAO;AACxE,MAAA;AAAA,IACF;AACA,IAAA,IAAI,YAAY,IAAA,EAAM;AACpB,MAAA,MAAM,EAAA,GAAK,QAAA,CAAS,UAAA,EAAY,GAAG,CAAA;AACnC,MAAA,GAAA,CAAI,IAAI,EAAE,CAAA;AACV,MAAA,OAAA,GAAU,EAAE,EAAA,EAAI,KAAA,EAAO,IAAA,CAAK,MAAA,EAAQ,KAAK,IAAA,CAAK,MAAA,GAAS,IAAA,CAAK,IAAA,CAAK,MAAA,EAAO;AAAA,IAC1E,CAAA,MAAO;AACL,MAAA,OAAA,CAAQ,GAAA,GAAM,IAAA,CAAK,MAAA,GAAS,IAAA,CAAK,IAAA,CAAK,MAAA;AAAA,IACxC;AAAA,EACF;AACA,EAAA,KAAA,CAAM,SAAS,MAAM,CAAA;AAErB,EAAA,IAAI,QAAA,CAAS,WAAW,CAAA,EAAG;AACzB,IAAA,OAAO,CAAC,EAAE,EAAA,EAAI,UAAA,EAAY,UAAA,EAAY,CAAA,EAAG,UAAA,EAAY,QAAA,CAAS,MAAA,EAAQ,IAAA,EAAM,QAAA,EAAU,CAAA;AAAA,EACxF;AACA,EAAA,OAAO,QAAA;AACT;AAEA,SAAS,gBAAgB,QAAA,EAAmC;AAC1D,EAAA,MAAM,MAAuB,EAAC;AAC9B,EAAA,MAAM,GAAA,uBAAU,GAAA,EAAY;AAC5B,EAAA,IAAI,MAAA,GAAS,CAAA;AACb,EAAA,IAAI,CAAA,GAAI,CAAA;AACR,EAAA,OAAO,MAAA,GAAS,SAAS,MAAA,EAAQ;AAC/B,IAAA,IAAI,GAAA,GAAM,QAAA,CAAS,OAAA,CAAQ,MAAA,EAAQ,MAAM,CAAA;AACzC,IAAA,IAAI,GAAA,KAAQ,EAAA,EAAI,GAAA,GAAM,QAAA,CAAS,MAAA;AAC/B,IAAA,MAAM,IAAA,GAAO,QAAA,CAAS,KAAA,CAAM,MAAA,EAAQ,GAAG,CAAA;AACvC,IAAA,IAAI,IAAA,CAAK,IAAA,EAAK,CAAE,MAAA,GAAS,CAAA,EAAG;AAC1B,MAAA,MAAM,EAAA,GAAK,QAAA,CAAS,OAAA,CAAQ,IAAA,CAAK,KAAA,CAAM,IAAI,CAAA,CAAE,CAAC,CAAA,IAAK,CAAA,EAAA,EAAK,CAAC,CAAA,CAAE,GAAG,GAAG,CAAA;AACjE,MAAA,GAAA,CAAI,IAAI,EAAE,CAAA;AACV,MAAA,GAAA,CAAI,IAAA,CAAK,EAAE,EAAA,EAAI,UAAA,EAAY,QAAQ,UAAA,EAAY,IAAA,CAAK,MAAA,EAAQ,IAAA,EAAM,CAAA;AAClE,MAAA,CAAA,IAAK,CAAA;AAAA,IACP;AACA,IAAA,MAAA,GAAS,GAAA,GAAM,CAAA;AAAA,EACjB;AACA,EAAA,IAAI,GAAA,CAAI,WAAW,CAAA,EAAG;AACpB,IAAA,OAAO,CAAC,EAAE,EAAA,EAAI,UAAA,EAAY,UAAA,EAAY,CAAA,EAAG,UAAA,EAAY,QAAA,CAAS,MAAA,EAAQ,IAAA,EAAM,QAAA,EAAU,CAAA;AAAA,EACxF;AACA,EAAA,OAAO,GAAA;AACT;AAEA,SAAS,iBAAiB,CAAA,EAA+C;AACvE,EAAA,MAAM,QAA4C,EAAC;AACnD,EAAA,IAAI,MAAA,GAAS,CAAA;AACb,EAAA,KAAA,MAAW,IAAA,IAAQ,CAAA,CAAE,KAAA,CAAM,IAAI,CAAA,EAAG;AAChC,IAAA,MAAM,SAAS,IAAA,IAAQ,MAAA,GAAS,KAAK,MAAA,GAAS,CAAA,CAAE,SAAS,IAAA,GAAO,EAAA,CAAA;AAChE,IAAA,KAAA,CAAM,IAAA,CAAK,EAAE,IAAA,EAAM,MAAA,EAAQ,QAAQ,CAAA;AACnC,IAAA,MAAA,IAAU,MAAA,CAAO,MAAA;AAAA,EACnB;AACA,EAAA,OAAO,KAAA;AACT;AAEA,SAAS,QAAQ,CAAA,EAAmB;AAClC,EAAA,OACE,CAAA,CACG,WAAA,EAAY,CACZ,OAAA,CAAQ,eAAe,GAAG,CAAA,CAC1B,OAAA,CAAQ,UAAA,EAAY,EAAE,CAAA,CACtB,KAAA,CAAM,CAAA,EAAG,EAAE,CAAA,IAAK,SAAA;AAEvB;AAEA,SAAS,QAAA,CAAS,MAAc,KAAA,EAA4B;AAC1D,EAAA,IAAI,CAAC,KAAA,CAAM,GAAA,CAAI,IAAI,GAAG,OAAO,IAAA;AAC7B,EAAA,IAAI,CAAA,GAAI,CAAA;AACR,EAAA,OAAO,KAAA,CAAM,IAAI,CAAA,EAAG,IAAI,IAAI,CAAC,CAAA,CAAE,GAAG,CAAA,IAAK,CAAA;AACvC,EAAA,OAAO,CAAA,EAAG,IAAI,CAAA,CAAA,EAAI,CAAC,CAAA,CAAA;AACrB;AC3GO,SAAS,0BAA0B,IAAA,EAAoD;AAC5F,EAAA,IAAI,IAAA,CAAK,sBAAsB,KAAA,EAAO;AACpC,IAAAA,gBAAA,CAAI,iBAAA,GAAoB,KAAA;AAAA,EAC1B;AAEA,EAAA,IAAI,eAAA,GAA2C,IAAA;AAC/C,EAAA,IAAI,iBAAA,GAAmC,IAAA;AAEvC,EAAA,SAAS,WAAA,GAAgC;AACvC,IAAA,IAAI,CAAC,eAAA,EAAiB;AACpB,MAAA,MAAM,SAAA,GAAqC,EAAE,KAAA,EAAO,IAAA,CAAK,SAAS,MAAA,EAAO;AACzE,MAAA,IAAI,IAAA,CAAK,MAAA,KAAW,MAAA,EAAW,SAAA,CAAU,SAAS,IAAA,CAAK,MAAA;AACvD,MAAA,eAAA,GAAkBC,qBAAA,CAAS,oBAAA,EAAsB,IAAA,CAAK,KAAA,EAAO,SAAkB,CAAA;AAAA,IACjF;AACA,IAAA,OAAO,eAAA;AAAA,EACT;AAEA,EAAA,MAAM,OAAA,GAA4B;AAAA,IAChC,OAAO,IAAA,CAAK,KAAA;AAAA,IACZ,aAAA,EAAe,KAAK,aAAA,IAAiB,MAAA;AAAA,IACrC,MAAA,EAAQ,KAAK,MAAA,IAAU,QAAA;AAAA,IACvB,UAAA,EAAY,IAAA;AAAA,IACZ,MAAM,MAAM,KAAA,EAA2C;AACrD,MAAA,IAAI,KAAA,CAAM,WAAW,CAAA,EAAG;AACtB,QAAA,OAAO,EAAE,OAAA,EAAS,IAAI,SAAA,EAAW,IAAA,CAAK,aAAa,CAAA,EAAE;AAAA,MACvD;AACA,MAAA,MAAM,IAAA,GAAQ,MAAM,WAAA,EAAY;AAIhC,MAAA,MAAM,MAAA,GAAS,MAAM,IAAA,CAAK,KAAA,EAAO,EAAE,OAAA,EAAS,MAAA,EAAQ,SAAA,EAAW,IAAA,EAAM,CAAA;AACrE,MAAA,MAAM,OAAO,MAAA,CAAO,IAAA;AACpB,MAAA,MAAM,YAAY,MAAA,CAAO,IAAA,CAAK,MAAA,CAAO,IAAA,CAAK,SAAS,CAAC,CAAA;AACpD,MAAA,IAAI,iBAAA,KAAsB,IAAA,IAAQ,iBAAA,KAAsB,SAAA,EAAW;AACjE,QAAA,MAAM,IAAI,KAAA,CAAM,CAAA,0CAAA,EAA6C,iBAAiB,CAAA,MAAA,EAAS,SAAS,CAAA,CAAE,CAAA;AAAA,MACpG;AACA,MAAA,iBAAA,GAAoB,SAAA;AACpB,MAAA,MAAM,UAA0B,EAAC;AACjC,MAAA,KAAA,IAAS,IAAI,CAAA,EAAG,CAAA,GAAI,KAAA,CAAM,MAAA,EAAQ,KAAK,CAAA,EAAG;AACxC,QAAA,OAAA,CAAQ,IAAA,CAAK,IAAI,YAAA,CAAa,IAAA,CAAK,MAAA,EAAQ,IAAA,CAAK,UAAA,GAAa,CAAA,GAAI,SAAA,GAAY,CAAA,EAAG,SAAS,CAAA,CAAE,OAAO,CAAA;AAAA,MACpG;AACA,MAAA,OAAO,EAAE,SAAS,SAAA,EAAU;AAAA,IAC9B;AAAA,GACF;AACA,EAAA,IAAI,IAAA,CAAK,SAAA,KAAc,MAAA,EAAW,OAAA,CAAQ,YAAY,IAAA,CAAK,SAAA;AAC3D,EAAA,OAAO,OAAA;AACT;;;ACvCO,IAAM,aAAA,GAAgB;AACtB,IAAM,uBAAA,GAA0B;;;ACNvC,eAAsB,KAAA,CAAM,QAAA,EAAkB,IAAA,GAAqB,EAAC,EAA+B;AACjG,EAAA,MAAM,MAAA,GAAS,cAAc,QAAA,EAAU,EAAE,MAAM,IAAA,CAAK,QAAA,IAAY,WAAW,CAAA;AAC3E,EAAA,MAAM,OAAA,GACJ,KAAK,OAAA,IACL,yBAAA;AAAA,IACE,KAAK,aAAA,KAAkB,MAAA,GACnB,EAAE,KAAA,EAAO,KAAK,KAAA,IAAS,aAAA,EAAe,aAAA,EAAe,IAAA,CAAK,eAAc,GACxE,EAAE,KAAA,EAAO,IAAA,CAAK,SAAS,aAAA;AAAc,GAC3C;AAEF,EAAA,MAAM,MAAA,GAAS,MAAM,OAAA,CAAQ,KAAA,CAAM,MAAA,CAAO,IAAI,CAAC,CAAA,KAAM,CAAA,CAAE,IAAI,CAAC,CAAA;AAC5D,EAAA,IAAI,MAAA,CAAO,OAAA,CAAQ,MAAA,KAAW,MAAA,CAAO,MAAA,EAAQ;AAC3C,IAAA,MAAM,IAAI,MAAM,CAAA,iBAAA,EAAoB,MAAA,CAAO,QAAQ,MAAM,CAAA,aAAA,EAAgB,MAAA,CAAO,MAAM,CAAA,OAAA,CAAS,CAAA;AAAA,EACjG;AAEA,EAAA,MAAM,eAAA,GAAoC,MAAA,CAAO,GAAA,CAAI,CAAC,GAAG,CAAA,MAAO;AAAA,IAC9D,IAAI,CAAA,CAAE,EAAA;AAAA,IACN,YAAY,CAAA,CAAE,UAAA;AAAA,IACd,YAAY,CAAA,CAAE,UAAA;AAAA,IACd,MAAA,EAAQ,MAAA,CAAO,OAAA,CAAQ,CAAC;AAAA,GAC1B,CAAE,CAAA;AAEF,EAAA,MAAM,KAAA,GAAwB;AAAA,IAC5B,OAAO,OAAA,CAAQ,KAAA;AAAA,IACf,eAAe,OAAA,CAAQ,aAAA;AAAA,IACvB,WAAW,MAAA,CAAO,SAAA;AAAA,IAClB,QAAQ,OAAA,CAAQ,MAAA;AAAA,IAChB,YAAY,OAAA,CAAQ,UAAA;AAAA,IACpB,QAAA,EAAU,KAAK,QAAA,IAAY,SAAA;AAAA,IAC3B,MAAA,EAAQ;AAAA,GACV;AAEA,EAAA,OAAO,EAAE,aAAA,EAAe,CAAA,EAAG,MAAA,EAAQ,CAAC,KAAK,CAAA,EAAE;AAC7C;;;AC/BO,SAAS,cAAA,CACd,OAAA,EACA,WAAA,EACA,IAAA,GAAsB,EAAC,EACV;AACb,EAAA,MAAM,KAAA,GAAQ,SAAA,CAAU,OAAA,EAAS,IAAA,CAAK,KAAK,CAAA;AAC3C,EAAA,IAAI,CAAC,KAAA,EAAO,MAAM,IAAI,MAAM,mCAAmC,CAAA;AAC/D,EAAA,IAAI,WAAA,CAAY,MAAA,KAAW,KAAA,CAAM,SAAA,EAAW;AAC1C,IAAA,MAAM,IAAI,KAAA,CAAM,CAAA,uBAAA,EAA0B,WAAA,CAAY,MAAM,CAAA,sBAAA,EAAyB,KAAA,CAAM,KAAK,CAAA,EAAA,EAAK,KAAA,CAAM,SAAS,CAAA,CAAA,CAAG,CAAA;AAAA,EACzH;AAEA,EAAA,MAAM,CAAA,GAAI,KAAK,CAAA,IAAK,CAAA;AACpB,EAAA,MAAM,MAAA,GAAsB,KAAA,CAAM,MAAA,CAAO,GAAA,CAAI,CAAC,CAAA,MAAO;AAAA,IACnD,YAAY,KAAA,CAAM,KAAA;AAAA,IAClB,SAAS,CAAA,CAAE,EAAA;AAAA,IACX,YAAY,CAAA,CAAE,UAAA;AAAA,IACd,YAAY,CAAA,CAAE,UAAA;AAAA,IACd,OAAO,UAAA,CAAW,WAAA,EAAa,CAAA,CAAE,MAAA,EAAQ,MAAM,MAAM;AAAA,GACvD,CAAE,CAAA;AAGF,EAAA,MAAM,KAAA,GAAQ,KAAA,CAAM,MAAA,KAAW,WAAA,GAAc,CAAA,GAAI,EAAA;AACjD,EAAA,MAAA,CAAO,IAAA,CAAK,CAAC,CAAA,EAAG,CAAA,KAAM,SAAS,CAAA,CAAE,KAAA,GAAQ,EAAE,KAAA,CAAM,CAAA;AACjD,EAAA,OAAO,MAAA,CAAO,KAAA,CAAM,CAAA,EAAG,CAAC,CAAA;AAC1B;AAEA,SAAS,SAAA,CAAU,SAA4B,SAAA,EAAgD;AAC7F,EAAA,IAAI,SAAA,SAAkB,OAAA,CAAQ,MAAA,CAAO,KAAK,CAAC,CAAA,KAAM,CAAA,CAAE,KAAA,KAAU,SAAS,CAAA;AACtE,EAAA,OAAO,OAAA,CAAQ,OAAO,CAAC,CAAA;AACzB;AAEA,SAAS,UAAA,CAAW,CAAA,EAAiB,CAAA,EAAiB,MAAA,EAAiC;AACrF,EAAA,IAAI,WAAW,WAAA,EAAa;AAC1B,IAAA,IAAI,GAAA,GAAM,CAAA;AACV,IAAA,KAAA,IAAS,IAAI,CAAA,EAAG,CAAA,GAAI,CAAA,CAAE,MAAA,EAAQ,KAAK,CAAA,EAAG;AACpC,MAAA,MAAM,CAAA,GAAI,CAAA,CAAE,CAAC,CAAA,GAAK,EAAE,CAAC,CAAA;AACrB,MAAA,GAAA,IAAO,CAAA,GAAI,CAAA;AAAA,IACb;AACA,IAAA,OAAO,IAAA,CAAK,KAAK,GAAG,CAAA;AAAA,EACtB;AACA,EAAA,IAAI,GAAA,GAAM,CAAA;AACV,EAAA,KAAA,IAAS,CAAA,GAAI,CAAA,EAAG,CAAA,GAAI,CAAA,CAAE,MAAA,EAAQ,CAAA,IAAK,CAAA,EAAG,GAAA,IAAO,CAAA,CAAE,CAAC,CAAA,GAAK,CAAA,CAAE,CAAC,CAAA;AACxD,EAAA,IAAI,MAAA,KAAW,OAAO,OAAO,GAAA;AAE7B,EAAA,IAAI,EAAA,GAAK,CAAA;AACT,EAAA,IAAI,EAAA,GAAK,CAAA;AACT,EAAA,KAAA,IAAS,IAAI,CAAA,EAAG,CAAA,GAAI,CAAA,CAAE,MAAA,EAAQ,KAAK,CAAA,EAAG;AACpC,IAAA,EAAA,IAAM,CAAA,CAAE,CAAC,CAAA,GAAK,CAAA,CAAE,CAAC,CAAA;AACjB,IAAA,EAAA,IAAM,CAAA,CAAE,CAAC,CAAA,GAAK,CAAA,CAAE,CAAC,CAAA;AAAA,EACnB;AACA,EAAA,MAAM,QAAQ,IAAA,CAAK,IAAA,CAAK,EAAE,CAAA,GAAI,IAAA,CAAK,KAAK,EAAE,CAAA;AAC1C,EAAA,OAAO,KAAA,KAAU,CAAA,GAAI,CAAA,GAAI,GAAA,GAAM,KAAA;AACjC;;;AClDA,IAAM,gBAAA,GAAmB,mDAAA;AAElB,SAAS,yBAAyB,IAAA,EAAmD;AAC1F,EAAA,MAAM,QAAQ,IAAA,CAAK,KAAA,IAAS,QAAQ,GAAA,CAAI,QAAA,IAAY,QAAQ,GAAA,CAAI,iBAAA;AAChE,EAAA,IAAI,CAAC,KAAA,EAAO;AACV,IAAA,MAAM,IAAI,MAAM,0EAA0E,CAAA;AAAA,EAC5F;AACA,EAAA,MAAM,OAAA,GAAU,KAAK,OAAA,IAAW,gBAAA;AAChC,EAAA,MAAM,MAAM,CAAA,EAAG,OAAO,IAAI,SAAA,CAAU,IAAA,CAAK,KAAK,CAAC,CAAA,4BAAA,CAAA;AAE/C,EAAA,MAAM,OAAA,GAA4B;AAAA,IAChC,OAAO,IAAA,CAAK,KAAA;AAAA,IACZ,aAAA,EAAe,KAAK,aAAA,IAAiB,MAAA;AAAA,IACrC,MAAA,EAAQ,KAAK,MAAA,IAAU,QAAA;AAAA,IACvB,UAAA,EAAY,IAAA;AAAA,IACZ,MAAM,MAAM,KAAA,EAA2C;AACrD,MAAA,IAAI,KAAA,CAAM,WAAW,CAAA,EAAG;AACtB,QAAA,OAAO,EAAE,OAAA,EAAS,IAAI,SAAA,EAAW,IAAA,CAAK,aAAa,CAAA,EAAE;AAAA,MACvD;AACA,MAAA,MAAM,IAAA,GAAO,IAAA,CAAK,SAAA,CAAU,EAAE,MAAA,EAAQ,KAAA,EAAO,OAAA,EAAS,EAAE,cAAA,EAAgB,IAAA,EAAK,EAAG,CAAA;AAChF,MAAA,MAAM,GAAA,GAAM,MAAM,KAAA,CAAM,GAAA,EAAK;AAAA,QAC3B,MAAA,EAAQ,MAAA;AAAA,QACR,OAAA,EAAS;AAAA,UACP,cAAA,EAAgB,kBAAA;AAAA,UAChB,aAAA,EAAe,UAAU,KAAK,CAAA;AAAA,SAChC;AAAA,QACA;AAAA,OACD,CAAA;AACD,MAAA,IAAI,CAAC,IAAI,EAAA,EAAI;AACX,QAAA,MAAM,SAAS,MAAM,GAAA,CAAI,MAAK,CAAE,KAAA,CAAM,MAAM,EAAE,CAAA;AAC9C,QAAA,MAAM,IAAI,KAAA,CAAM,CAAA,iBAAA,EAAoB,GAAA,CAAI,MAAM,CAAA,KAAA,EAAQ,IAAA,CAAK,KAAK,CAAA,EAAA,EAAK,MAAA,IAAU,GAAA,CAAI,UAAU,CAAA,CAAE,CAAA;AAAA,MACjG;AACA,MAAA,MAAM,GAAA,GAAO,MAAM,GAAA,CAAI,IAAA,EAAK;AAC5B,MAAA,MAAM,MAAA,GAAS,eAAA,CAAgB,GAAA,EAAK,KAAA,CAAM,MAAM,CAAA;AAChD,MAAA,MAAM,YAAY,MAAA,CAAO,CAAC,CAAA,EAAG,MAAA,IAAU,KAAK,SAAA,IAAa,CAAA;AACzD,MAAA,MAAM,OAAA,GAAU,MAAA,CAAO,GAAA,CAAI,CAAC,CAAA,KAAM,UAAU,YAAA,CAAa,IAAA,CAAK,CAAC,CAAC,CAAC,CAAA;AACjE,MAAA,OAAO,EAAE,SAAS,SAAA,EAAU;AAAA,IAC9B;AAAA,GACF;AACA,EAAA,IAAI,IAAA,CAAK,SAAA,KAAc,MAAA,EAAW,OAAA,CAAQ,YAAY,IAAA,CAAK,SAAA;AAC3D,EAAA,OAAO,OAAA;AACT;AAWA,SAAS,eAAA,CAAgB,KAAc,aAAA,EAAmC;AACxE,EAAA,IAAI,CAAC,KAAA,CAAM,OAAA,CAAQ,GAAG,CAAA,EAAG;AACvB,IAAA,MAAM,IAAI,MAAM,2CAA2C,CAAA;AAAA,EAC7D;AACA,EAAA,IAAI,GAAA,CAAI,MAAA,KAAW,CAAA,EAAG,OAAO,EAAC;AAE9B,EAAA,MAAM,KAAA,GAAQ,IAAI,CAAC,CAAA;AACnB,EAAA,IAAI,OAAO,UAAU,QAAA,EAAU;AAC7B,IAAA,IAAI,kBAAkB,CAAA,EAAG;AACvB,MAAA,MAAM,IAAI,KAAA,CAAM,CAAA,yCAAA,EAA4C,aAAa,CAAA,CAAE,CAAA;AAAA,IAC7E;AACA,IAAA,OAAO,CAAC,GAAe,CAAA;AAAA,EACzB;AACA,EAAA,IAAI,KAAA,CAAM,OAAA,CAAQ,KAAK,CAAA,KAAM,KAAA,CAAM,MAAA,KAAW,CAAA,IAAK,OAAO,KAAA,CAAM,CAAC,CAAA,KAAM,QAAA,CAAA,EAAW;AAChF,IAAA,OAAO,GAAA;AAAA,EACT;AACA,EAAA,IAAI,KAAA,CAAM,QAAQ,KAAK,CAAA,IAAK,MAAM,OAAA,CAAQ,KAAA,CAAM,CAAC,CAAC,CAAA,EAAG;AAEnD,IAAA,OAAQ,GAAA,CAAqB,IAAI,QAAQ,CAAA;AAAA,EAC3C;AACA,EAAA,MAAM,IAAI,MAAM,+CAA+C,CAAA;AACjE;AAEA,SAAS,SAAS,MAAA,EAA8B;AAC9C,EAAA,IAAI,MAAA,CAAO,MAAA,KAAW,CAAA,EAAG,OAAO,EAAC;AACjC,EAAA,MAAM,GAAA,GAAM,MAAA,CAAO,CAAC,CAAA,CAAG,MAAA;AACvB,EAAA,MAAM,MAAM,IAAI,KAAA,CAAc,GAAG,CAAA,CAAE,KAAK,CAAC,CAAA;AACzC,EAAA,KAAA,MAAW,KAAK,MAAA,EAAQ;AACtB,IAAA,KAAA,IAAS,CAAA,GAAI,CAAA,EAAG,CAAA,GAAI,GAAA,EAAK,CAAA,IAAK,GAAG,GAAA,CAAI,CAAC,CAAA,IAAM,CAAA,CAAE,CAAC,CAAA;AAAA,EACjD;AACA,EAAA,KAAA,IAAS,CAAA,GAAI,GAAG,CAAA,GAAI,GAAA,EAAK,KAAK,CAAA,EAAG,GAAA,CAAI,CAAC,CAAA,IAAM,MAAA,CAAO,MAAA;AACnD,EAAA,OAAO,GAAA;AACT;AAEA,SAAS,UAAU,CAAA,EAA+B;AAChD,EAAA,IAAI,GAAA,GAAM,CAAA;AACV,EAAA,KAAA,IAAS,CAAA,GAAI,CAAA,EAAG,CAAA,GAAI,CAAA,CAAE,MAAA,EAAQ,CAAA,IAAK,CAAA,EAAG,GAAA,IAAO,CAAA,CAAE,CAAC,CAAA,GAAK,CAAA,CAAE,CAAC,CAAA;AACxD,EAAA,MAAM,IAAA,GAAO,IAAA,CAAK,IAAA,CAAK,GAAG,CAAA;AAC1B,EAAA,IAAI,IAAA,KAAS,GAAG,OAAO,CAAA;AACvB,EAAA,KAAA,IAAS,CAAA,GAAI,CAAA,EAAG,CAAA,GAAI,CAAA,CAAE,MAAA,EAAQ,CAAA,IAAK,CAAA,EAAG,CAAA,CAAE,CAAC,CAAA,GAAI,CAAA,CAAE,CAAC,CAAA,GAAK,IAAA;AACrD,EAAA,OAAO,CAAA;AACT","file":"index.cjs","sourcesContent":["/**\n * Section-based markdown chunker.\n *\n * Splits a markdown document on ATX headings (`#`, `##`, ...). Each chunk\n * carries the byte offset and length into the original UTF-8 source so a\n * downstream consumer can map a vector hit back to the exact substring\n * without re-tokenising. Pre-heading content becomes a \"preamble\" chunk.\n */\n\nexport type ChunkingMode = 'document' | 'section' | 'paragraph';\n\nexport interface MarkdownChunk {\n  id: string;\n  textOffset: number;\n  textLength: number;\n  text: string;\n}\n\nexport interface ChunkOptions {\n  mode?: ChunkingMode;\n}\n\nconst HEADING = /^(#{1,6})\\s+(.+?)\\s*$/;\n\nexport function chunkMarkdown(markdown: string, opts: ChunkOptions = {}): MarkdownChunk[] {\n  const mode = opts.mode ?? 'section';\n  if (mode === 'document') {\n    return [{ id: 'document', textOffset: 0, textLength: markdown.length, text: markdown }];\n  }\n  if (mode === 'paragraph') {\n    return paragraphChunks(markdown);\n  }\n  return sectionChunks(markdown);\n}\n\nfunction sectionChunks(markdown: string): MarkdownChunk[] {\n  const lines = splitWithOffsets(markdown);\n  const sections: MarkdownChunk[] = [];\n  let current: { id: string; start: number; end: number } | null = null;\n  const ids = new Set<string>();\n\n  function flush(end: number): void {\n    if (!current) return;\n    const text = markdown.slice(current.start, end);\n    if (text.trim().length === 0) {\n      current = null;\n      return;\n    }\n    sections.push({ id: current.id, textOffset: current.start, textLength: end - current.start, text });\n    current = null;\n  }\n\n  for (const line of lines) {\n    const match = HEADING.exec(line.text);\n    if (match) {\n      flush(line.offset);\n      const id = uniqueId(slugify(match[2] ?? `section-${sections.length + 1}`), ids);\n      ids.add(id);\n      current = { id, start: line.offset, end: line.offset + line.text.length };\n      continue;\n    }\n    if (current === null) {\n      const id = uniqueId('preamble', ids);\n      ids.add(id);\n      current = { id, start: line.offset, end: line.offset + line.text.length };\n    } else {\n      current.end = line.offset + line.text.length;\n    }\n  }\n  flush(markdown.length);\n\n  if (sections.length === 0) {\n    return [{ id: 'document', textOffset: 0, textLength: markdown.length, text: markdown }];\n  }\n  return sections;\n}\n\nfunction paragraphChunks(markdown: string): MarkdownChunk[] {\n  const out: MarkdownChunk[] = [];\n  const ids = new Set<string>();\n  let cursor = 0;\n  let i = 0;\n  while (cursor < markdown.length) {\n    let end = markdown.indexOf('\\n\\n', cursor);\n    if (end === -1) end = markdown.length;\n    const text = markdown.slice(cursor, end);\n    if (text.trim().length > 0) {\n      const id = uniqueId(slugify(text.split('\\n')[0] ?? `p-${i}`), ids);\n      ids.add(id);\n      out.push({ id, textOffset: cursor, textLength: text.length, text });\n      i += 1;\n    }\n    cursor = end + 2;\n  }\n  if (out.length === 0) {\n    return [{ id: 'document', textOffset: 0, textLength: markdown.length, text: markdown }];\n  }\n  return out;\n}\n\nfunction splitWithOffsets(s: string): { text: string; offset: number }[] {\n  const lines: { text: string; offset: number }[] = [];\n  let offset = 0;\n  for (const line of s.split('\\n')) {\n    const withNl = line + (offset + line.length < s.length ? '\\n' : '');\n    lines.push({ text: withNl, offset });\n    offset += withNl.length;\n  }\n  return lines;\n}\n\nfunction slugify(s: string): string {\n  return (\n    s\n      .toLowerCase()\n      .replace(/[^a-z0-9]+/g, '-')\n      .replace(/^-+|-+$/g, '')\n      .slice(0, 64) || 'section'\n  );\n}\n\nfunction uniqueId(base: string, taken: Set<string>): string {\n  if (!taken.has(base)) return base;\n  let n = 2;\n  while (taken.has(`${base}-${n}`)) n += 1;\n  return `${base}-${n}`;\n}\n","/**\n * transformers.js backend (browser + Node + Bun via @huggingface/transformers).\n *\n * Loads the model lazily and reuses the pipeline across calls. Models are\n * cached on disk by transformers.js itself (HF cache layout).\n */\n\nimport { pipeline, env } from '@huggingface/transformers';\nimport type { EmbeddingBackend, EmbeddingBackendOptions, EmbeddingMatrix } from './types.js';\n\ninterface TransformersBackendOptions extends EmbeddingBackendOptions {\n  /** Override Xenova quantisation. Defaults to fp32 for closest fidelity. */\n  dtype?: 'fp32' | 'fp16' | 'q8' | 'q4';\n  /** Force a backend device (e.g. 'cpu', 'gpu', 'wasm'). Defaults to auto. */\n  device?: 'cpu' | 'gpu' | 'wasm' | 'webgpu';\n  /** Allow remote model downloads. true by default. */\n  allowRemoteModels?: boolean;\n}\n\nexport function createTransformersBackend(opts: TransformersBackendOptions): EmbeddingBackend {\n  if (opts.allowRemoteModels === false) {\n    env.allowRemoteModels = false;\n  }\n\n  let pipelinePromise: Promise<unknown> | null = null;\n  let resolvedDimension: number | null = null;\n\n  function getPipeline(): Promise<unknown> {\n    if (!pipelinePromise) {\n      const modelOpts: Record<string, unknown> = { dtype: opts.dtype ?? 'fp32' };\n      if (opts.device !== undefined) modelOpts.device = opts.device;\n      pipelinePromise = pipeline('feature-extraction', opts.model, modelOpts as never);\n    }\n    return pipelinePromise;\n  }\n\n  const backend: EmbeddingBackend = {\n    model: opts.model,\n    modelRevision: opts.modelRevision ?? 'main',\n    metric: opts.metric ?? 'cosine',\n    normalized: true,\n    async embed(texts: string[]): Promise<EmbeddingMatrix> {\n      if (texts.length === 0) {\n        return { vectors: [], dimension: opts.dimension ?? 0 };\n      }\n      const pipe = (await getPipeline()) as (\n        texts: string[],\n        opts: { pooling: 'mean'; normalize: boolean },\n      ) => Promise<{ data: Float32Array; dims: number[] }>;\n      const tensor = await pipe(texts, { pooling: 'mean', normalize: true });\n      const data = tensor.data;\n      const dimension = tensor.dims[tensor.dims.length - 1] as number;\n      if (resolvedDimension !== null && resolvedDimension !== dimension) {\n        throw new Error(`Model emitted inconsistent dimension: had ${resolvedDimension}, now ${dimension}`);\n      }\n      resolvedDimension = dimension;\n      const vectors: Float32Array[] = [];\n      for (let i = 0; i < texts.length; i += 1) {\n        vectors.push(new Float32Array(data.buffer, data.byteOffset + i * dimension * 4, dimension).slice());\n      }\n      return { vectors, dimension };\n    },\n  };\n  if (opts.dimension !== undefined) backend.dimension = opts.dimension;\n  return backend;\n}\n","import type { EmbeddingMetric } from '@cvfile/sdk';\n\nexport interface EmbeddingMatrix {\n  vectors: Float32Array[];\n  dimension: number;\n}\n\nexport interface EmbeddingBackend {\n  model: string;\n  modelRevision: string;\n  /** Optional pre-declared dimension; the backend may override after first call. */\n  dimension?: number;\n  metric: EmbeddingMetric;\n  normalized: boolean;\n  embed(texts: string[]): Promise<EmbeddingMatrix>;\n}\n\nexport interface EmbeddingBackendOptions {\n  model: string;\n  modelRevision?: string;\n  /** Optional pre-known dimension (e.g. 1024 for bge-m3). */\n  dimension?: number;\n  metric?: EmbeddingMetric;\n}\n\n/** Recommended default per spec §5: BAAI BGE-M3, MIT, multilingual, 1024-dim. */\nexport const DEFAULT_MODEL = 'Xenova/bge-m3';\nexport const DEFAULT_MODEL_DIMENSION = 1024;\n","/**\n * High-level embed() API: markdown in, EmbeddingsPayload out, ready to drop\n * into pack({ embeddings: ... }).\n */\n\nimport type { EmbeddingChunk, EmbeddingSpace, EmbeddingsPayload } from '@cvfile/sdk';\nimport { chunkMarkdown, type ChunkingMode } from './chunk.js';\nimport { createTransformersBackend } from './transformers-backend.js';\nimport { DEFAULT_MODEL, type EmbeddingBackend } from './types.js';\n\nexport interface EmbedOptions {\n  /** HF model id; defaults to BGE-M3 (Xenova/bge-m3). */\n  model?: string;\n  /** Pinned model revision; recorded in the payload for reproducibility. */\n  modelRevision?: string;\n  /** Chunking strategy; default 'section'. */\n  chunking?: ChunkingMode;\n  /** Bring-your-own backend (e.g. an OpenAI/Voyage adapter). */\n  backend?: EmbeddingBackend;\n}\n\nexport async function embed(markdown: string, opts: EmbedOptions = {}): Promise<EmbeddingsPayload> {\n  const chunks = chunkMarkdown(markdown, { mode: opts.chunking ?? 'section' });\n  const backend =\n    opts.backend ??\n    createTransformersBackend(\n      opts.modelRevision !== undefined\n        ? { model: opts.model ?? DEFAULT_MODEL, modelRevision: opts.modelRevision }\n        : { model: opts.model ?? DEFAULT_MODEL },\n    );\n\n  const matrix = await backend.embed(chunks.map((c) => c.text));\n  if (matrix.vectors.length !== chunks.length) {\n    throw new Error(`Backend returned ${matrix.vectors.length} vectors for ${chunks.length} chunks`);\n  }\n\n  const embeddingChunks: EmbeddingChunk[] = chunks.map((c, i) => ({\n    id: c.id,\n    textOffset: c.textOffset,\n    textLength: c.textLength,\n    vector: matrix.vectors[i]!,\n  }));\n\n  const space: EmbeddingSpace = {\n    model: backend.model,\n    modelRevision: backend.modelRevision,\n    dimension: matrix.dimension,\n    metric: backend.metric,\n    normalized: backend.normalized,\n    chunking: opts.chunking ?? 'section',\n    chunks: embeddingChunks,\n  };\n\n  return { formatVersion: 1, spaces: [space] };\n}\n","/**\n * Pure similarity search over an EmbeddingsPayload. Vector input is the\n * caller's responsibility (encode the query with the same model that was\n * used to populate the space, ideally pulled from space.model).\n */\n\nimport type { EmbeddingMetric, EmbeddingSpace, EmbeddingsPayload } from '@cvfile/sdk';\n\nexport interface SearchHit {\n  spaceModel: string;\n  chunkId: string;\n  textOffset: number;\n  textLength: number;\n  score: number;\n}\n\nexport interface SearchOptions {\n  /** Restrict to a specific embedding space; default = first space. */\n  model?: string;\n  /** Top-k results; default 5. */\n  k?: number;\n}\n\nexport function searchSemantic(\n  payload: EmbeddingsPayload,\n  queryVector: Float32Array,\n  opts: SearchOptions = {},\n): SearchHit[] {\n  const space = pickSpace(payload, opts.model);\n  if (!space) throw new Error('No matching embedding space found');\n  if (queryVector.length !== space.dimension) {\n    throw new Error(`Query vector dimension ${queryVector.length} does not match space ${space.model} (${space.dimension})`);\n  }\n\n  const k = opts.k ?? 5;\n  const scored: SearchHit[] = space.chunks.map((c) => ({\n    spaceModel: space.model,\n    chunkId: c.id,\n    textOffset: c.textOffset,\n    textLength: c.textLength,\n    score: similarity(queryVector, c.vector, space.metric),\n  }));\n\n  // Higher score = better for cosine/dot; lower for euclidean.\n  const order = space.metric === 'euclidean' ? 1 : -1;\n  scored.sort((a, b) => order * (a.score - b.score));\n  return scored.slice(0, k);\n}\n\nfunction pickSpace(payload: EmbeddingsPayload, modelHint?: string): EmbeddingSpace | undefined {\n  if (modelHint) return payload.spaces.find((s) => s.model === modelHint);\n  return payload.spaces[0];\n}\n\nfunction similarity(a: Float32Array, b: Float32Array, metric: EmbeddingMetric): number {\n  if (metric === 'euclidean') {\n    let sum = 0;\n    for (let i = 0; i < a.length; i += 1) {\n      const d = a[i]! - b[i]!;\n      sum += d * d;\n    }\n    return Math.sqrt(sum);\n  }\n  let dot = 0;\n  for (let i = 0; i < a.length; i += 1) dot += a[i]! * b[i]!;\n  if (metric === 'dot') return dot;\n  // cosine: assume normalized vectors when produced by our backend\n  let na = 0;\n  let nb = 0;\n  for (let i = 0; i < a.length; i += 1) {\n    na += a[i]! * a[i]!;\n    nb += b[i]! * b[i]!;\n  }\n  const denom = Math.sqrt(na) * Math.sqrt(nb);\n  return denom === 0 ? 0 : dot / denom;\n}\n","/**\n * Hugging Face Inference API backend.\n *\n * Calls https://router.huggingface.co/hf-inference/models/<model>/pipeline/feature-extraction\n * with HF_TOKEN. For sentence-transformers models (BGE-M3, MiniLM, etc.) the\n * response is already mean-pooled per input — one vector per text. We\n * normalise client-side so cosine math is consistent across backends.\n */\n\nimport type { EmbeddingBackend, EmbeddingMatrix } from './types.js';\nimport type { EmbeddingMetric } from '@cvfile/sdk';\n\nexport interface HuggingFaceBackendOptions {\n  model: string;\n  /** HF token. Defaults to `process.env.HF_TOKEN`. */\n  token?: string;\n  /** Pinned revision; recorded in the payload. Default 'main'. */\n  modelRevision?: string;\n  /** Pre-known dimension. Optional; inferred from first response otherwise. */\n  dimension?: number;\n  metric?: EmbeddingMetric;\n  /** Override base URL (e.g. for self-hosted TEI). */\n  baseUrl?: string;\n}\n\nconst DEFAULT_BASE_URL = 'https://router.huggingface.co/hf-inference/models';\n\nexport function createHuggingFaceBackend(opts: HuggingFaceBackendOptions): EmbeddingBackend {\n  const token = opts.token ?? process.env.HF_TOKEN ?? process.env.HUGGINGFACE_TOKEN;\n  if (!token) {\n    throw new Error('HF_TOKEN (or HUGGINGFACE_TOKEN) is required for the Hugging Face backend');\n  }\n  const baseUrl = opts.baseUrl ?? DEFAULT_BASE_URL;\n  const url = `${baseUrl}/${encodeURI(opts.model)}/pipeline/feature-extraction`;\n\n  const backend: EmbeddingBackend = {\n    model: opts.model,\n    modelRevision: opts.modelRevision ?? 'main',\n    metric: opts.metric ?? 'cosine',\n    normalized: true,\n    async embed(texts: string[]): Promise<EmbeddingMatrix> {\n      if (texts.length === 0) {\n        return { vectors: [], dimension: opts.dimension ?? 0 };\n      }\n      const body = JSON.stringify({ inputs: texts, options: { wait_for_model: true } });\n      const res = await fetch(url, {\n        method: 'POST',\n        headers: {\n          'content-type': 'application/json',\n          authorization: `Bearer ${token}`,\n        },\n        body,\n      });\n      if (!res.ok) {\n        const detail = await res.text().catch(() => '');\n        throw new Error(`HF Inference API ${res.status} for ${opts.model}: ${detail || res.statusText}`);\n      }\n      const raw = (await res.json()) as unknown;\n      const matrix = parseHfResponse(raw, texts.length);\n      const dimension = matrix[0]?.length ?? opts.dimension ?? 0;\n      const vectors = matrix.map((v) => normalize(Float32Array.from(v)));\n      return { vectors, dimension };\n    },\n  };\n  if (opts.dimension !== undefined) backend.dimension = opts.dimension;\n  return backend;\n}\n\n/**\n * Coerce the variety of shapes the HF Inference API returns into a flat\n * `number[][]`: one mean-pooled vector per input.\n *\n * Observed shapes:\n *   - sentence-transformers (BGE-M3, MiniLM): `[[...vec], [...vec]]`\n *   - feature-extraction without pooling: `[[[...token0], [...token1], ...], ...]`\n *   - single-input convenience form: `[...vec]`\n */\nfunction parseHfResponse(raw: unknown, expectedCount: number): number[][] {\n  if (!Array.isArray(raw)) {\n    throw new Error('HF Inference API: expected array response');\n  }\n  if (raw.length === 0) return [];\n\n  const first = raw[0];\n  if (typeof first === 'number') {\n    if (expectedCount !== 1) {\n      throw new Error(`HF Inference API: got 1 vector, expected ${expectedCount}`);\n    }\n    return [raw as number[]];\n  }\n  if (Array.isArray(first) && (first.length === 0 || typeof first[0] === 'number')) {\n    return raw as number[][];\n  }\n  if (Array.isArray(first) && Array.isArray(first[0])) {\n    // token-level embeddings: mean-pool per input\n    return (raw as number[][][]).map(meanPool);\n  }\n  throw new Error('HF Inference API: unrecognised response shape');\n}\n\nfunction meanPool(tokens: number[][]): number[] {\n  if (tokens.length === 0) return [];\n  const dim = tokens[0]!.length;\n  const out = new Array<number>(dim).fill(0);\n  for (const t of tokens) {\n    for (let i = 0; i < dim; i += 1) out[i]! += t[i]!;\n  }\n  for (let i = 0; i < dim; i += 1) out[i]! /= tokens.length;\n  return out;\n}\n\nfunction normalize(v: Float32Array): Float32Array {\n  let sum = 0;\n  for (let i = 0; i < v.length; i += 1) sum += v[i]! * v[i]!;\n  const norm = Math.sqrt(sum);\n  if (norm === 0) return v;\n  for (let i = 0; i < v.length; i += 1) v[i] = v[i]! / norm;\n  return v;\n}\n"]}
+{"version":3,"sources":["../src/chunk.ts","../src/transformers-backend.ts","../src/types.ts","../src/embed.ts","../src/search.ts","../src/huggingface-backend.ts"],"names":["env","pipeline"],"mappings":";;;;;AAgBA,IAAM,OAAA,GAAU,IAAI,WAAA,EAAY;AAChC,IAAM,OAAA,GAAU,IAAI,WAAA,EAAY;AAehC,IAAM,OAAA,GAAU,uBAAA;AAST,SAAS,aAAA,CAAc,QAAA,EAAkB,IAAA,GAAqB,EAAC,EAAoB;AACxF,EAAA,MAAM,IAAA,GAAO,KAAK,IAAA,IAAQ,SAAA;AAC1B,EAAA,MAAM,KAAA,GAAQ,OAAA,CAAQ,MAAA,CAAO,QAAQ,CAAA;AACrC,EAAA,IAAI,SAAS,UAAA,EAAY;AACvB,IAAA,OAAO,CAAC,aAAA,CAAc,KAAK,CAAC,CAAA;AAAA,EAC9B;AACA,EAAA,IAAI,SAAS,WAAA,EAAa;AACxB,IAAA,OAAO,gBAAgB,KAAK,CAAA;AAAA,EAC9B;AACA,EAAA,OAAO,cAAc,KAAK,CAAA;AAC5B;AAEA,SAAS,cAAc,KAAA,EAAkC;AACvD,EAAA,OAAO,EAAE,EAAA,EAAI,UAAA,EAAY,UAAA,EAAY,GAAG,UAAA,EAAY,KAAA,CAAM,UAAA,EAAY,IAAA,EAAM,SAAA,CAAU,KAAA,EAAO,CAAA,EAAG,KAAA,CAAM,UAAU,CAAA,EAAE;AACpH;AAEA,SAAS,cAAc,KAAA,EAAoC;AACzD,EAAA,MAAM,KAAA,GAAQ,qBAAqB,KAAK,CAAA;AACxC,EAAA,MAAM,WAA4B,EAAC;AACnC,EAAA,IAAI,OAAA,GAA6D,IAAA;AACjE,EAAA,MAAM,GAAA,uBAAU,GAAA,EAAY;AAE5B,EAAA,SAAS,MAAM,GAAA,EAAmB;AAChC,IAAA,IAAI,CAAC,OAAA,EAAS;AACd,IAAA,MAAM,IAAA,GAAO,SAAA,CAAU,KAAA,EAAO,OAAA,CAAQ,OAAO,GAAG,CAAA;AAChD,IAAA,IAAI,IAAA,CAAK,IAAA,EAAK,CAAE,MAAA,KAAW,CAAA,EAAG;AAC5B,MAAA,OAAA,GAAU,IAAA;AACV,MAAA;AAAA,IACF;AACA,IAAA,QAAA,CAAS,IAAA,CAAK,EAAE,EAAA,EAAI,OAAA,CAAQ,EAAA,EAAI,UAAA,EAAY,OAAA,CAAQ,KAAA,EAAO,UAAA,EAAY,GAAA,GAAM,OAAA,CAAQ,KAAA,EAAO,MAAM,CAAA;AAClG,IAAA,OAAA,GAAU,IAAA;AAAA,EACZ;AAEA,EAAA,KAAA,MAAW,QAAQ,KAAA,EAAO;AACxB,IAAA,MAAM,KAAA,GAAQ,OAAA,CAAQ,IAAA,CAAK,IAAA,CAAK,IAAI,CAAA;AACpC,IAAA,MAAM,OAAA,GAAU,IAAA,CAAK,MAAA,GAAS,IAAA,CAAK,UAAA;AACnC,IAAA,IAAI,KAAA,EAAO;AACT,MAAA,KAAA,CAAM,KAAK,MAAM,CAAA;AACjB,MAAA,MAAM,EAAA,GAAK,QAAA,CAAS,OAAA,CAAQ,KAAA,CAAM,CAAC,CAAA,IAAK,CAAA,QAAA,EAAW,QAAA,CAAS,MAAA,GAAS,CAAC,CAAA,CAAE,CAAA,EAAG,GAAG,CAAA;AAC9E,MAAA,GAAA,CAAI,IAAI,EAAE,CAAA;AACV,MAAA,OAAA,GAAU,EAAE,EAAA,EAAI,KAAA,EAAO,IAAA,CAAK,MAAA,EAAQ,KAAK,OAAA,EAAQ;AACjD,MAAA;AAAA,IACF;AACA,IAAA,IAAI,YAAY,IAAA,EAAM;AACpB,MAAA,MAAM,EAAA,GAAK,QAAA,CAAS,UAAA,EAAY,GAAG,CAAA;AACnC,MAAA,GAAA,CAAI,IAAI,EAAE,CAAA;AACV,MAAA,OAAA,GAAU,EAAE,EAAA,EAAI,KAAA,EAAO,IAAA,CAAK,MAAA,EAAQ,KAAK,OAAA,EAAQ;AAAA,IACnD,CAAA,MAAO;AACL,MAAA,OAAA,CAAQ,GAAA,GAAM,OAAA;AAAA,IAChB;AAAA,EACF;AACA,EAAA,KAAA,CAAM,MAAM,UAAU,CAAA;AAEtB,EAAA,IAAI,QAAA,CAAS,WAAW,CAAA,EAAG;AACzB,IAAA,OAAO,CAAC,aAAA,CAAc,KAAK,CAAC,CAAA;AAAA,EAC9B;AACA,EAAA,OAAO,QAAA;AACT;AAEA,SAAS,gBAAgB,KAAA,EAAoC;AAC3D,EAAA,MAAM,MAAuB,EAAC;AAC9B,EAAA,MAAM,GAAA,uBAAU,GAAA,EAAY;AAC5B,EAAA,MAAM,SAAA,GAAY,OAAA,CAAQ,MAAA,CAAO,MAAM,CAAA;AACvC,EAAA,IAAI,MAAA,GAAS,CAAA;AACb,EAAA,IAAI,CAAA,GAAI,CAAA;AACR,EAAA,OAAO,MAAA,GAAS,MAAM,UAAA,EAAY;AAChC,IAAA,IAAI,GAAA,GAAM,YAAA,CAAa,KAAA,EAAO,SAAA,EAAW,MAAM,CAAA;AAC/C,IAAA,IAAI,GAAA,KAAQ,EAAA,EAAI,GAAA,GAAM,KAAA,CAAM,UAAA;AAC5B,IAAA,MAAM,IAAA,GAAO,SAAA,CAAU,KAAA,EAAO,MAAA,EAAQ,GAAG,CAAA;AACzC,IAAA,IAAI,IAAA,CAAK,IAAA,EAAK,CAAE,MAAA,GAAS,CAAA,EAAG;AAC1B,MAAA,MAAM,EAAA,GAAK,QAAA,CAAS,OAAA,CAAQ,IAAA,CAAK,KAAA,CAAM,IAAI,CAAA,CAAE,CAAC,CAAA,IAAK,CAAA,EAAA,EAAK,CAAC,CAAA,CAAE,GAAG,GAAG,CAAA;AACjE,MAAA,GAAA,CAAI,IAAI,EAAE,CAAA;AACV,MAAA,GAAA,CAAI,IAAA,CAAK,EAAE,EAAA,EAAI,UAAA,EAAY,QAAQ,UAAA,EAAY,GAAA,GAAM,MAAA,EAAQ,IAAA,EAAM,CAAA;AACnE,MAAA,CAAA,IAAK,CAAA;AAAA,IACP;AACA,IAAA,MAAA,GAAS,MAAM,SAAA,CAAU,UAAA;AAAA,EAC3B;AACA,EAAA,IAAI,GAAA,CAAI,WAAW,CAAA,EAAG;AACpB,IAAA,OAAO,CAAC,aAAA,CAAc,KAAK,CAAC,CAAA;AAAA,EAC9B;AACA,EAAA,OAAO,GAAA;AACT;AAGA,SAAS,SAAA,CAAU,KAAA,EAAmB,KAAA,EAAe,GAAA,EAAqB;AACxE,EAAA,OAAO,QAAQ,MAAA,CAAO,KAAA,CAAM,QAAA,CAAS,KAAA,EAAO,GAAG,CAAC,CAAA;AAClD;AAGA,SAAS,qBAAqB,KAAA,EAA+B;AAC3D,EAAA,MAAM,OAAA,GAAU,EAAA;AAChB,EAAA,MAAM,QAAoB,EAAC;AAC3B,EAAA,IAAI,KAAA,GAAQ,CAAA;AACZ,EAAA,KAAA,IAAS,IAAI,CAAA,EAAG,CAAA,GAAI,KAAA,CAAM,UAAA,EAAY,KAAK,CAAA,EAAG;AAC5C,IAAA,IAAI,KAAA,CAAM,CAAC,CAAA,KAAM,OAAA,EAAS;AACxB,MAAA,MAAM,UAAA,GAAa,IAAI,KAAA,GAAQ,CAAA;AAC/B,MAAA,KAAA,CAAM,IAAA,CAAK,EAAE,IAAA,EAAM,SAAA,CAAU,KAAA,EAAO,KAAA,EAAO,CAAA,GAAI,CAAC,CAAA,EAAG,MAAA,EAAQ,KAAA,EAAO,UAAA,EAAY,CAAA;AAC9E,MAAA,KAAA,GAAQ,CAAA,GAAI,CAAA;AAAA,IACd;AAAA,EACF;AACA,EAAA,IAAI,KAAA,GAAQ,MAAM,UAAA,EAAY;AAC5B,IAAA,KAAA,CAAM,IAAA,CAAK,EAAE,IAAA,EAAM,SAAA,CAAU,OAAO,KAAA,EAAO,KAAA,CAAM,UAAU,CAAA,EAAG,QAAQ,KAAA,EAAO,UAAA,EAAY,KAAA,CAAM,UAAA,GAAa,OAAO,CAAA;AAAA,EACrH;AACA,EAAA,OAAO,KAAA;AACT;AAGA,SAAS,YAAA,CAAa,QAAA,EAAsB,MAAA,EAAoB,IAAA,EAAsB;AACpF,EAAA,MAAM,IAAA,GAAO,QAAA,CAAS,UAAA,GAAa,MAAA,CAAO,UAAA;AAC1C,EAAA,KAAA,IAAS,CAAA,GAAI,IAAA,EAAM,CAAA,IAAK,IAAA,EAAM,KAAK,CAAA,EAAG;AACpC,IAAA,IAAI,OAAA,GAAU,IAAA;AACd,IAAA,KAAA,IAAS,IAAI,CAAA,EAAG,CAAA,GAAI,MAAA,CAAO,UAAA,EAAY,KAAK,CAAA,EAAG;AAC7C,MAAA,IAAI,SAAS,CAAA,GAAI,CAAC,CAAA,KAAM,MAAA,CAAO,CAAC,CAAA,EAAG;AACjC,QAAA,OAAA,GAAU,KAAA;AACV,QAAA;AAAA,MACF;AAAA,IACF;AACA,IAAA,IAAI,SAAS,OAAO,CAAA;AAAA,EACtB;AACA,EAAA,OAAO,EAAA;AACT;AAEA,SAAS,QAAQ,CAAA,EAAmB;AAClC,EAAA,OACE,CAAA,CACG,WAAA,EAAY,CACZ,OAAA,CAAQ,eAAe,GAAG,CAAA,CAC1B,OAAA,CAAQ,UAAA,EAAY,EAAE,CAAA,CACtB,KAAA,CAAM,CAAA,EAAG,EAAE,CAAA,IAAK,SAAA;AAEvB;AAEA,SAAS,QAAA,CAAS,MAAc,KAAA,EAA4B;AAC1D,EAAA,IAAI,CAAC,KAAA,CAAM,GAAA,CAAI,IAAI,GAAG,OAAO,IAAA;AAC7B,EAAA,IAAI,CAAA,GAAI,CAAA;AACR,EAAA,OAAO,KAAA,CAAM,IAAI,CAAA,EAAG,IAAI,IAAI,CAAC,CAAA,CAAE,GAAG,CAAA,IAAK,CAAA;AACvC,EAAA,OAAO,CAAA,EAAG,IAAI,CAAA,CAAA,EAAI,CAAC,CAAA,CAAA;AACrB;AC/JO,SAAS,0BAA0B,IAAA,EAAoD;AAC5F,EAAA,IAAI,IAAA,CAAK,sBAAsB,KAAA,EAAO;AACpC,IAAAA,gBAAA,CAAI,iBAAA,GAAoB,KAAA;AAAA,EAC1B;AAEA,EAAA,IAAI,eAAA,GAA2C,IAAA;AAC/C,EAAA,IAAI,iBAAA,GAAmC,IAAA;AAEvC,EAAA,SAAS,WAAA,GAAgC;AACvC,IAAA,IAAI,CAAC,eAAA,EAAiB;AACpB,MAAA,MAAM,SAAA,GAAqC,EAAE,KAAA,EAAO,IAAA,CAAK,SAAS,MAAA,EAAO;AACzE,MAAA,IAAI,IAAA,CAAK,MAAA,KAAW,MAAA,EAAW,SAAA,CAAU,SAAS,IAAA,CAAK,MAAA;AACvD,MAAA,eAAA,GAAkBC,qBAAA,CAAS,oBAAA,EAAsB,IAAA,CAAK,KAAA,EAAO,SAAkB,CAAA;AAAA,IACjF;AACA,IAAA,OAAO,eAAA;AAAA,EACT;AAEA,EAAA,MAAM,OAAA,GAA4B;AAAA,IAChC,OAAO,IAAA,CAAK,KAAA;AAAA,IACZ,aAAA,EAAe,KAAK,aAAA,IAAiB,MAAA;AAAA,IACrC,MAAA,EAAQ,KAAK,MAAA,IAAU,QAAA;AAAA,IACvB,UAAA,EAAY,IAAA;AAAA,IACZ,MAAM,MAAM,KAAA,EAA2C;AACrD,MAAA,IAAI,KAAA,CAAM,WAAW,CAAA,EAAG;AACtB,QAAA,OAAO,EAAE,OAAA,EAAS,IAAI,SAAA,EAAW,IAAA,CAAK,aAAa,CAAA,EAAE;AAAA,MACvD;AACA,MAAA,MAAM,IAAA,GAAQ,MAAM,WAAA,EAAY;AAIhC,MAAA,MAAM,MAAA,GAAS,MAAM,IAAA,CAAK,KAAA,EAAO,EAAE,OAAA,EAAS,MAAA,EAAQ,SAAA,EAAW,IAAA,EAAM,CAAA;AACrE,MAAA,MAAM,OAAO,MAAA,CAAO,IAAA;AACpB,MAAA,MAAM,YAAY,MAAA,CAAO,IAAA,CAAK,MAAA,CAAO,IAAA,CAAK,SAAS,CAAC,CAAA;AACpD,MAAA,IAAI,iBAAA,KAAsB,IAAA,IAAQ,iBAAA,KAAsB,SAAA,EAAW;AACjE,QAAA,MAAM,IAAI,KAAA,CAAM,CAAA,0CAAA,EAA6C,iBAAiB,CAAA,MAAA,EAAS,SAAS,CAAA,CAAE,CAAA;AAAA,MACpG;AACA,MAAA,iBAAA,GAAoB,SAAA;AACpB,MAAA,MAAM,UAA0B,EAAC;AACjC,MAAA,KAAA,IAAS,IAAI,CAAA,EAAG,CAAA,GAAI,KAAA,CAAM,MAAA,EAAQ,KAAK,CAAA,EAAG;AACxC,QAAA,OAAA,CAAQ,IAAA,CAAK,IAAI,YAAA,CAAa,IAAA,CAAK,MAAA,EAAQ,IAAA,CAAK,UAAA,GAAa,CAAA,GAAI,SAAA,GAAY,CAAA,EAAG,SAAS,CAAA,CAAE,OAAO,CAAA;AAAA,MACpG;AACA,MAAA,OAAO,EAAE,SAAS,SAAA,EAAU;AAAA,IAC9B;AAAA,GACF;AACA,EAAA,IAAI,IAAA,CAAK,SAAA,KAAc,MAAA,EAAW,OAAA,CAAQ,YAAY,IAAA,CAAK,SAAA;AAC3D,EAAA,OAAO,OAAA;AACT;;;ACvCO,IAAM,aAAA,GAAgB;AACtB,IAAM,uBAAA,GAA0B;;;ACNvC,eAAsB,KAAA,CAAM,QAAA,EAAkB,IAAA,GAAqB,EAAC,EAA+B;AACjG,EAAA,MAAM,MAAA,GAAS,cAAc,QAAA,EAAU,EAAE,MAAM,IAAA,CAAK,QAAA,IAAY,WAAW,CAAA;AAC3E,EAAA,MAAM,OAAA,GACJ,KAAK,OAAA,IACL,yBAAA;AAAA,IACE,KAAK,aAAA,KAAkB,MAAA,GACnB,EAAE,KAAA,EAAO,KAAK,KAAA,IAAS,aAAA,EAAe,aAAA,EAAe,IAAA,CAAK,eAAc,GACxE,EAAE,KAAA,EAAO,IAAA,CAAK,SAAS,aAAA;AAAc,GAC3C;AAEF,EAAA,MAAM,MAAA,GAAS,MAAM,OAAA,CAAQ,KAAA,CAAM,MAAA,CAAO,IAAI,CAAC,CAAA,KAAM,CAAA,CAAE,IAAI,CAAC,CAAA;AAC5D,EAAA,IAAI,MAAA,CAAO,OAAA,CAAQ,MAAA,KAAW,MAAA,CAAO,MAAA,EAAQ;AAC3C,IAAA,MAAM,IAAI,MAAM,CAAA,iBAAA,EAAoB,MAAA,CAAO,QAAQ,MAAM,CAAA,aAAA,EAAgB,MAAA,CAAO,MAAM,CAAA,OAAA,CAAS,CAAA;AAAA,EACjG;AAEA,EAAA,MAAM,eAAA,GAAoC,MAAA,CAAO,GAAA,CAAI,CAAC,GAAG,CAAA,MAAO;AAAA,IAC9D,IAAI,CAAA,CAAE,EAAA;AAAA,IACN,YAAY,CAAA,CAAE,UAAA;AAAA,IACd,YAAY,CAAA,CAAE,UAAA;AAAA,IACd,MAAA,EAAQ,MAAA,CAAO,OAAA,CAAQ,CAAC;AAAA,GAC1B,CAAE,CAAA;AAEF,EAAA,MAAM,KAAA,GAAwB;AAAA,IAC5B,OAAO,OAAA,CAAQ,KAAA;AAAA,IACf,eAAe,OAAA,CAAQ,aAAA;AAAA,IACvB,WAAW,MAAA,CAAO,SAAA;AAAA,IAClB,QAAQ,OAAA,CAAQ,MAAA;AAAA,IAChB,YAAY,OAAA,CAAQ,UAAA;AAAA,IACpB,QAAA,EAAU,KAAK,QAAA,IAAY,SAAA;AAAA,IAC3B,MAAA,EAAQ;AAAA,GACV;AAEA,EAAA,OAAO,EAAE,aAAA,EAAe,CAAA,EAAG,MAAA,EAAQ,CAAC,KAAK,CAAA,EAAE;AAC7C;;;AC/BO,SAAS,cAAA,CACd,OAAA,EACA,WAAA,EACA,IAAA,GAAsB,EAAC,EACV;AACb,EAAA,MAAM,KAAA,GAAQ,SAAA,CAAU,OAAA,EAAS,IAAA,CAAK,KAAK,CAAA;AAC3C,EAAA,IAAI,CAAC,KAAA,EAAO,MAAM,IAAI,MAAM,mCAAmC,CAAA;AAC/D,EAAA,IAAI,WAAA,CAAY,MAAA,KAAW,KAAA,CAAM,SAAA,EAAW;AAC1C,IAAA,MAAM,IAAI,KAAA,CAAM,CAAA,uBAAA,EAA0B,WAAA,CAAY,MAAM,CAAA,sBAAA,EAAyB,KAAA,CAAM,KAAK,CAAA,EAAA,EAAK,KAAA,CAAM,SAAS,CAAA,CAAA,CAAG,CAAA;AAAA,EACzH;AAEA,EAAA,MAAM,CAAA,GAAI,KAAK,CAAA,IAAK,CAAA;AACpB,EAAA,MAAM,MAAA,GAAsB,KAAA,CAAM,MAAA,CAAO,GAAA,CAAI,CAAC,CAAA,MAAO;AAAA,IACnD,YAAY,KAAA,CAAM,KAAA;AAAA,IAClB,SAAS,CAAA,CAAE,EAAA;AAAA,IACX,YAAY,CAAA,CAAE,UAAA;AAAA,IACd,YAAY,CAAA,CAAE,UAAA;AAAA,IACd,OAAO,UAAA,CAAW,WAAA,EAAa,CAAA,CAAE,MAAA,EAAQ,MAAM,MAAM;AAAA,GACvD,CAAE,CAAA;AAGF,EAAA,MAAM,KAAA,GAAQ,KAAA,CAAM,MAAA,KAAW,WAAA,GAAc,CAAA,GAAI,EAAA;AACjD,EAAA,MAAA,CAAO,IAAA,CAAK,CAAC,CAAA,EAAG,CAAA,KAAM,SAAS,CAAA,CAAE,KAAA,GAAQ,EAAE,KAAA,CAAM,CAAA;AACjD,EAAA,OAAO,MAAA,CAAO,KAAA,CAAM,CAAA,EAAG,CAAC,CAAA;AAC1B;AAEA,SAAS,SAAA,CAAU,SAA4B,SAAA,EAAgD;AAC7F,EAAA,IAAI,SAAA,SAAkB,OAAA,CAAQ,MAAA,CAAO,KAAK,CAAC,CAAA,KAAM,CAAA,CAAE,KAAA,KAAU,SAAS,CAAA;AACtE,EAAA,OAAO,OAAA,CAAQ,OAAO,CAAC,CAAA;AACzB;AAEA,SAAS,UAAA,CAAW,CAAA,EAAiB,CAAA,EAAiB,MAAA,EAAiC;AACrF,EAAA,IAAI,WAAW,WAAA,EAAa;AAC1B,IAAA,IAAI,GAAA,GAAM,CAAA;AACV,IAAA,KAAA,IAAS,IAAI,CAAA,EAAG,CAAA,GAAI,CAAA,CAAE,MAAA,EAAQ,KAAK,CAAA,EAAG;AACpC,MAAA,MAAM,CAAA,GAAI,CAAA,CAAE,CAAC,CAAA,GAAK,EAAE,CAAC,CAAA;AACrB,MAAA,GAAA,IAAO,CAAA,GAAI,CAAA;AAAA,IACb;AACA,IAAA,OAAO,IAAA,CAAK,KAAK,GAAG,CAAA;AAAA,EACtB;AACA,EAAA,IAAI,GAAA,GAAM,CAAA;AACV,EAAA,KAAA,IAAS,CAAA,GAAI,CAAA,EAAG,CAAA,GAAI,CAAA,CAAE,MAAA,EAAQ,CAAA,IAAK,CAAA,EAAG,GAAA,IAAO,CAAA,CAAE,CAAC,CAAA,GAAK,CAAA,CAAE,CAAC,CAAA;AACxD,EAAA,IAAI,MAAA,KAAW,OAAO,OAAO,GAAA;AAE7B,EAAA,IAAI,EAAA,GAAK,CAAA;AACT,EAAA,IAAI,EAAA,GAAK,CAAA;AACT,EAAA,KAAA,IAAS,IAAI,CAAA,EAAG,CAAA,GAAI,CAAA,CAAE,MAAA,EAAQ,KAAK,CAAA,EAAG;AACpC,IAAA,EAAA,IAAM,CAAA,CAAE,CAAC,CAAA,GAAK,CAAA,CAAE,CAAC,CAAA;AACjB,IAAA,EAAA,IAAM,CAAA,CAAE,CAAC,CAAA,GAAK,CAAA,CAAE,CAAC,CAAA;AAAA,EACnB;AACA,EAAA,MAAM,QAAQ,IAAA,CAAK,IAAA,CAAK,EAAE,CAAA,GAAI,IAAA,CAAK,KAAK,EAAE,CAAA;AAC1C,EAAA,OAAO,KAAA,KAAU,CAAA,GAAI,CAAA,GAAI,GAAA,GAAM,KAAA;AACjC;;;AClDA,IAAM,gBAAA,GAAmB,mDAAA;AAElB,SAAS,yBAAyB,IAAA,EAAmD;AAC1F,EAAA,MAAM,QAAQ,IAAA,CAAK,KAAA,IAAS,QAAQ,GAAA,CAAI,QAAA,IAAY,QAAQ,GAAA,CAAI,iBAAA;AAChE,EAAA,IAAI,CAAC,KAAA,EAAO;AACV,IAAA,MAAM,IAAI,MAAM,0EAA0E,CAAA;AAAA,EAC5F;AACA,EAAA,MAAM,OAAA,GAAU,KAAK,OAAA,IAAW,gBAAA;AAChC,EAAA,MAAM,MAAM,CAAA,EAAG,OAAO,IAAI,SAAA,CAAU,IAAA,CAAK,KAAK,CAAC,CAAA,4BAAA,CAAA;AAE/C,EAAA,MAAM,OAAA,GAA4B;AAAA,IAChC,OAAO,IAAA,CAAK,KAAA;AAAA,IACZ,aAAA,EAAe,KAAK,aAAA,IAAiB,MAAA;AAAA,IACrC,MAAA,EAAQ,KAAK,MAAA,IAAU,QAAA;AAAA,IACvB,UAAA,EAAY,IAAA;AAAA,IACZ,MAAM,MAAM,KAAA,EAA2C;AACrD,MAAA,IAAI,KAAA,CAAM,WAAW,CAAA,EAAG;AACtB,QAAA,OAAO,EAAE,OAAA,EAAS,IAAI,SAAA,EAAW,IAAA,CAAK,aAAa,CAAA,EAAE;AAAA,MACvD;AACA,MAAA,MAAM,IAAA,GAAO,IAAA,CAAK,SAAA,CAAU,EAAE,MAAA,EAAQ,KAAA,EAAO,OAAA,EAAS,EAAE,cAAA,EAAgB,IAAA,EAAK,EAAG,CAAA;AAChF,MAAA,MAAM,GAAA,GAAM,MAAM,KAAA,CAAM,GAAA,EAAK;AAAA,QAC3B,MAAA,EAAQ,MAAA;AAAA,QACR,OAAA,EAAS;AAAA,UACP,cAAA,EAAgB,kBAAA;AAAA,UAChB,aAAA,EAAe,UAAU,KAAK,CAAA;AAAA,SAChC;AAAA,QACA;AAAA,OACD,CAAA;AACD,MAAA,IAAI,CAAC,IAAI,EAAA,EAAI;AACX,QAAA,MAAM,SAAS,MAAM,GAAA,CAAI,MAAK,CAAE,KAAA,CAAM,MAAM,EAAE,CAAA;AAC9C,QAAA,MAAM,IAAI,KAAA,CAAM,CAAA,iBAAA,EAAoB,GAAA,CAAI,MAAM,CAAA,KAAA,EAAQ,IAAA,CAAK,KAAK,CAAA,EAAA,EAAK,MAAA,IAAU,GAAA,CAAI,UAAU,CAAA,CAAE,CAAA;AAAA,MACjG;AACA,MAAA,MAAM,GAAA,GAAO,MAAM,GAAA,CAAI,IAAA,EAAK;AAC5B,MAAA,MAAM,MAAA,GAAS,eAAA,CAAgB,GAAA,EAAK,KAAA,CAAM,MAAM,CAAA;AAChD,MAAA,MAAM,YAAY,MAAA,CAAO,CAAC,CAAA,EAAG,MAAA,IAAU,KAAK,SAAA,IAAa,CAAA;AACzD,MAAA,MAAM,OAAA,GAAU,MAAA,CAAO,GAAA,CAAI,CAAC,CAAA,KAAM,UAAU,YAAA,CAAa,IAAA,CAAK,CAAC,CAAC,CAAC,CAAA;AACjE,MAAA,OAAO,EAAE,SAAS,SAAA,EAAU;AAAA,IAC9B;AAAA,GACF;AACA,EAAA,IAAI,IAAA,CAAK,SAAA,KAAc,MAAA,EAAW,OAAA,CAAQ,YAAY,IAAA,CAAK,SAAA;AAC3D,EAAA,OAAO,OAAA;AACT;AAWA,SAAS,eAAA,CAAgB,KAAc,aAAA,EAAmC;AACxE,EAAA,IAAI,CAAC,KAAA,CAAM,OAAA,CAAQ,GAAG,CAAA,EAAG;AACvB,IAAA,MAAM,IAAI,MAAM,2CAA2C,CAAA;AAAA,EAC7D;AACA,EAAA,IAAI,GAAA,CAAI,MAAA,KAAW,CAAA,EAAG,OAAO,EAAC;AAE9B,EAAA,MAAM,KAAA,GAAQ,IAAI,CAAC,CAAA;AACnB,EAAA,IAAI,OAAO,UAAU,QAAA,EAAU;AAC7B,IAAA,IAAI,kBAAkB,CAAA,EAAG;AACvB,MAAA,MAAM,IAAI,KAAA,CAAM,CAAA,yCAAA,EAA4C,aAAa,CAAA,CAAE,CAAA;AAAA,IAC7E;AACA,IAAA,OAAO,CAAC,GAAe,CAAA;AAAA,EACzB;AACA,EAAA,IAAI,KAAA,CAAM,OAAA,CAAQ,KAAK,CAAA,KAAM,KAAA,CAAM,MAAA,KAAW,CAAA,IAAK,OAAO,KAAA,CAAM,CAAC,CAAA,KAAM,QAAA,CAAA,EAAW;AAChF,IAAA,OAAO,GAAA;AAAA,EACT;AACA,EAAA,IAAI,KAAA,CAAM,QAAQ,KAAK,CAAA,IAAK,MAAM,OAAA,CAAQ,KAAA,CAAM,CAAC,CAAC,CAAA,EAAG;AAEnD,IAAA,OAAQ,GAAA,CAAqB,IAAI,QAAQ,CAAA;AAAA,EAC3C;AACA,EAAA,MAAM,IAAI,MAAM,+CAA+C,CAAA;AACjE;AAEA,SAAS,SAAS,MAAA,EAA8B;AAC9C,EAAA,IAAI,MAAA,CAAO,MAAA,KAAW,CAAA,EAAG,OAAO,EAAC;AACjC,EAAA,MAAM,GAAA,GAAM,MAAA,CAAO,CAAC,CAAA,CAAG,MAAA;AACvB,EAAA,MAAM,MAAM,IAAI,KAAA,CAAc,GAAG,CAAA,CAAE,KAAK,CAAC,CAAA;AACzC,EAAA,KAAA,MAAW,KAAK,MAAA,EAAQ;AACtB,IAAA,KAAA,IAAS,CAAA,GAAI,CAAA,EAAG,CAAA,GAAI,GAAA,EAAK,CAAA,IAAK,GAAG,GAAA,CAAI,CAAC,CAAA,IAAM,CAAA,CAAE,CAAC,CAAA;AAAA,EACjD;AACA,EAAA,KAAA,IAAS,CAAA,GAAI,GAAG,CAAA,GAAI,GAAA,EAAK,KAAK,CAAA,EAAG,GAAA,CAAI,CAAC,CAAA,IAAM,MAAA,CAAO,MAAA;AACnD,EAAA,OAAO,GAAA;AACT;AAEA,SAAS,UAAU,CAAA,EAA+B;AAChD,EAAA,IAAI,GAAA,GAAM,CAAA;AACV,EAAA,KAAA,IAAS,CAAA,GAAI,CAAA,EAAG,CAAA,GAAI,CAAA,CAAE,MAAA,EAAQ,CAAA,IAAK,CAAA,EAAG,GAAA,IAAO,CAAA,CAAE,CAAC,CAAA,GAAK,CAAA,CAAE,CAAC,CAAA;AACxD,EAAA,MAAM,IAAA,GAAO,IAAA,CAAK,IAAA,CAAK,GAAG,CAAA;AAC1B,EAAA,IAAI,IAAA,KAAS,GAAG,OAAO,CAAA;AACvB,EAAA,KAAA,IAAS,CAAA,GAAI,CAAA,EAAG,CAAA,GAAI,CAAA,CAAE,MAAA,EAAQ,CAAA,IAAK,CAAA,EAAG,CAAA,CAAE,CAAC,CAAA,GAAI,CAAA,CAAE,CAAC,CAAA,GAAK,IAAA;AACrD,EAAA,OAAO,CAAA;AACT","file":"index.cjs","sourcesContent":["/**\n * Section-based markdown chunker.\n *\n * Splits a markdown document on ATX headings (`#`, `##`, ...). Each chunk\n * carries the byte offset and length into the original UTF-8 source so a\n * downstream consumer can map a vector hit back to the exact substring\n * without re-tokenising. Pre-heading content becomes a \"preamble\" chunk.\n *\n * Per spec §5.1, `textOffset`/`textLength` are UTF-8 *byte* offsets into the\n * markdown source. We encode the document once with `TextEncoder`, track a\n * byte cursor while iterating lines (counting the trailing `\\n` byte), and\n * derive each chunk's `text` by decoding the corresponding byte slice. This\n * keeps the offsets in agreement with the Go and Python SDKs for any\n * non-ASCII résumé.\n */\n\nconst encoder = new TextEncoder();\nconst decoder = new TextDecoder();\n\nexport type ChunkingMode = 'document' | 'section' | 'paragraph';\n\nexport interface MarkdownChunk {\n  id: string;\n  textOffset: number;\n  textLength: number;\n  text: string;\n}\n\nexport interface ChunkOptions {\n  mode?: ChunkingMode;\n}\n\nconst HEADING = /^(#{1,6})\\s+(.+?)\\s*$/;\n\n/** A source line plus its UTF-8 byte offset and byte length (including any trailing `\\n`). */\ninterface ByteLine {\n  text: string;\n  offset: number;\n  byteLength: number;\n}\n\nexport function chunkMarkdown(markdown: string, opts: ChunkOptions = {}): MarkdownChunk[] {\n  const mode = opts.mode ?? 'section';\n  const bytes = encoder.encode(markdown);\n  if (mode === 'document') {\n    return [documentChunk(bytes)];\n  }\n  if (mode === 'paragraph') {\n    return paragraphChunks(bytes);\n  }\n  return sectionChunks(bytes);\n}\n\nfunction documentChunk(bytes: Uint8Array): MarkdownChunk {\n  return { id: 'document', textOffset: 0, textLength: bytes.byteLength, text: sliceText(bytes, 0, bytes.byteLength) };\n}\n\nfunction sectionChunks(bytes: Uint8Array): MarkdownChunk[] {\n  const lines = splitWithByteOffsets(bytes);\n  const sections: MarkdownChunk[] = [];\n  let current: { id: string; start: number; end: number } | null = null;\n  const ids = new Set<string>();\n\n  function flush(end: number): void {\n    if (!current) return;\n    const text = sliceText(bytes, current.start, end);\n    if (text.trim().length === 0) {\n      current = null;\n      return;\n    }\n    sections.push({ id: current.id, textOffset: current.start, textLength: end - current.start, text });\n    current = null;\n  }\n\n  for (const line of lines) {\n    const match = HEADING.exec(line.text);\n    const lineEnd = line.offset + line.byteLength;\n    if (match) {\n      flush(line.offset);\n      const id = uniqueId(slugify(match[2] ?? `section-${sections.length + 1}`), ids);\n      ids.add(id);\n      current = { id, start: line.offset, end: lineEnd };\n      continue;\n    }\n    if (current === null) {\n      const id = uniqueId('preamble', ids);\n      ids.add(id);\n      current = { id, start: line.offset, end: lineEnd };\n    } else {\n      current.end = lineEnd;\n    }\n  }\n  flush(bytes.byteLength);\n\n  if (sections.length === 0) {\n    return [documentChunk(bytes)];\n  }\n  return sections;\n}\n\nfunction paragraphChunks(bytes: Uint8Array): MarkdownChunk[] {\n  const out: MarkdownChunk[] = [];\n  const ids = new Set<string>();\n  const separator = encoder.encode('\\n\\n');\n  let cursor = 0;\n  let i = 0;\n  while (cursor < bytes.byteLength) {\n    let end = indexOfBytes(bytes, separator, cursor);\n    if (end === -1) end = bytes.byteLength;\n    const text = sliceText(bytes, cursor, end);\n    if (text.trim().length > 0) {\n      const id = uniqueId(slugify(text.split('\\n')[0] ?? `p-${i}`), ids);\n      ids.add(id);\n      out.push({ id, textOffset: cursor, textLength: end - cursor, text });\n      i += 1;\n    }\n    cursor = end + separator.byteLength;\n  }\n  if (out.length === 0) {\n    return [documentChunk(bytes)];\n  }\n  return out;\n}\n\n/** Decode the UTF-8 byte slice `[start, end)` back into a string. */\nfunction sliceText(bytes: Uint8Array, start: number, end: number): string {\n  return decoder.decode(bytes.subarray(start, end));\n}\n\n/** Split UTF-8 bytes into lines, each tagged with its byte offset and byte length (newline included). */\nfunction splitWithByteOffsets(bytes: Uint8Array): ByteLine[] {\n  const newline = 0x0a; // '\\n'\n  const lines: ByteLine[] = [];\n  let start = 0;\n  for (let i = 0; i < bytes.byteLength; i += 1) {\n    if (bytes[i] === newline) {\n      const byteLength = i - start + 1;\n      lines.push({ text: sliceText(bytes, start, i + 1), offset: start, byteLength });\n      start = i + 1;\n    }\n  }\n  if (start < bytes.byteLength) {\n    lines.push({ text: sliceText(bytes, start, bytes.byteLength), offset: start, byteLength: bytes.byteLength - start });\n  }\n  return lines;\n}\n\n/** Find the byte index of `needle` in `haystack` at or after `from`, or -1. */\nfunction indexOfBytes(haystack: Uint8Array, needle: Uint8Array, from: number): number {\n  const last = haystack.byteLength - needle.byteLength;\n  for (let i = from; i <= last; i += 1) {\n    let matched = true;\n    for (let j = 0; j < needle.byteLength; j += 1) {\n      if (haystack[i + j] !== needle[j]) {\n        matched = false;\n        break;\n      }\n    }\n    if (matched) return i;\n  }\n  return -1;\n}\n\nfunction slugify(s: string): string {\n  return (\n    s\n      .toLowerCase()\n      .replace(/[^a-z0-9]+/g, '-')\n      .replace(/^-+|-+$/g, '')\n      .slice(0, 64) || 'section'\n  );\n}\n\nfunction uniqueId(base: string, taken: Set<string>): string {\n  if (!taken.has(base)) return base;\n  let n = 2;\n  while (taken.has(`${base}-${n}`)) n += 1;\n  return `${base}-${n}`;\n}\n","/**\n * transformers.js backend (browser + Node + Bun via @huggingface/transformers).\n *\n * Loads the model lazily and reuses the pipeline across calls. Models are\n * cached on disk by transformers.js itself (HF cache layout).\n */\n\nimport { pipeline, env } from '@huggingface/transformers';\nimport type { EmbeddingBackend, EmbeddingBackendOptions, EmbeddingMatrix } from './types.js';\n\ninterface TransformersBackendOptions extends EmbeddingBackendOptions {\n  /** Override Xenova quantisation. Defaults to fp32 for closest fidelity. */\n  dtype?: 'fp32' | 'fp16' | 'q8' | 'q4';\n  /** Force a backend device (e.g. 'cpu', 'gpu', 'wasm'). Defaults to auto. */\n  device?: 'cpu' | 'gpu' | 'wasm' | 'webgpu';\n  /** Allow remote model downloads. true by default. */\n  allowRemoteModels?: boolean;\n}\n\nexport function createTransformersBackend(opts: TransformersBackendOptions): EmbeddingBackend {\n  if (opts.allowRemoteModels === false) {\n    env.allowRemoteModels = false;\n  }\n\n  let pipelinePromise: Promise<unknown> | null = null;\n  let resolvedDimension: number | null = null;\n\n  function getPipeline(): Promise<unknown> {\n    if (!pipelinePromise) {\n      const modelOpts: Record<string, unknown> = { dtype: opts.dtype ?? 'fp32' };\n      if (opts.device !== undefined) modelOpts.device = opts.device;\n      pipelinePromise = pipeline('feature-extraction', opts.model, modelOpts as never);\n    }\n    return pipelinePromise;\n  }\n\n  const backend: EmbeddingBackend = {\n    model: opts.model,\n    modelRevision: opts.modelRevision ?? 'main',\n    metric: opts.metric ?? 'cosine',\n    normalized: true,\n    async embed(texts: string[]): Promise<EmbeddingMatrix> {\n      if (texts.length === 0) {\n        return { vectors: [], dimension: opts.dimension ?? 0 };\n      }\n      const pipe = (await getPipeline()) as (\n        texts: string[],\n        opts: { pooling: 'mean'; normalize: boolean },\n      ) => Promise<{ data: Float32Array; dims: number[] }>;\n      const tensor = await pipe(texts, { pooling: 'mean', normalize: true });\n      const data = tensor.data;\n      const dimension = tensor.dims[tensor.dims.length - 1] as number;\n      if (resolvedDimension !== null && resolvedDimension !== dimension) {\n        throw new Error(`Model emitted inconsistent dimension: had ${resolvedDimension}, now ${dimension}`);\n      }\n      resolvedDimension = dimension;\n      const vectors: Float32Array[] = [];\n      for (let i = 0; i < texts.length; i += 1) {\n        vectors.push(new Float32Array(data.buffer, data.byteOffset + i * dimension * 4, dimension).slice());\n      }\n      return { vectors, dimension };\n    },\n  };\n  if (opts.dimension !== undefined) backend.dimension = opts.dimension;\n  return backend;\n}\n","import type { EmbeddingMetric } from '@cvfile/sdk';\n\nexport interface EmbeddingMatrix {\n  vectors: Float32Array[];\n  dimension: number;\n}\n\nexport interface EmbeddingBackend {\n  model: string;\n  modelRevision: string;\n  /** Optional pre-declared dimension; the backend may override after first call. */\n  dimension?: number;\n  metric: EmbeddingMetric;\n  normalized: boolean;\n  embed(texts: string[]): Promise<EmbeddingMatrix>;\n}\n\nexport interface EmbeddingBackendOptions {\n  model: string;\n  modelRevision?: string;\n  /** Optional pre-known dimension (e.g. 1024 for bge-m3). */\n  dimension?: number;\n  metric?: EmbeddingMetric;\n}\n\n/** Recommended default per spec §5: BAAI BGE-M3, MIT, multilingual, 1024-dim. */\nexport const DEFAULT_MODEL = 'Xenova/bge-m3';\nexport const DEFAULT_MODEL_DIMENSION = 1024;\n","/**\n * High-level embed() API: markdown in, EmbeddingsPayload out, ready to drop\n * into pack({ embeddings: ... }).\n */\n\nimport type { EmbeddingChunk, EmbeddingSpace, EmbeddingsPayload } from '@cvfile/sdk';\nimport { chunkMarkdown, type ChunkingMode } from './chunk.js';\nimport { createTransformersBackend } from './transformers-backend.js';\nimport { DEFAULT_MODEL, type EmbeddingBackend } from './types.js';\n\nexport interface EmbedOptions {\n  /** HF model id; defaults to BGE-M3 (Xenova/bge-m3). */\n  model?: string;\n  /** Pinned model revision; recorded in the payload for reproducibility. */\n  modelRevision?: string;\n  /** Chunking strategy; default 'section'. */\n  chunking?: ChunkingMode;\n  /** Bring-your-own backend (e.g. an OpenAI/Voyage adapter). */\n  backend?: EmbeddingBackend;\n}\n\nexport async function embed(markdown: string, opts: EmbedOptions = {}): Promise<EmbeddingsPayload> {\n  const chunks = chunkMarkdown(markdown, { mode: opts.chunking ?? 'section' });\n  const backend =\n    opts.backend ??\n    createTransformersBackend(\n      opts.modelRevision !== undefined\n        ? { model: opts.model ?? DEFAULT_MODEL, modelRevision: opts.modelRevision }\n        : { model: opts.model ?? DEFAULT_MODEL },\n    );\n\n  const matrix = await backend.embed(chunks.map((c) => c.text));\n  if (matrix.vectors.length !== chunks.length) {\n    throw new Error(`Backend returned ${matrix.vectors.length} vectors for ${chunks.length} chunks`);\n  }\n\n  const embeddingChunks: EmbeddingChunk[] = chunks.map((c, i) => ({\n    id: c.id,\n    textOffset: c.textOffset,\n    textLength: c.textLength,\n    vector: matrix.vectors[i]!,\n  }));\n\n  const space: EmbeddingSpace = {\n    model: backend.model,\n    modelRevision: backend.modelRevision,\n    dimension: matrix.dimension,\n    metric: backend.metric,\n    normalized: backend.normalized,\n    chunking: opts.chunking ?? 'section',\n    chunks: embeddingChunks,\n  };\n\n  return { formatVersion: 1, spaces: [space] };\n}\n","/**\n * Pure similarity search over an EmbeddingsPayload. Vector input is the\n * caller's responsibility (encode the query with the same model that was\n * used to populate the space, ideally pulled from space.model).\n */\n\nimport type { EmbeddingMetric, EmbeddingSpace, EmbeddingsPayload } from '@cvfile/sdk';\n\nexport interface SearchHit {\n  spaceModel: string;\n  chunkId: string;\n  textOffset: number;\n  textLength: number;\n  score: number;\n}\n\nexport interface SearchOptions {\n  /** Restrict to a specific embedding space; default = first space. */\n  model?: string;\n  /** Top-k results; default 5. */\n  k?: number;\n}\n\nexport function searchSemantic(\n  payload: EmbeddingsPayload,\n  queryVector: Float32Array,\n  opts: SearchOptions = {},\n): SearchHit[] {\n  const space = pickSpace(payload, opts.model);\n  if (!space) throw new Error('No matching embedding space found');\n  if (queryVector.length !== space.dimension) {\n    throw new Error(`Query vector dimension ${queryVector.length} does not match space ${space.model} (${space.dimension})`);\n  }\n\n  const k = opts.k ?? 5;\n  const scored: SearchHit[] = space.chunks.map((c) => ({\n    spaceModel: space.model,\n    chunkId: c.id,\n    textOffset: c.textOffset,\n    textLength: c.textLength,\n    score: similarity(queryVector, c.vector, space.metric),\n  }));\n\n  // Higher score = better for cosine/dot; lower for euclidean.\n  const order = space.metric === 'euclidean' ? 1 : -1;\n  scored.sort((a, b) => order * (a.score - b.score));\n  return scored.slice(0, k);\n}\n\nfunction pickSpace(payload: EmbeddingsPayload, modelHint?: string): EmbeddingSpace | undefined {\n  if (modelHint) return payload.spaces.find((s) => s.model === modelHint);\n  return payload.spaces[0];\n}\n\nfunction similarity(a: Float32Array, b: Float32Array, metric: EmbeddingMetric): number {\n  if (metric === 'euclidean') {\n    let sum = 0;\n    for (let i = 0; i < a.length; i += 1) {\n      const d = a[i]! - b[i]!;\n      sum += d * d;\n    }\n    return Math.sqrt(sum);\n  }\n  let dot = 0;\n  for (let i = 0; i < a.length; i += 1) dot += a[i]! * b[i]!;\n  if (metric === 'dot') return dot;\n  // cosine: assume normalized vectors when produced by our backend\n  let na = 0;\n  let nb = 0;\n  for (let i = 0; i < a.length; i += 1) {\n    na += a[i]! * a[i]!;\n    nb += b[i]! * b[i]!;\n  }\n  const denom = Math.sqrt(na) * Math.sqrt(nb);\n  return denom === 0 ? 0 : dot / denom;\n}\n","/**\n * Hugging Face Inference API backend.\n *\n * Calls https://router.huggingface.co/hf-inference/models/<model>/pipeline/feature-extraction\n * with HF_TOKEN. For sentence-transformers models (BGE-M3, MiniLM, etc.) the\n * response is already mean-pooled per input — one vector per text. We\n * normalise client-side so cosine math is consistent across backends.\n */\n\nimport type { EmbeddingBackend, EmbeddingMatrix } from './types.js';\nimport type { EmbeddingMetric } from '@cvfile/sdk';\n\nexport interface HuggingFaceBackendOptions {\n  model: string;\n  /** HF token. Defaults to `process.env.HF_TOKEN`. */\n  token?: string;\n  /** Pinned revision; recorded in the payload. Default 'main'. */\n  modelRevision?: string;\n  /** Pre-known dimension. Optional; inferred from first response otherwise. */\n  dimension?: number;\n  metric?: EmbeddingMetric;\n  /** Override base URL (e.g. for self-hosted TEI). */\n  baseUrl?: string;\n}\n\nconst DEFAULT_BASE_URL = 'https://router.huggingface.co/hf-inference/models';\n\nexport function createHuggingFaceBackend(opts: HuggingFaceBackendOptions): EmbeddingBackend {\n  const token = opts.token ?? process.env.HF_TOKEN ?? process.env.HUGGINGFACE_TOKEN;\n  if (!token) {\n    throw new Error('HF_TOKEN (or HUGGINGFACE_TOKEN) is required for the Hugging Face backend');\n  }\n  const baseUrl = opts.baseUrl ?? DEFAULT_BASE_URL;\n  const url = `${baseUrl}/${encodeURI(opts.model)}/pipeline/feature-extraction`;\n\n  const backend: EmbeddingBackend = {\n    model: opts.model,\n    modelRevision: opts.modelRevision ?? 'main',\n    metric: opts.metric ?? 'cosine',\n    normalized: true,\n    async embed(texts: string[]): Promise<EmbeddingMatrix> {\n      if (texts.length === 0) {\n        return { vectors: [], dimension: opts.dimension ?? 0 };\n      }\n      const body = JSON.stringify({ inputs: texts, options: { wait_for_model: true } });\n      const res = await fetch(url, {\n        method: 'POST',\n        headers: {\n          'content-type': 'application/json',\n          authorization: `Bearer ${token}`,\n        },\n        body,\n      });\n      if (!res.ok) {\n        const detail = await res.text().catch(() => '');\n        throw new Error(`HF Inference API ${res.status} for ${opts.model}: ${detail || res.statusText}`);\n      }\n      const raw = (await res.json()) as unknown;\n      const matrix = parseHfResponse(raw, texts.length);\n      const dimension = matrix[0]?.length ?? opts.dimension ?? 0;\n      const vectors = matrix.map((v) => normalize(Float32Array.from(v)));\n      return { vectors, dimension };\n    },\n  };\n  if (opts.dimension !== undefined) backend.dimension = opts.dimension;\n  return backend;\n}\n\n/**\n * Coerce the variety of shapes the HF Inference API returns into a flat\n * `number[][]`: one mean-pooled vector per input.\n *\n * Observed shapes:\n *   - sentence-transformers (BGE-M3, MiniLM): `[[...vec], [...vec]]`\n *   - feature-extraction without pooling: `[[[...token0], [...token1], ...], ...]`\n *   - single-input convenience form: `[...vec]`\n */\nfunction parseHfResponse(raw: unknown, expectedCount: number): number[][] {\n  if (!Array.isArray(raw)) {\n    throw new Error('HF Inference API: expected array response');\n  }\n  if (raw.length === 0) return [];\n\n  const first = raw[0];\n  if (typeof first === 'number') {\n    if (expectedCount !== 1) {\n      throw new Error(`HF Inference API: got 1 vector, expected ${expectedCount}`);\n    }\n    return [raw as number[]];\n  }\n  if (Array.isArray(first) && (first.length === 0 || typeof first[0] === 'number')) {\n    return raw as number[][];\n  }\n  if (Array.isArray(first) && Array.isArray(first[0])) {\n    // token-level embeddings: mean-pool per input\n    return (raw as number[][][]).map(meanPool);\n  }\n  throw new Error('HF Inference API: unrecognised response shape');\n}\n\nfunction meanPool(tokens: number[][]): number[] {\n  if (tokens.length === 0) return [];\n  const dim = tokens[0]!.length;\n  const out = new Array<number>(dim).fill(0);\n  for (const t of tokens) {\n    for (let i = 0; i < dim; i += 1) out[i]! += t[i]!;\n  }\n  for (let i = 0; i < dim; i += 1) out[i]! /= tokens.length;\n  return out;\n}\n\nfunction normalize(v: Float32Array): Float32Array {\n  let sum = 0;\n  for (let i = 0; i < v.length; i += 1) sum += v[i]! * v[i]!;\n  const norm = Math.sqrt(sum);\n  if (norm === 0) return v;\n  for (let i = 0; i < v.length; i += 1) v[i] = v[i]! / norm;\n  return v;\n}\n"]}

package/dist/index.d.cts

@@ -7,6 +7,13 @@ import { EmbeddingMetric, EmbeddingsPayload } from '@cvfile/sdk';
  * carries the byte offset and length into the original UTF-8 source so a
  * downstream consumer can map a vector hit back to the exact substring
  * without re-tokenising. Pre-heading content becomes a "preamble" chunk.
+ *
+ * Per spec §5.1, `textOffset`/`textLength` are UTF-8 *byte* offsets into the
+ * markdown source. We encode the document once with `TextEncoder`, track a
+ * byte cursor while iterating lines (counting the trailing `\n` byte), and
+ * derive each chunk's `text` by decoding the corresponding byte slice. This
+ * keeps the offsets in agreement with the Go and Python SDKs for any
+ * non-ASCII résumé.
  */
 type ChunkingMode = 'document' | 'section' | 'paragraph';
 interface MarkdownChunk {

package/dist/index.d.ts

@@ -7,6 +7,13 @@ import { EmbeddingMetric, EmbeddingsPayload } from '@cvfile/sdk';
  * carries the byte offset and length into the original UTF-8 source so a
  * downstream consumer can map a vector hit back to the exact substring
  * without re-tokenising. Pre-heading content becomes a "preamble" chunk.
+ *
+ * Per spec §5.1, `textOffset`/`textLength` are UTF-8 *byte* offsets into the
+ * markdown source. We encode the document once with `TextEncoder`, track a
+ * byte cursor while iterating lines (counting the trailing `\n` byte), and
+ * derive each chunk's `text` by decoding the corresponding byte slice. This
+ * keeps the offsets in agreement with the Go and Python SDKs for any
+ * non-ASCII résumé.
  */
 type ChunkingMode = 'document' | 'section' | 'paragraph';
 interface MarkdownChunk {

package/dist/index.js

@@ -1,25 +1,31 @@
 import { env, pipeline } from '@huggingface/transformers';
 
 // src/chunk.ts
+var encoder = new TextEncoder();
+var decoder = new TextDecoder();
 var HEADING = /^(#{1,6})\s+(.+?)\s*$/;
 function chunkMarkdown(markdown, opts = {}) {
   const mode = opts.mode ?? "section";
+  const bytes = encoder.encode(markdown);
   if (mode === "document") {
-    return [{ id: "document", textOffset: 0, textLength: markdown.length, text: markdown }];
+    return [documentChunk(bytes)];
   }
   if (mode === "paragraph") {
-    return paragraphChunks(markdown);
+    return paragraphChunks(bytes);
   }
-  return sectionChunks(markdown);
+  return sectionChunks(bytes);
 }
-function sectionChunks(markdown) {
-  const lines = splitWithOffsets(markdown);
+function documentChunk(bytes) {
+  return { id: "document", textOffset: 0, textLength: bytes.byteLength, text: sliceText(bytes, 0, bytes.byteLength) };
+}
+function sectionChunks(bytes) {
+  const lines = splitWithByteOffsets(bytes);
   const sections = [];
   let current = null;
   const ids = /* @__PURE__ */ new Set();
   function flush(end) {
     if (!current) return;
-    const text = markdown.slice(current.start, end);
+    const text = sliceText(bytes, current.start, end);
     if (text.trim().length === 0) {
       current = null;
       return;
@@ -29,59 +35,84 @@ function sectionChunks(markdown) {
   }
   for (const line of lines) {
     const match = HEADING.exec(line.text);
+    const lineEnd = line.offset + line.byteLength;
     if (match) {
       flush(line.offset);
       const id = uniqueId(slugify(match[2] ?? `section-${sections.length + 1}`), ids);
       ids.add(id);
-      current = { id, start: line.offset, end: line.offset + line.text.length };
+      current = { id, start: line.offset, end: lineEnd };
       continue;
     }
     if (current === null) {
       const id = uniqueId("preamble", ids);
       ids.add(id);
-      current = { id, start: line.offset, end: line.offset + line.text.length };
+      current = { id, start: line.offset, end: lineEnd };
     } else {
-      current.end = line.offset + line.text.length;
+      current.end = lineEnd;
     }
   }
-  flush(markdown.length);
+  flush(bytes.byteLength);
   if (sections.length === 0) {
-    return [{ id: "document", textOffset: 0, textLength: markdown.length, text: markdown }];
+    return [documentChunk(bytes)];
   }
   return sections;
 }
-function paragraphChunks(markdown) {
+function paragraphChunks(bytes) {
   const out = [];
   const ids = /* @__PURE__ */ new Set();
+  const separator = encoder.encode("\n\n");
   let cursor = 0;
   let i = 0;
-  while (cursor < markdown.length) {
-    let end = markdown.indexOf("\n\n", cursor);
-    if (end === -1) end = markdown.length;
-    const text = markdown.slice(cursor, end);
+  while (cursor < bytes.byteLength) {
+    let end = indexOfBytes(bytes, separator, cursor);
+    if (end === -1) end = bytes.byteLength;
+    const text = sliceText(bytes, cursor, end);
     if (text.trim().length > 0) {
       const id = uniqueId(slugify(text.split("\n")[0] ?? `p-${i}`), ids);
       ids.add(id);
-      out.push({ id, textOffset: cursor, textLength: text.length, text });
+      out.push({ id, textOffset: cursor, textLength: end - cursor, text });
       i += 1;
     }
-    cursor = end + 2;
+    cursor = end + separator.byteLength;
   }
   if (out.length === 0) {
-    return [{ id: "document", textOffset: 0, textLength: markdown.length, text: markdown }];
+    return [documentChunk(bytes)];
   }
   return out;
 }
-function splitWithOffsets(s) {
+function sliceText(bytes, start, end) {
+  return decoder.decode(bytes.subarray(start, end));
+}
+function splitWithByteOffsets(bytes) {
+  const newline = 10;
   const lines = [];
-  let offset = 0;
-  for (const line of s.split("\n")) {
-    const withNl = line + (offset + line.length < s.length ? "\n" : "");
-    lines.push({ text: withNl, offset });
-    offset += withNl.length;
+  let start = 0;
+  for (let i = 0; i < bytes.byteLength; i += 1) {
+    if (bytes[i] === newline) {
+      const byteLength = i - start + 1;
+      lines.push({ text: sliceText(bytes, start, i + 1), offset: start, byteLength });
+      start = i + 1;
+    }
+  }
+  if (start < bytes.byteLength) {
+    lines.push({ text: sliceText(bytes, start, bytes.byteLength), offset: start, byteLength: bytes.byteLength - start });
   }
   return lines;
 }
+function indexOfBytes(haystack, needle, from) {
+  const last = haystack.byteLength - needle.byteLength;
+  for (let i = from; i <= last; i += 1) {
+    let matched = true;
+    for (let j = 0; j < needle.byteLength; j += 1) {
+      if (haystack[i + j] !== needle[j]) {
+        matched = false;
+        break;
+      }
+    }
+    if (matched) return i;
+  }
+  return -1;
+}
 function slugify(s) {
   return s.toLowerCase().replace(/[^a-z0-9]+/g, "-").replace(/^-+|-+$/g, "").slice(0, 64) || "section";
 }

package/dist/index.js.map

@@ -1 +1 @@
-{"version":3,"sources":["../src/chunk.ts","../src/transformers-backend.ts","../src/types.ts","../src/embed.ts","../src/search.ts","../src/huggingface-backend.ts"],"names":[],"mappings":";;;AAsBA,IAAM,OAAA,GAAU,uBAAA;AAET,SAAS,aAAA,CAAc,QAAA,EAAkB,IAAA,GAAqB,EAAC,EAAoB;AACxF,EAAA,MAAM,IAAA,GAAO,KAAK,IAAA,IAAQ,SAAA;AAC1B,EAAA,IAAI,SAAS,UAAA,EAAY;AACvB,IAAA,OAAO,CAAC,EAAE,EAAA,EAAI,UAAA,EAAY,UAAA,EAAY,CAAA,EAAG,UAAA,EAAY,QAAA,CAAS,MAAA,EAAQ,IAAA,EAAM,QAAA,EAAU,CAAA;AAAA,EACxF;AACA,EAAA,IAAI,SAAS,WAAA,EAAa;AACxB,IAAA,OAAO,gBAAgB,QAAQ,CAAA;AAAA,EACjC;AACA,EAAA,OAAO,cAAc,QAAQ,CAAA;AAC/B;AAEA,SAAS,cAAc,QAAA,EAAmC;AACxD,EAAA,MAAM,KAAA,GAAQ,iBAAiB,QAAQ,CAAA;AACvC,EAAA,MAAM,WAA4B,EAAC;AACnC,EAAA,IAAI,OAAA,GAA6D,IAAA;AACjE,EAAA,MAAM,GAAA,uBAAU,GAAA,EAAY;AAE5B,EAAA,SAAS,MAAM,GAAA,EAAmB;AAChC,IAAA,IAAI,CAAC,OAAA,EAAS;AACd,IAAA,MAAM,IAAA,GAAO,QAAA,CAAS,KAAA,CAAM,OAAA,CAAQ,OAAO,GAAG,CAAA;AAC9C,IAAA,IAAI,IAAA,CAAK,IAAA,EAAK,CAAE,MAAA,KAAW,CAAA,EAAG;AAC5B,MAAA,OAAA,GAAU,IAAA;AACV,MAAA;AAAA,IACF;AACA,IAAA,QAAA,CAAS,IAAA,CAAK,EAAE,EAAA,EAAI,OAAA,CAAQ,EAAA,EAAI,UAAA,EAAY,OAAA,CAAQ,KAAA,EAAO,UAAA,EAAY,GAAA,GAAM,OAAA,CAAQ,KAAA,EAAO,MAAM,CAAA;AAClG,IAAA,OAAA,GAAU,IAAA;AAAA,EACZ;AAEA,EAAA,KAAA,MAAW,QAAQ,KAAA,EAAO;AACxB,IAAA,MAAM,KAAA,GAAQ,OAAA,CAAQ,IAAA,CAAK,IAAA,CAAK,IAAI,CAAA;AACpC,IAAA,IAAI,KAAA,EAAO;AACT,MAAA,KAAA,CAAM,KAAK,MAAM,CAAA;AACjB,MAAA,MAAM,EAAA,GAAK,QAAA,CAAS,OAAA,CAAQ,KAAA,CAAM,CAAC,CAAA,IAAK,CAAA,QAAA,EAAW,QAAA,CAAS,MAAA,GAAS,CAAC,CAAA,CAAE,CAAA,EAAG,GAAG,CAAA;AAC9E,MAAA,GAAA,CAAI,IAAI,EAAE,CAAA;AACV,MAAA,OAAA,GAAU,EAAE,EAAA,EAAI,KAAA,EAAO,IAAA,CAAK,MAAA,EAAQ,KAAK,IAAA,CAAK,MAAA,GAAS,IAAA,CAAK,IAAA,CAAK,MAAA,EAAO;AACxE,MAAA;AAAA,IACF;AACA,IAAA,IAAI,YAAY,IAAA,EAAM;AACpB,MAAA,MAAM,EAAA,GAAK,QAAA,CAAS,UAAA,EAAY,GAAG,CAAA;AACnC,MAAA,GAAA,CAAI,IAAI,EAAE,CAAA;AACV,MAAA,OAAA,GAAU,EAAE,EAAA,EAAI,KAAA,EAAO,IAAA,CAAK,MAAA,EAAQ,KAAK,IAAA,CAAK,MAAA,GAAS,IAAA,CAAK,IAAA,CAAK,MAAA,EAAO;AAAA,IAC1E,CAAA,MAAO;AACL,MAAA,OAAA,CAAQ,GAAA,GAAM,IAAA,CAAK,MAAA,GAAS,IAAA,CAAK,IAAA,CAAK,MAAA;AAAA,IACxC;AAAA,EACF;AACA,EAAA,KAAA,CAAM,SAAS,MAAM,CAAA;AAErB,EAAA,IAAI,QAAA,CAAS,WAAW,CAAA,EAAG;AACzB,IAAA,OAAO,CAAC,EAAE,EAAA,EAAI,UAAA,EAAY,UAAA,EAAY,CAAA,EAAG,UAAA,EAAY,QAAA,CAAS,MAAA,EAAQ,IAAA,EAAM,QAAA,EAAU,CAAA;AAAA,EACxF;AACA,EAAA,OAAO,QAAA;AACT;AAEA,SAAS,gBAAgB,QAAA,EAAmC;AAC1D,EAAA,MAAM,MAAuB,EAAC;AAC9B,EAAA,MAAM,GAAA,uBAAU,GAAA,EAAY;AAC5B,EAAA,IAAI,MAAA,GAAS,CAAA;AACb,EAAA,IAAI,CAAA,GAAI,CAAA;AACR,EAAA,OAAO,MAAA,GAAS,SAAS,MAAA,EAAQ;AAC/B,IAAA,IAAI,GAAA,GAAM,QAAA,CAAS,OAAA,CAAQ,MAAA,EAAQ,MAAM,CAAA;AACzC,IAAA,IAAI,GAAA,KAAQ,EAAA,EAAI,GAAA,GAAM,QAAA,CAAS,MAAA;AAC/B,IAAA,MAAM,IAAA,GAAO,QAAA,CAAS,KAAA,CAAM,MAAA,EAAQ,GAAG,CAAA;AACvC,IAAA,IAAI,IAAA,CAAK,IAAA,EAAK,CAAE,MAAA,GAAS,CAAA,EAAG;AAC1B,MAAA,MAAM,EAAA,GAAK,QAAA,CAAS,OAAA,CAAQ,IAAA,CAAK,KAAA,CAAM,IAAI,CAAA,CAAE,CAAC,CAAA,IAAK,CAAA,EAAA,EAAK,CAAC,CAAA,CAAE,GAAG,GAAG,CAAA;AACjE,MAAA,GAAA,CAAI,IAAI,EAAE,CAAA;AACV,MAAA,GAAA,CAAI,IAAA,CAAK,EAAE,EAAA,EAAI,UAAA,EAAY,QAAQ,UAAA,EAAY,IAAA,CAAK,MAAA,EAAQ,IAAA,EAAM,CAAA;AAClE,MAAA,CAAA,IAAK,CAAA;AAAA,IACP;AACA,IAAA,MAAA,GAAS,GAAA,GAAM,CAAA;AAAA,EACjB;AACA,EAAA,IAAI,GAAA,CAAI,WAAW,CAAA,EAAG;AACpB,IAAA,OAAO,CAAC,EAAE,EAAA,EAAI,UAAA,EAAY,UAAA,EAAY,CAAA,EAAG,UAAA,EAAY,QAAA,CAAS,MAAA,EAAQ,IAAA,EAAM,QAAA,EAAU,CAAA;AAAA,EACxF;AACA,EAAA,OAAO,GAAA;AACT;AAEA,SAAS,iBAAiB,CAAA,EAA+C;AACvE,EAAA,MAAM,QAA4C,EAAC;AACnD,EAAA,IAAI,MAAA,GAAS,CAAA;AACb,EAAA,KAAA,MAAW,IAAA,IAAQ,CAAA,CAAE,KAAA,CAAM,IAAI,CAAA,EAAG;AAChC,IAAA,MAAM,SAAS,IAAA,IAAQ,MAAA,GAAS,KAAK,MAAA,GAAS,CAAA,CAAE,SAAS,IAAA,GAAO,EAAA,CAAA;AAChE,IAAA,KAAA,CAAM,IAAA,CAAK,EAAE,IAAA,EAAM,MAAA,EAAQ,QAAQ,CAAA;AACnC,IAAA,MAAA,IAAU,MAAA,CAAO,MAAA;AAAA,EACnB;AACA,EAAA,OAAO,KAAA;AACT;AAEA,SAAS,QAAQ,CAAA,EAAmB;AAClC,EAAA,OACE,CAAA,CACG,WAAA,EAAY,CACZ,OAAA,CAAQ,eAAe,GAAG,CAAA,CAC1B,OAAA,CAAQ,UAAA,EAAY,EAAE,CAAA,CACtB,KAAA,CAAM,CAAA,EAAG,EAAE,CAAA,IAAK,SAAA;AAEvB;AAEA,SAAS,QAAA,CAAS,MAAc,KAAA,EAA4B;AAC1D,EAAA,IAAI,CAAC,KAAA,CAAM,GAAA,CAAI,IAAI,GAAG,OAAO,IAAA;AAC7B,EAAA,IAAI,CAAA,GAAI,CAAA;AACR,EAAA,OAAO,KAAA,CAAM,IAAI,CAAA,EAAG,IAAI,IAAI,CAAC,CAAA,CAAE,GAAG,CAAA,IAAK,CAAA;AACvC,EAAA,OAAO,CAAA,EAAG,IAAI,CAAA,CAAA,EAAI,CAAC,CAAA,CAAA;AACrB;AC3GO,SAAS,0BAA0B,IAAA,EAAoD;AAC5F,EAAA,IAAI,IAAA,CAAK,sBAAsB,KAAA,EAAO;AACpC,IAAA,GAAA,CAAI,iBAAA,GAAoB,KAAA;AAAA,EAC1B;AAEA,EAAA,IAAI,eAAA,GAA2C,IAAA;AAC/C,EAAA,IAAI,iBAAA,GAAmC,IAAA;AAEvC,EAAA,SAAS,WAAA,GAAgC;AACvC,IAAA,IAAI,CAAC,eAAA,EAAiB;AACpB,MAAA,MAAM,SAAA,GAAqC,EAAE,KAAA,EAAO,IAAA,CAAK,SAAS,MAAA,EAAO;AACzE,MAAA,IAAI,IAAA,CAAK,MAAA,KAAW,MAAA,EAAW,SAAA,CAAU,SAAS,IAAA,CAAK,MAAA;AACvD,MAAA,eAAA,GAAkB,QAAA,CAAS,oBAAA,EAAsB,IAAA,CAAK,KAAA,EAAO,SAAkB,CAAA;AAAA,IACjF;AACA,IAAA,OAAO,eAAA;AAAA,EACT;AAEA,EAAA,MAAM,OAAA,GAA4B;AAAA,IAChC,OAAO,IAAA,CAAK,KAAA;AAAA,IACZ,aAAA,EAAe,KAAK,aAAA,IAAiB,MAAA;AAAA,IACrC,MAAA,EAAQ,KAAK,MAAA,IAAU,QAAA;AAAA,IACvB,UAAA,EAAY,IAAA;AAAA,IACZ,MAAM,MAAM,KAAA,EAA2C;AACrD,MAAA,IAAI,KAAA,CAAM,WAAW,CAAA,EAAG;AACtB,QAAA,OAAO,EAAE,OAAA,EAAS,IAAI,SAAA,EAAW,IAAA,CAAK,aAAa,CAAA,EAAE;AAAA,MACvD;AACA,MAAA,MAAM,IAAA,GAAQ,MAAM,WAAA,EAAY;AAIhC,MAAA,MAAM,MAAA,GAAS,MAAM,IAAA,CAAK,KAAA,EAAO,EAAE,OAAA,EAAS,MAAA,EAAQ,SAAA,EAAW,IAAA,EAAM,CAAA;AACrE,MAAA,MAAM,OAAO,MAAA,CAAO,IAAA;AACpB,MAAA,MAAM,YAAY,MAAA,CAAO,IAAA,CAAK,MAAA,CAAO,IAAA,CAAK,SAAS,CAAC,CAAA;AACpD,MAAA,IAAI,iBAAA,KAAsB,IAAA,IAAQ,iBAAA,KAAsB,SAAA,EAAW;AACjE,QAAA,MAAM,IAAI,KAAA,CAAM,CAAA,0CAAA,EAA6C,iBAAiB,CAAA,MAAA,EAAS,SAAS,CAAA,CAAE,CAAA;AAAA,MACpG;AACA,MAAA,iBAAA,GAAoB,SAAA;AACpB,MAAA,MAAM,UAA0B,EAAC;AACjC,MAAA,KAAA,IAAS,IAAI,CAAA,EAAG,CAAA,GAAI,KAAA,CAAM,MAAA,EAAQ,KAAK,CAAA,EAAG;AACxC,QAAA,OAAA,CAAQ,IAAA,CAAK,IAAI,YAAA,CAAa,IAAA,CAAK,MAAA,EAAQ,IAAA,CAAK,UAAA,GAAa,CAAA,GAAI,SAAA,GAAY,CAAA,EAAG,SAAS,CAAA,CAAE,OAAO,CAAA;AAAA,MACpG;AACA,MAAA,OAAO,EAAE,SAAS,SAAA,EAAU;AAAA,IAC9B;AAAA,GACF;AACA,EAAA,IAAI,IAAA,CAAK,SAAA,KAAc,MAAA,EAAW,OAAA,CAAQ,YAAY,IAAA,CAAK,SAAA;AAC3D,EAAA,OAAO,OAAA;AACT;;;ACvCO,IAAM,aAAA,GAAgB;AACtB,IAAM,uBAAA,GAA0B;;;ACNvC,eAAsB,KAAA,CAAM,QAAA,EAAkB,IAAA,GAAqB,EAAC,EAA+B;AACjG,EAAA,MAAM,MAAA,GAAS,cAAc,QAAA,EAAU,EAAE,MAAM,IAAA,CAAK,QAAA,IAAY,WAAW,CAAA;AAC3E,EAAA,MAAM,OAAA,GACJ,KAAK,OAAA,IACL,yBAAA;AAAA,IACE,KAAK,aAAA,KAAkB,MAAA,GACnB,EAAE,KAAA,EAAO,KAAK,KAAA,IAAS,aAAA,EAAe,aAAA,EAAe,IAAA,CAAK,eAAc,GACxE,EAAE,KAAA,EAAO,IAAA,CAAK,SAAS,aAAA;AAAc,GAC3C;AAEF,EAAA,MAAM,MAAA,GAAS,MAAM,OAAA,CAAQ,KAAA,CAAM,MAAA,CAAO,IAAI,CAAC,CAAA,KAAM,CAAA,CAAE,IAAI,CAAC,CAAA;AAC5D,EAAA,IAAI,MAAA,CAAO,OAAA,CAAQ,MAAA,KAAW,MAAA,CAAO,MAAA,EAAQ;AAC3C,IAAA,MAAM,IAAI,MAAM,CAAA,iBAAA,EAAoB,MAAA,CAAO,QAAQ,MAAM,CAAA,aAAA,EAAgB,MAAA,CAAO,MAAM,CAAA,OAAA,CAAS,CAAA;AAAA,EACjG;AAEA,EAAA,MAAM,eAAA,GAAoC,MAAA,CAAO,GAAA,CAAI,CAAC,GAAG,CAAA,MAAO;AAAA,IAC9D,IAAI,CAAA,CAAE,EAAA;AAAA,IACN,YAAY,CAAA,CAAE,UAAA;AAAA,IACd,YAAY,CAAA,CAAE,UAAA;AAAA,IACd,MAAA,EAAQ,MAAA,CAAO,OAAA,CAAQ,CAAC;AAAA,GAC1B,CAAE,CAAA;AAEF,EAAA,MAAM,KAAA,GAAwB;AAAA,IAC5B,OAAO,OAAA,CAAQ,KAAA;AAAA,IACf,eAAe,OAAA,CAAQ,aAAA;AAAA,IACvB,WAAW,MAAA,CAAO,SAAA;AAAA,IAClB,QAAQ,OAAA,CAAQ,MAAA;AAAA,IAChB,YAAY,OAAA,CAAQ,UAAA;AAAA,IACpB,QAAA,EAAU,KAAK,QAAA,IAAY,SAAA;AAAA,IAC3B,MAAA,EAAQ;AAAA,GACV;AAEA,EAAA,OAAO,EAAE,aAAA,EAAe,CAAA,EAAG,MAAA,EAAQ,CAAC,KAAK,CAAA,EAAE;AAC7C;;;AC/BO,SAAS,cAAA,CACd,OAAA,EACA,WAAA,EACA,IAAA,GAAsB,EAAC,EACV;AACb,EAAA,MAAM,KAAA,GAAQ,SAAA,CAAU,OAAA,EAAS,IAAA,CAAK,KAAK,CAAA;AAC3C,EAAA,IAAI,CAAC,KAAA,EAAO,MAAM,IAAI,MAAM,mCAAmC,CAAA;AAC/D,EAAA,IAAI,WAAA,CAAY,MAAA,KAAW,KAAA,CAAM,SAAA,EAAW;AAC1C,IAAA,MAAM,IAAI,KAAA,CAAM,CAAA,uBAAA,EAA0B,WAAA,CAAY,MAAM,CAAA,sBAAA,EAAyB,KAAA,CAAM,KAAK,CAAA,EAAA,EAAK,KAAA,CAAM,SAAS,CAAA,CAAA,CAAG,CAAA;AAAA,EACzH;AAEA,EAAA,MAAM,CAAA,GAAI,KAAK,CAAA,IAAK,CAAA;AACpB,EAAA,MAAM,MAAA,GAAsB,KAAA,CAAM,MAAA,CAAO,GAAA,CAAI,CAAC,CAAA,MAAO;AAAA,IACnD,YAAY,KAAA,CAAM,KAAA;AAAA,IAClB,SAAS,CAAA,CAAE,EAAA;AAAA,IACX,YAAY,CAAA,CAAE,UAAA;AAAA,IACd,YAAY,CAAA,CAAE,UAAA;AAAA,IACd,OAAO,UAAA,CAAW,WAAA,EAAa,CAAA,CAAE,MAAA,EAAQ,MAAM,MAAM;AAAA,GACvD,CAAE,CAAA;AAGF,EAAA,MAAM,KAAA,GAAQ,KAAA,CAAM,MAAA,KAAW,WAAA,GAAc,CAAA,GAAI,EAAA;AACjD,EAAA,MAAA,CAAO,IAAA,CAAK,CAAC,CAAA,EAAG,CAAA,KAAM,SAAS,CAAA,CAAE,KAAA,GAAQ,EAAE,KAAA,CAAM,CAAA;AACjD,EAAA,OAAO,MAAA,CAAO,KAAA,CAAM,CAAA,EAAG,CAAC,CAAA;AAC1B;AAEA,SAAS,SAAA,CAAU,SAA4B,SAAA,EAAgD;AAC7F,EAAA,IAAI,SAAA,SAAkB,OAAA,CAAQ,MAAA,CAAO,KAAK,CAAC,CAAA,KAAM,CAAA,CAAE,KAAA,KAAU,SAAS,CAAA;AACtE,EAAA,OAAO,OAAA,CAAQ,OAAO,CAAC,CAAA;AACzB;AAEA,SAAS,UAAA,CAAW,CAAA,EAAiB,CAAA,EAAiB,MAAA,EAAiC;AACrF,EAAA,IAAI,WAAW,WAAA,EAAa;AAC1B,IAAA,IAAI,GAAA,GAAM,CAAA;AACV,IAAA,KAAA,IAAS,IAAI,CAAA,EAAG,CAAA,GAAI,CAAA,CAAE,MAAA,EAAQ,KAAK,CAAA,EAAG;AACpC,MAAA,MAAM,CAAA,GAAI,CAAA,CAAE,CAAC,CAAA,GAAK,EAAE,CAAC,CAAA;AACrB,MAAA,GAAA,IAAO,CAAA,GAAI,CAAA;AAAA,IACb;AACA,IAAA,OAAO,IAAA,CAAK,KAAK,GAAG,CAAA;AAAA,EACtB;AACA,EAAA,IAAI,GAAA,GAAM,CAAA;AACV,EAAA,KAAA,IAAS,CAAA,GAAI,CAAA,EAAG,CAAA,GAAI,CAAA,CAAE,MAAA,EAAQ,CAAA,IAAK,CAAA,EAAG,GAAA,IAAO,CAAA,CAAE,CAAC,CAAA,GAAK,CAAA,CAAE,CAAC,CAAA;AACxD,EAAA,IAAI,MAAA,KAAW,OAAO,OAAO,GAAA;AAE7B,EAAA,IAAI,EAAA,GAAK,CAAA;AACT,EAAA,IAAI,EAAA,GAAK,CAAA;AACT,EAAA,KAAA,IAAS,IAAI,CAAA,EAAG,CAAA,GAAI,CAAA,CAAE,MAAA,EAAQ,KAAK,CAAA,EAAG;AACpC,IAAA,EAAA,IAAM,CAAA,CAAE,CAAC,CAAA,GAAK,CAAA,CAAE,CAAC,CAAA;AACjB,IAAA,EAAA,IAAM,CAAA,CAAE,CAAC,CAAA,GAAK,CAAA,CAAE,CAAC,CAAA;AAAA,EACnB;AACA,EAAA,MAAM,QAAQ,IAAA,CAAK,IAAA,CAAK,EAAE,CAAA,GAAI,IAAA,CAAK,KAAK,EAAE,CAAA;AAC1C,EAAA,OAAO,KAAA,KAAU,CAAA,GAAI,CAAA,GAAI,GAAA,GAAM,KAAA;AACjC;;;AClDA,IAAM,gBAAA,GAAmB,mDAAA;AAElB,SAAS,yBAAyB,IAAA,EAAmD;AAC1F,EAAA,MAAM,QAAQ,IAAA,CAAK,KAAA,IAAS,QAAQ,GAAA,CAAI,QAAA,IAAY,QAAQ,GAAA,CAAI,iBAAA;AAChE,EAAA,IAAI,CAAC,KAAA,EAAO;AACV,IAAA,MAAM,IAAI,MAAM,0EAA0E,CAAA;AAAA,EAC5F;AACA,EAAA,MAAM,OAAA,GAAU,KAAK,OAAA,IAAW,gBAAA;AAChC,EAAA,MAAM,MAAM,CAAA,EAAG,OAAO,IAAI,SAAA,CAAU,IAAA,CAAK,KAAK,CAAC,CAAA,4BAAA,CAAA;AAE/C,EAAA,MAAM,OAAA,GAA4B;AAAA,IAChC,OAAO,IAAA,CAAK,KAAA;AAAA,IACZ,aAAA,EAAe,KAAK,aAAA,IAAiB,MAAA;AAAA,IACrC,MAAA,EAAQ,KAAK,MAAA,IAAU,QAAA;AAAA,IACvB,UAAA,EAAY,IAAA;AAAA,IACZ,MAAM,MAAM,KAAA,EAA2C;AACrD,MAAA,IAAI,KAAA,CAAM,WAAW,CAAA,EAAG;AACtB,QAAA,OAAO,EAAE,OAAA,EAAS,IAAI,SAAA,EAAW,IAAA,CAAK,aAAa,CAAA,EAAE;AAAA,MACvD;AACA,MAAA,MAAM,IAAA,GAAO,IAAA,CAAK,SAAA,CAAU,EAAE,MAAA,EAAQ,KAAA,EAAO,OAAA,EAAS,EAAE,cAAA,EAAgB,IAAA,EAAK,EAAG,CAAA;AAChF,MAAA,MAAM,GAAA,GAAM,MAAM,KAAA,CAAM,GAAA,EAAK;AAAA,QAC3B,MAAA,EAAQ,MAAA;AAAA,QACR,OAAA,EAAS;AAAA,UACP,cAAA,EAAgB,kBAAA;AAAA,UAChB,aAAA,EAAe,UAAU,KAAK,CAAA;AAAA,SAChC;AAAA,QACA;AAAA,OACD,CAAA;AACD,MAAA,IAAI,CAAC,IAAI,EAAA,EAAI;AACX,QAAA,MAAM,SAAS,MAAM,GAAA,CAAI,MAAK,CAAE,KAAA,CAAM,MAAM,EAAE,CAAA;AAC9C,QAAA,MAAM,IAAI,KAAA,CAAM,CAAA,iBAAA,EAAoB,GAAA,CAAI,MAAM,CAAA,KAAA,EAAQ,IAAA,CAAK,KAAK,CAAA,EAAA,EAAK,MAAA,IAAU,GAAA,CAAI,UAAU,CAAA,CAAE,CAAA;AAAA,MACjG;AACA,MAAA,MAAM,GAAA,GAAO,MAAM,GAAA,CAAI,IAAA,EAAK;AAC5B,MAAA,MAAM,MAAA,GAAS,eAAA,CAAgB,GAAA,EAAK,KAAA,CAAM,MAAM,CAAA;AAChD,MAAA,MAAM,YAAY,MAAA,CAAO,CAAC,CAAA,EAAG,MAAA,IAAU,KAAK,SAAA,IAAa,CAAA;AACzD,MAAA,MAAM,OAAA,GAAU,MAAA,CAAO,GAAA,CAAI,CAAC,CAAA,KAAM,UAAU,YAAA,CAAa,IAAA,CAAK,CAAC,CAAC,CAAC,CAAA;AACjE,MAAA,OAAO,EAAE,SAAS,SAAA,EAAU;AAAA,IAC9B;AAAA,GACF;AACA,EAAA,IAAI,IAAA,CAAK,SAAA,KAAc,MAAA,EAAW,OAAA,CAAQ,YAAY,IAAA,CAAK,SAAA;AAC3D,EAAA,OAAO,OAAA;AACT;AAWA,SAAS,eAAA,CAAgB,KAAc,aAAA,EAAmC;AACxE,EAAA,IAAI,CAAC,KAAA,CAAM,OAAA,CAAQ,GAAG,CAAA,EAAG;AACvB,IAAA,MAAM,IAAI,MAAM,2CAA2C,CAAA;AAAA,EAC7D;AACA,EAAA,IAAI,GAAA,CAAI,MAAA,KAAW,CAAA,EAAG,OAAO,EAAC;AAE9B,EAAA,MAAM,KAAA,GAAQ,IAAI,CAAC,CAAA;AACnB,EAAA,IAAI,OAAO,UAAU,QAAA,EAAU;AAC7B,IAAA,IAAI,kBAAkB,CAAA,EAAG;AACvB,MAAA,MAAM,IAAI,KAAA,CAAM,CAAA,yCAAA,EAA4C,aAAa,CAAA,CAAE,CAAA;AAAA,IAC7E;AACA,IAAA,OAAO,CAAC,GAAe,CAAA;AAAA,EACzB;AACA,EAAA,IAAI,KAAA,CAAM,OAAA,CAAQ,KAAK,CAAA,KAAM,KAAA,CAAM,MAAA,KAAW,CAAA,IAAK,OAAO,KAAA,CAAM,CAAC,CAAA,KAAM,QAAA,CAAA,EAAW;AAChF,IAAA,OAAO,GAAA;AAAA,EACT;AACA,EAAA,IAAI,KAAA,CAAM,QAAQ,KAAK,CAAA,IAAK,MAAM,OAAA,CAAQ,KAAA,CAAM,CAAC,CAAC,CAAA,EAAG;AAEnD,IAAA,OAAQ,GAAA,CAAqB,IAAI,QAAQ,CAAA;AAAA,EAC3C;AACA,EAAA,MAAM,IAAI,MAAM,+CAA+C,CAAA;AACjE;AAEA,SAAS,SAAS,MAAA,EAA8B;AAC9C,EAAA,IAAI,MAAA,CAAO,MAAA,KAAW,CAAA,EAAG,OAAO,EAAC;AACjC,EAAA,MAAM,GAAA,GAAM,MAAA,CAAO,CAAC,CAAA,CAAG,MAAA;AACvB,EAAA,MAAM,MAAM,IAAI,KAAA,CAAc,GAAG,CAAA,CAAE,KAAK,CAAC,CAAA;AACzC,EAAA,KAAA,MAAW,KAAK,MAAA,EAAQ;AACtB,IAAA,KAAA,IAAS,CAAA,GAAI,CAAA,EAAG,CAAA,GAAI,GAAA,EAAK,CAAA,IAAK,GAAG,GAAA,CAAI,CAAC,CAAA,IAAM,CAAA,CAAE,CAAC,CAAA;AAAA,EACjD;AACA,EAAA,KAAA,IAAS,CAAA,GAAI,GAAG,CAAA,GAAI,GAAA,EAAK,KAAK,CAAA,EAAG,GAAA,CAAI,CAAC,CAAA,IAAM,MAAA,CAAO,MAAA;AACnD,EAAA,OAAO,GAAA;AACT;AAEA,SAAS,UAAU,CAAA,EAA+B;AAChD,EAAA,IAAI,GAAA,GAAM,CAAA;AACV,EAAA,KAAA,IAAS,CAAA,GAAI,CAAA,EAAG,CAAA,GAAI,CAAA,CAAE,MAAA,EAAQ,CAAA,IAAK,CAAA,EAAG,GAAA,IAAO,CAAA,CAAE,CAAC,CAAA,GAAK,CAAA,CAAE,CAAC,CAAA;AACxD,EAAA,MAAM,IAAA,GAAO,IAAA,CAAK,IAAA,CAAK,GAAG,CAAA;AAC1B,EAAA,IAAI,IAAA,KAAS,GAAG,OAAO,CAAA;AACvB,EAAA,KAAA,IAAS,CAAA,GAAI,CAAA,EAAG,CAAA,GAAI,CAAA,CAAE,MAAA,EAAQ,CAAA,IAAK,CAAA,EAAG,CAAA,CAAE,CAAC,CAAA,GAAI,CAAA,CAAE,CAAC,CAAA,GAAK,IAAA;AACrD,EAAA,OAAO,CAAA;AACT","file":"index.js","sourcesContent":["/**\n * Section-based markdown chunker.\n *\n * Splits a markdown document on ATX headings (`#`, `##`, ...). Each chunk\n * carries the byte offset and length into the original UTF-8 source so a\n * downstream consumer can map a vector hit back to the exact substring\n * without re-tokenising. Pre-heading content becomes a \"preamble\" chunk.\n */\n\nexport type ChunkingMode = 'document' | 'section' | 'paragraph';\n\nexport interface MarkdownChunk {\n  id: string;\n  textOffset: number;\n  textLength: number;\n  text: string;\n}\n\nexport interface ChunkOptions {\n  mode?: ChunkingMode;\n}\n\nconst HEADING = /^(#{1,6})\\s+(.+?)\\s*$/;\n\nexport function chunkMarkdown(markdown: string, opts: ChunkOptions = {}): MarkdownChunk[] {\n  const mode = opts.mode ?? 'section';\n  if (mode === 'document') {\n    return [{ id: 'document', textOffset: 0, textLength: markdown.length, text: markdown }];\n  }\n  if (mode === 'paragraph') {\n    return paragraphChunks(markdown);\n  }\n  return sectionChunks(markdown);\n}\n\nfunction sectionChunks(markdown: string): MarkdownChunk[] {\n  const lines = splitWithOffsets(markdown);\n  const sections: MarkdownChunk[] = [];\n  let current: { id: string; start: number; end: number } | null = null;\n  const ids = new Set<string>();\n\n  function flush(end: number): void {\n    if (!current) return;\n    const text = markdown.slice(current.start, end);\n    if (text.trim().length === 0) {\n      current = null;\n      return;\n    }\n    sections.push({ id: current.id, textOffset: current.start, textLength: end - current.start, text });\n    current = null;\n  }\n\n  for (const line of lines) {\n    const match = HEADING.exec(line.text);\n    if (match) {\n      flush(line.offset);\n      const id = uniqueId(slugify(match[2] ?? `section-${sections.length + 1}`), ids);\n      ids.add(id);\n      current = { id, start: line.offset, end: line.offset + line.text.length };\n      continue;\n    }\n    if (current === null) {\n      const id = uniqueId('preamble', ids);\n      ids.add(id);\n      current = { id, start: line.offset, end: line.offset + line.text.length };\n    } else {\n      current.end = line.offset + line.text.length;\n    }\n  }\n  flush(markdown.length);\n\n  if (sections.length === 0) {\n    return [{ id: 'document', textOffset: 0, textLength: markdown.length, text: markdown }];\n  }\n  return sections;\n}\n\nfunction paragraphChunks(markdown: string): MarkdownChunk[] {\n  const out: MarkdownChunk[] = [];\n  const ids = new Set<string>();\n  let cursor = 0;\n  let i = 0;\n  while (cursor < markdown.length) {\n    let end = markdown.indexOf('\\n\\n', cursor);\n    if (end === -1) end = markdown.length;\n    const text = markdown.slice(cursor, end);\n    if (text.trim().length > 0) {\n      const id = uniqueId(slugify(text.split('\\n')[0] ?? `p-${i}`), ids);\n      ids.add(id);\n      out.push({ id, textOffset: cursor, textLength: text.length, text });\n      i += 1;\n    }\n    cursor = end + 2;\n  }\n  if (out.length === 0) {\n    return [{ id: 'document', textOffset: 0, textLength: markdown.length, text: markdown }];\n  }\n  return out;\n}\n\nfunction splitWithOffsets(s: string): { text: string; offset: number }[] {\n  const lines: { text: string; offset: number }[] = [];\n  let offset = 0;\n  for (const line of s.split('\\n')) {\n    const withNl = line + (offset + line.length < s.length ? '\\n' : '');\n    lines.push({ text: withNl, offset });\n    offset += withNl.length;\n  }\n  return lines;\n}\n\nfunction slugify(s: string): string {\n  return (\n    s\n      .toLowerCase()\n      .replace(/[^a-z0-9]+/g, '-')\n      .replace(/^-+|-+$/g, '')\n      .slice(0, 64) || 'section'\n  );\n}\n\nfunction uniqueId(base: string, taken: Set<string>): string {\n  if (!taken.has(base)) return base;\n  let n = 2;\n  while (taken.has(`${base}-${n}`)) n += 1;\n  return `${base}-${n}`;\n}\n","/**\n * transformers.js backend (browser + Node + Bun via @huggingface/transformers).\n *\n * Loads the model lazily and reuses the pipeline across calls. Models are\n * cached on disk by transformers.js itself (HF cache layout).\n */\n\nimport { pipeline, env } from '@huggingface/transformers';\nimport type { EmbeddingBackend, EmbeddingBackendOptions, EmbeddingMatrix } from './types.js';\n\ninterface TransformersBackendOptions extends EmbeddingBackendOptions {\n  /** Override Xenova quantisation. Defaults to fp32 for closest fidelity. */\n  dtype?: 'fp32' | 'fp16' | 'q8' | 'q4';\n  /** Force a backend device (e.g. 'cpu', 'gpu', 'wasm'). Defaults to auto. */\n  device?: 'cpu' | 'gpu' | 'wasm' | 'webgpu';\n  /** Allow remote model downloads. true by default. */\n  allowRemoteModels?: boolean;\n}\n\nexport function createTransformersBackend(opts: TransformersBackendOptions): EmbeddingBackend {\n  if (opts.allowRemoteModels === false) {\n    env.allowRemoteModels = false;\n  }\n\n  let pipelinePromise: Promise<unknown> | null = null;\n  let resolvedDimension: number | null = null;\n\n  function getPipeline(): Promise<unknown> {\n    if (!pipelinePromise) {\n      const modelOpts: Record<string, unknown> = { dtype: opts.dtype ?? 'fp32' };\n      if (opts.device !== undefined) modelOpts.device = opts.device;\n      pipelinePromise = pipeline('feature-extraction', opts.model, modelOpts as never);\n    }\n    return pipelinePromise;\n  }\n\n  const backend: EmbeddingBackend = {\n    model: opts.model,\n    modelRevision: opts.modelRevision ?? 'main',\n    metric: opts.metric ?? 'cosine',\n    normalized: true,\n    async embed(texts: string[]): Promise<EmbeddingMatrix> {\n      if (texts.length === 0) {\n        return { vectors: [], dimension: opts.dimension ?? 0 };\n      }\n      const pipe = (await getPipeline()) as (\n        texts: string[],\n        opts: { pooling: 'mean'; normalize: boolean },\n      ) => Promise<{ data: Float32Array; dims: number[] }>;\n      const tensor = await pipe(texts, { pooling: 'mean', normalize: true });\n      const data = tensor.data;\n      const dimension = tensor.dims[tensor.dims.length - 1] as number;\n      if (resolvedDimension !== null && resolvedDimension !== dimension) {\n        throw new Error(`Model emitted inconsistent dimension: had ${resolvedDimension}, now ${dimension}`);\n      }\n      resolvedDimension = dimension;\n      const vectors: Float32Array[] = [];\n      for (let i = 0; i < texts.length; i += 1) {\n        vectors.push(new Float32Array(data.buffer, data.byteOffset + i * dimension * 4, dimension).slice());\n      }\n      return { vectors, dimension };\n    },\n  };\n  if (opts.dimension !== undefined) backend.dimension = opts.dimension;\n  return backend;\n}\n","import type { EmbeddingMetric } from '@cvfile/sdk';\n\nexport interface EmbeddingMatrix {\n  vectors: Float32Array[];\n  dimension: number;\n}\n\nexport interface EmbeddingBackend {\n  model: string;\n  modelRevision: string;\n  /** Optional pre-declared dimension; the backend may override after first call. */\n  dimension?: number;\n  metric: EmbeddingMetric;\n  normalized: boolean;\n  embed(texts: string[]): Promise<EmbeddingMatrix>;\n}\n\nexport interface EmbeddingBackendOptions {\n  model: string;\n  modelRevision?: string;\n  /** Optional pre-known dimension (e.g. 1024 for bge-m3). */\n  dimension?: number;\n  metric?: EmbeddingMetric;\n}\n\n/** Recommended default per spec §5: BAAI BGE-M3, MIT, multilingual, 1024-dim. */\nexport const DEFAULT_MODEL = 'Xenova/bge-m3';\nexport const DEFAULT_MODEL_DIMENSION = 1024;\n","/**\n * High-level embed() API: markdown in, EmbeddingsPayload out, ready to drop\n * into pack({ embeddings: ... }).\n */\n\nimport type { EmbeddingChunk, EmbeddingSpace, EmbeddingsPayload } from '@cvfile/sdk';\nimport { chunkMarkdown, type ChunkingMode } from './chunk.js';\nimport { createTransformersBackend } from './transformers-backend.js';\nimport { DEFAULT_MODEL, type EmbeddingBackend } from './types.js';\n\nexport interface EmbedOptions {\n  /** HF model id; defaults to BGE-M3 (Xenova/bge-m3). */\n  model?: string;\n  /** Pinned model revision; recorded in the payload for reproducibility. */\n  modelRevision?: string;\n  /** Chunking strategy; default 'section'. */\n  chunking?: ChunkingMode;\n  /** Bring-your-own backend (e.g. an OpenAI/Voyage adapter). */\n  backend?: EmbeddingBackend;\n}\n\nexport async function embed(markdown: string, opts: EmbedOptions = {}): Promise<EmbeddingsPayload> {\n  const chunks = chunkMarkdown(markdown, { mode: opts.chunking ?? 'section' });\n  const backend =\n    opts.backend ??\n    createTransformersBackend(\n      opts.modelRevision !== undefined\n        ? { model: opts.model ?? DEFAULT_MODEL, modelRevision: opts.modelRevision }\n        : { model: opts.model ?? DEFAULT_MODEL },\n    );\n\n  const matrix = await backend.embed(chunks.map((c) => c.text));\n  if (matrix.vectors.length !== chunks.length) {\n    throw new Error(`Backend returned ${matrix.vectors.length} vectors for ${chunks.length} chunks`);\n  }\n\n  const embeddingChunks: EmbeddingChunk[] = chunks.map((c, i) => ({\n    id: c.id,\n    textOffset: c.textOffset,\n    textLength: c.textLength,\n    vector: matrix.vectors[i]!,\n  }));\n\n  const space: EmbeddingSpace = {\n    model: backend.model,\n    modelRevision: backend.modelRevision,\n    dimension: matrix.dimension,\n    metric: backend.metric,\n    normalized: backend.normalized,\n    chunking: opts.chunking ?? 'section',\n    chunks: embeddingChunks,\n  };\n\n  return { formatVersion: 1, spaces: [space] };\n}\n","/**\n * Pure similarity search over an EmbeddingsPayload. Vector input is the\n * caller's responsibility (encode the query with the same model that was\n * used to populate the space, ideally pulled from space.model).\n */\n\nimport type { EmbeddingMetric, EmbeddingSpace, EmbeddingsPayload } from '@cvfile/sdk';\n\nexport interface SearchHit {\n  spaceModel: string;\n  chunkId: string;\n  textOffset: number;\n  textLength: number;\n  score: number;\n}\n\nexport interface SearchOptions {\n  /** Restrict to a specific embedding space; default = first space. */\n  model?: string;\n  /** Top-k results; default 5. */\n  k?: number;\n}\n\nexport function searchSemantic(\n  payload: EmbeddingsPayload,\n  queryVector: Float32Array,\n  opts: SearchOptions = {},\n): SearchHit[] {\n  const space = pickSpace(payload, opts.model);\n  if (!space) throw new Error('No matching embedding space found');\n  if (queryVector.length !== space.dimension) {\n    throw new Error(`Query vector dimension ${queryVector.length} does not match space ${space.model} (${space.dimension})`);\n  }\n\n  const k = opts.k ?? 5;\n  const scored: SearchHit[] = space.chunks.map((c) => ({\n    spaceModel: space.model,\n    chunkId: c.id,\n    textOffset: c.textOffset,\n    textLength: c.textLength,\n    score: similarity(queryVector, c.vector, space.metric),\n  }));\n\n  // Higher score = better for cosine/dot; lower for euclidean.\n  const order = space.metric === 'euclidean' ? 1 : -1;\n  scored.sort((a, b) => order * (a.score - b.score));\n  return scored.slice(0, k);\n}\n\nfunction pickSpace(payload: EmbeddingsPayload, modelHint?: string): EmbeddingSpace | undefined {\n  if (modelHint) return payload.spaces.find((s) => s.model === modelHint);\n  return payload.spaces[0];\n}\n\nfunction similarity(a: Float32Array, b: Float32Array, metric: EmbeddingMetric): number {\n  if (metric === 'euclidean') {\n    let sum = 0;\n    for (let i = 0; i < a.length; i += 1) {\n      const d = a[i]! - b[i]!;\n      sum += d * d;\n    }\n    return Math.sqrt(sum);\n  }\n  let dot = 0;\n  for (let i = 0; i < a.length; i += 1) dot += a[i]! * b[i]!;\n  if (metric === 'dot') return dot;\n  // cosine: assume normalized vectors when produced by our backend\n  let na = 0;\n  let nb = 0;\n  for (let i = 0; i < a.length; i += 1) {\n    na += a[i]! * a[i]!;\n    nb += b[i]! * b[i]!;\n  }\n  const denom = Math.sqrt(na) * Math.sqrt(nb);\n  return denom === 0 ? 0 : dot / denom;\n}\n","/**\n * Hugging Face Inference API backend.\n *\n * Calls https://router.huggingface.co/hf-inference/models/<model>/pipeline/feature-extraction\n * with HF_TOKEN. For sentence-transformers models (BGE-M3, MiniLM, etc.) the\n * response is already mean-pooled per input — one vector per text. We\n * normalise client-side so cosine math is consistent across backends.\n */\n\nimport type { EmbeddingBackend, EmbeddingMatrix } from './types.js';\nimport type { EmbeddingMetric } from '@cvfile/sdk';\n\nexport interface HuggingFaceBackendOptions {\n  model: string;\n  /** HF token. Defaults to `process.env.HF_TOKEN`. */\n  token?: string;\n  /** Pinned revision; recorded in the payload. Default 'main'. */\n  modelRevision?: string;\n  /** Pre-known dimension. Optional; inferred from first response otherwise. */\n  dimension?: number;\n  metric?: EmbeddingMetric;\n  /** Override base URL (e.g. for self-hosted TEI). */\n  baseUrl?: string;\n}\n\nconst DEFAULT_BASE_URL = 'https://router.huggingface.co/hf-inference/models';\n\nexport function createHuggingFaceBackend(opts: HuggingFaceBackendOptions): EmbeddingBackend {\n  const token = opts.token ?? process.env.HF_TOKEN ?? process.env.HUGGINGFACE_TOKEN;\n  if (!token) {\n    throw new Error('HF_TOKEN (or HUGGINGFACE_TOKEN) is required for the Hugging Face backend');\n  }\n  const baseUrl = opts.baseUrl ?? DEFAULT_BASE_URL;\n  const url = `${baseUrl}/${encodeURI(opts.model)}/pipeline/feature-extraction`;\n\n  const backend: EmbeddingBackend = {\n    model: opts.model,\n    modelRevision: opts.modelRevision ?? 'main',\n    metric: opts.metric ?? 'cosine',\n    normalized: true,\n    async embed(texts: string[]): Promise<EmbeddingMatrix> {\n      if (texts.length === 0) {\n        return { vectors: [], dimension: opts.dimension ?? 0 };\n      }\n      const body = JSON.stringify({ inputs: texts, options: { wait_for_model: true } });\n      const res = await fetch(url, {\n        method: 'POST',\n        headers: {\n          'content-type': 'application/json',\n          authorization: `Bearer ${token}`,\n        },\n        body,\n      });\n      if (!res.ok) {\n        const detail = await res.text().catch(() => '');\n        throw new Error(`HF Inference API ${res.status} for ${opts.model}: ${detail || res.statusText}`);\n      }\n      const raw = (await res.json()) as unknown;\n      const matrix = parseHfResponse(raw, texts.length);\n      const dimension = matrix[0]?.length ?? opts.dimension ?? 0;\n      const vectors = matrix.map((v) => normalize(Float32Array.from(v)));\n      return { vectors, dimension };\n    },\n  };\n  if (opts.dimension !== undefined) backend.dimension = opts.dimension;\n  return backend;\n}\n\n/**\n * Coerce the variety of shapes the HF Inference API returns into a flat\n * `number[][]`: one mean-pooled vector per input.\n *\n * Observed shapes:\n *   - sentence-transformers (BGE-M3, MiniLM): `[[...vec], [...vec]]`\n *   - feature-extraction without pooling: `[[[...token0], [...token1], ...], ...]`\n *   - single-input convenience form: `[...vec]`\n */\nfunction parseHfResponse(raw: unknown, expectedCount: number): number[][] {\n  if (!Array.isArray(raw)) {\n    throw new Error('HF Inference API: expected array response');\n  }\n  if (raw.length === 0) return [];\n\n  const first = raw[0];\n  if (typeof first === 'number') {\n    if (expectedCount !== 1) {\n      throw new Error(`HF Inference API: got 1 vector, expected ${expectedCount}`);\n    }\n    return [raw as number[]];\n  }\n  if (Array.isArray(first) && (first.length === 0 || typeof first[0] === 'number')) {\n    return raw as number[][];\n  }\n  if (Array.isArray(first) && Array.isArray(first[0])) {\n    // token-level embeddings: mean-pool per input\n    return (raw as number[][][]).map(meanPool);\n  }\n  throw new Error('HF Inference API: unrecognised response shape');\n}\n\nfunction meanPool(tokens: number[][]): number[] {\n  if (tokens.length === 0) return [];\n  const dim = tokens[0]!.length;\n  const out = new Array<number>(dim).fill(0);\n  for (const t of tokens) {\n    for (let i = 0; i < dim; i += 1) out[i]! += t[i]!;\n  }\n  for (let i = 0; i < dim; i += 1) out[i]! /= tokens.length;\n  return out;\n}\n\nfunction normalize(v: Float32Array): Float32Array {\n  let sum = 0;\n  for (let i = 0; i < v.length; i += 1) sum += v[i]! * v[i]!;\n  const norm = Math.sqrt(sum);\n  if (norm === 0) return v;\n  for (let i = 0; i < v.length; i += 1) v[i] = v[i]! / norm;\n  return v;\n}\n"]}
+{"version":3,"sources":["../src/chunk.ts","../src/transformers-backend.ts","../src/types.ts","../src/embed.ts","../src/search.ts","../src/huggingface-backend.ts"],"names":[],"mappings":";;;AAgBA,IAAM,OAAA,GAAU,IAAI,WAAA,EAAY;AAChC,IAAM,OAAA,GAAU,IAAI,WAAA,EAAY;AAehC,IAAM,OAAA,GAAU,uBAAA;AAST,SAAS,aAAA,CAAc,QAAA,EAAkB,IAAA,GAAqB,EAAC,EAAoB;AACxF,EAAA,MAAM,IAAA,GAAO,KAAK,IAAA,IAAQ,SAAA;AAC1B,EAAA,MAAM,KAAA,GAAQ,OAAA,CAAQ,MAAA,CAAO,QAAQ,CAAA;AACrC,EAAA,IAAI,SAAS,UAAA,EAAY;AACvB,IAAA,OAAO,CAAC,aAAA,CAAc,KAAK,CAAC,CAAA;AAAA,EAC9B;AACA,EAAA,IAAI,SAAS,WAAA,EAAa;AACxB,IAAA,OAAO,gBAAgB,KAAK,CAAA;AAAA,EAC9B;AACA,EAAA,OAAO,cAAc,KAAK,CAAA;AAC5B;AAEA,SAAS,cAAc,KAAA,EAAkC;AACvD,EAAA,OAAO,EAAE,EAAA,EAAI,UAAA,EAAY,UAAA,EAAY,GAAG,UAAA,EAAY,KAAA,CAAM,UAAA,EAAY,IAAA,EAAM,SAAA,CAAU,KAAA,EAAO,CAAA,EAAG,KAAA,CAAM,UAAU,CAAA,EAAE;AACpH;AAEA,SAAS,cAAc,KAAA,EAAoC;AACzD,EAAA,MAAM,KAAA,GAAQ,qBAAqB,KAAK,CAAA;AACxC,EAAA,MAAM,WAA4B,EAAC;AACnC,EAAA,IAAI,OAAA,GAA6D,IAAA;AACjE,EAAA,MAAM,GAAA,uBAAU,GAAA,EAAY;AAE5B,EAAA,SAAS,MAAM,GAAA,EAAmB;AAChC,IAAA,IAAI,CAAC,OAAA,EAAS;AACd,IAAA,MAAM,IAAA,GAAO,SAAA,CAAU,KAAA,EAAO,OAAA,CAAQ,OAAO,GAAG,CAAA;AAChD,IAAA,IAAI,IAAA,CAAK,IAAA,EAAK,CAAE,MAAA,KAAW,CAAA,EAAG;AAC5B,MAAA,OAAA,GAAU,IAAA;AACV,MAAA;AAAA,IACF;AACA,IAAA,QAAA,CAAS,IAAA,CAAK,EAAE,EAAA,EAAI,OAAA,CAAQ,EAAA,EAAI,UAAA,EAAY,OAAA,CAAQ,KAAA,EAAO,UAAA,EAAY,GAAA,GAAM,OAAA,CAAQ,KAAA,EAAO,MAAM,CAAA;AAClG,IAAA,OAAA,GAAU,IAAA;AAAA,EACZ;AAEA,EAAA,KAAA,MAAW,QAAQ,KAAA,EAAO;AACxB,IAAA,MAAM,KAAA,GAAQ,OAAA,CAAQ,IAAA,CAAK,IAAA,CAAK,IAAI,CAAA;AACpC,IAAA,MAAM,OAAA,GAAU,IAAA,CAAK,MAAA,GAAS,IAAA,CAAK,UAAA;AACnC,IAAA,IAAI,KAAA,EAAO;AACT,MAAA,KAAA,CAAM,KAAK,MAAM,CAAA;AACjB,MAAA,MAAM,EAAA,GAAK,QAAA,CAAS,OAAA,CAAQ,KAAA,CAAM,CAAC,CAAA,IAAK,CAAA,QAAA,EAAW,QAAA,CAAS,MAAA,GAAS,CAAC,CAAA,CAAE,CAAA,EAAG,GAAG,CAAA;AAC9E,MAAA,GAAA,CAAI,IAAI,EAAE,CAAA;AACV,MAAA,OAAA,GAAU,EAAE,EAAA,EAAI,KAAA,EAAO,IAAA,CAAK,MAAA,EAAQ,KAAK,OAAA,EAAQ;AACjD,MAAA;AAAA,IACF;AACA,IAAA,IAAI,YAAY,IAAA,EAAM;AACpB,MAAA,MAAM,EAAA,GAAK,QAAA,CAAS,UAAA,EAAY,GAAG,CAAA;AACnC,MAAA,GAAA,CAAI,IAAI,EAAE,CAAA;AACV,MAAA,OAAA,GAAU,EAAE,EAAA,EAAI,KAAA,EAAO,IAAA,CAAK,MAAA,EAAQ,KAAK,OAAA,EAAQ;AAAA,IACnD,CAAA,MAAO;AACL,MAAA,OAAA,CAAQ,GAAA,GAAM,OAAA;AAAA,IAChB;AAAA,EACF;AACA,EAAA,KAAA,CAAM,MAAM,UAAU,CAAA;AAEtB,EAAA,IAAI,QAAA,CAAS,WAAW,CAAA,EAAG;AACzB,IAAA,OAAO,CAAC,aAAA,CAAc,KAAK,CAAC,CAAA;AAAA,EAC9B;AACA,EAAA,OAAO,QAAA;AACT;AAEA,SAAS,gBAAgB,KAAA,EAAoC;AAC3D,EAAA,MAAM,MAAuB,EAAC;AAC9B,EAAA,MAAM,GAAA,uBAAU,GAAA,EAAY;AAC5B,EAAA,MAAM,SAAA,GAAY,OAAA,CAAQ,MAAA,CAAO,MAAM,CAAA;AACvC,EAAA,IAAI,MAAA,GAAS,CAAA;AACb,EAAA,IAAI,CAAA,GAAI,CAAA;AACR,EAAA,OAAO,MAAA,GAAS,MAAM,UAAA,EAAY;AAChC,IAAA,IAAI,GAAA,GAAM,YAAA,CAAa,KAAA,EAAO,SAAA,EAAW,MAAM,CAAA;AAC/C,IAAA,IAAI,GAAA,KAAQ,EAAA,EAAI,GAAA,GAAM,KAAA,CAAM,UAAA;AAC5B,IAAA,MAAM,IAAA,GAAO,SAAA,CAAU,KAAA,EAAO,MAAA,EAAQ,GAAG,CAAA;AACzC,IAAA,IAAI,IAAA,CAAK,IAAA,EAAK,CAAE,MAAA,GAAS,CAAA,EAAG;AAC1B,MAAA,MAAM,EAAA,GAAK,QAAA,CAAS,OAAA,CAAQ,IAAA,CAAK,KAAA,CAAM,IAAI,CAAA,CAAE,CAAC,CAAA,IAAK,CAAA,EAAA,EAAK,CAAC,CAAA,CAAE,GAAG,GAAG,CAAA;AACjE,MAAA,GAAA,CAAI,IAAI,EAAE,CAAA;AACV,MAAA,GAAA,CAAI,IAAA,CAAK,EAAE,EAAA,EAAI,UAAA,EAAY,QAAQ,UAAA,EAAY,GAAA,GAAM,MAAA,EAAQ,IAAA,EAAM,CAAA;AACnE,MAAA,CAAA,IAAK,CAAA;AAAA,IACP;AACA,IAAA,MAAA,GAAS,MAAM,SAAA,CAAU,UAAA;AAAA,EAC3B;AACA,EAAA,IAAI,GAAA,CAAI,WAAW,CAAA,EAAG;AACpB,IAAA,OAAO,CAAC,aAAA,CAAc,KAAK,CAAC,CAAA;AAAA,EAC9B;AACA,EAAA,OAAO,GAAA;AACT;AAGA,SAAS,SAAA,CAAU,KAAA,EAAmB,KAAA,EAAe,GAAA,EAAqB;AACxE,EAAA,OAAO,QAAQ,MAAA,CAAO,KAAA,CAAM,QAAA,CAAS,KAAA,EAAO,GAAG,CAAC,CAAA;AAClD;AAGA,SAAS,qBAAqB,KAAA,EAA+B;AAC3D,EAAA,MAAM,OAAA,GAAU,EAAA;AAChB,EAAA,MAAM,QAAoB,EAAC;AAC3B,EAAA,IAAI,KAAA,GAAQ,CAAA;AACZ,EAAA,KAAA,IAAS,IAAI,CAAA,EAAG,CAAA,GAAI,KAAA,CAAM,UAAA,EAAY,KAAK,CAAA,EAAG;AAC5C,IAAA,IAAI,KAAA,CAAM,CAAC,CAAA,KAAM,OAAA,EAAS;AACxB,MAAA,MAAM,UAAA,GAAa,IAAI,KAAA,GAAQ,CAAA;AAC/B,MAAA,KAAA,CAAM,IAAA,CAAK,EAAE,IAAA,EAAM,SAAA,CAAU,KAAA,EAAO,KAAA,EAAO,CAAA,GAAI,CAAC,CAAA,EAAG,MAAA,EAAQ,KAAA,EAAO,UAAA,EAAY,CAAA;AAC9E,MAAA,KAAA,GAAQ,CAAA,GAAI,CAAA;AAAA,IACd;AAAA,EACF;AACA,EAAA,IAAI,KAAA,GAAQ,MAAM,UAAA,EAAY;AAC5B,IAAA,KAAA,CAAM,IAAA,CAAK,EAAE,IAAA,EAAM,SAAA,CAAU,OAAO,KAAA,EAAO,KAAA,CAAM,UAAU,CAAA,EAAG,QAAQ,KAAA,EAAO,UAAA,EAAY,KAAA,CAAM,UAAA,GAAa,OAAO,CAAA;AAAA,EACrH;AACA,EAAA,OAAO,KAAA;AACT;AAGA,SAAS,YAAA,CAAa,QAAA,EAAsB,MAAA,EAAoB,IAAA,EAAsB;AACpF,EAAA,MAAM,IAAA,GAAO,QAAA,CAAS,UAAA,GAAa,MAAA,CAAO,UAAA;AAC1C,EAAA,KAAA,IAAS,CAAA,GAAI,IAAA,EAAM,CAAA,IAAK,IAAA,EAAM,KAAK,CAAA,EAAG;AACpC,IAAA,IAAI,OAAA,GAAU,IAAA;AACd,IAAA,KAAA,IAAS,IAAI,CAAA,EAAG,CAAA,GAAI,MAAA,CAAO,UAAA,EAAY,KAAK,CAAA,EAAG;AAC7C,MAAA,IAAI,SAAS,CAAA,GAAI,CAAC,CAAA,KAAM,MAAA,CAAO,CAAC,CAAA,EAAG;AACjC,QAAA,OAAA,GAAU,KAAA;AACV,QAAA;AAAA,MACF;AAAA,IACF;AACA,IAAA,IAAI,SAAS,OAAO,CAAA;AAAA,EACtB;AACA,EAAA,OAAO,EAAA;AACT;AAEA,SAAS,QAAQ,CAAA,EAAmB;AAClC,EAAA,OACE,CAAA,CACG,WAAA,EAAY,CACZ,OAAA,CAAQ,eAAe,GAAG,CAAA,CAC1B,OAAA,CAAQ,UAAA,EAAY,EAAE,CAAA,CACtB,KAAA,CAAM,CAAA,EAAG,EAAE,CAAA,IAAK,SAAA;AAEvB;AAEA,SAAS,QAAA,CAAS,MAAc,KAAA,EAA4B;AAC1D,EAAA,IAAI,CAAC,KAAA,CAAM,GAAA,CAAI,IAAI,GAAG,OAAO,IAAA;AAC7B,EAAA,IAAI,CAAA,GAAI,CAAA;AACR,EAAA,OAAO,KAAA,CAAM,IAAI,CAAA,EAAG,IAAI,IAAI,CAAC,CAAA,CAAE,GAAG,CAAA,IAAK,CAAA;AACvC,EAAA,OAAO,CAAA,EAAG,IAAI,CAAA,CAAA,EAAI,CAAC,CAAA,CAAA;AACrB;AC/JO,SAAS,0BAA0B,IAAA,EAAoD;AAC5F,EAAA,IAAI,IAAA,CAAK,sBAAsB,KAAA,EAAO;AACpC,IAAA,GAAA,CAAI,iBAAA,GAAoB,KAAA;AAAA,EAC1B;AAEA,EAAA,IAAI,eAAA,GAA2C,IAAA;AAC/C,EAAA,IAAI,iBAAA,GAAmC,IAAA;AAEvC,EAAA,SAAS,WAAA,GAAgC;AACvC,IAAA,IAAI,CAAC,eAAA,EAAiB;AACpB,MAAA,MAAM,SAAA,GAAqC,EAAE,KAAA,EAAO,IAAA,CAAK,SAAS,MAAA,EAAO;AACzE,MAAA,IAAI,IAAA,CAAK,MAAA,KAAW,MAAA,EAAW,SAAA,CAAU,SAAS,IAAA,CAAK,MAAA;AACvD,MAAA,eAAA,GAAkB,QAAA,CAAS,oBAAA,EAAsB,IAAA,CAAK,KAAA,EAAO,SAAkB,CAAA;AAAA,IACjF;AACA,IAAA,OAAO,eAAA;AAAA,EACT;AAEA,EAAA,MAAM,OAAA,GAA4B;AAAA,IAChC,OAAO,IAAA,CAAK,KAAA;AAAA,IACZ,aAAA,EAAe,KAAK,aAAA,IAAiB,MAAA;AAAA,IACrC,MAAA,EAAQ,KAAK,MAAA,IAAU,QAAA;AAAA,IACvB,UAAA,EAAY,IAAA;AAAA,IACZ,MAAM,MAAM,KAAA,EAA2C;AACrD,MAAA,IAAI,KAAA,CAAM,WAAW,CAAA,EAAG;AACtB,QAAA,OAAO,EAAE,OAAA,EAAS,IAAI,SAAA,EAAW,IAAA,CAAK,aAAa,CAAA,EAAE;AAAA,MACvD;AACA,MAAA,MAAM,IAAA,GAAQ,MAAM,WAAA,EAAY;AAIhC,MAAA,MAAM,MAAA,GAAS,MAAM,IAAA,CAAK,KAAA,EAAO,EAAE,OAAA,EAAS,MAAA,EAAQ,SAAA,EAAW,IAAA,EAAM,CAAA;AACrE,MAAA,MAAM,OAAO,MAAA,CAAO,IAAA;AACpB,MAAA,MAAM,YAAY,MAAA,CAAO,IAAA,CAAK,MAAA,CAAO,IAAA,CAAK,SAAS,CAAC,CAAA;AACpD,MAAA,IAAI,iBAAA,KAAsB,IAAA,IAAQ,iBAAA,KAAsB,SAAA,EAAW;AACjE,QAAA,MAAM,IAAI,KAAA,CAAM,CAAA,0CAAA,EAA6C,iBAAiB,CAAA,MAAA,EAAS,SAAS,CAAA,CAAE,CAAA;AAAA,MACpG;AACA,MAAA,iBAAA,GAAoB,SAAA;AACpB,MAAA,MAAM,UAA0B,EAAC;AACjC,MAAA,KAAA,IAAS,IAAI,CAAA,EAAG,CAAA,GAAI,KAAA,CAAM,MAAA,EAAQ,KAAK,CAAA,EAAG;AACxC,QAAA,OAAA,CAAQ,IAAA,CAAK,IAAI,YAAA,CAAa,IAAA,CAAK,MAAA,EAAQ,IAAA,CAAK,UAAA,GAAa,CAAA,GAAI,SAAA,GAAY,CAAA,EAAG,SAAS,CAAA,CAAE,OAAO,CAAA;AAAA,MACpG;AACA,MAAA,OAAO,EAAE,SAAS,SAAA,EAAU;AAAA,IAC9B;AAAA,GACF;AACA,EAAA,IAAI,IAAA,CAAK,SAAA,KAAc,MAAA,EAAW,OAAA,CAAQ,YAAY,IAAA,CAAK,SAAA;AAC3D,EAAA,OAAO,OAAA;AACT;;;ACvCO,IAAM,aAAA,GAAgB;AACtB,IAAM,uBAAA,GAA0B;;;ACNvC,eAAsB,KAAA,CAAM,QAAA,EAAkB,IAAA,GAAqB,EAAC,EAA+B;AACjG,EAAA,MAAM,MAAA,GAAS,cAAc,QAAA,EAAU,EAAE,MAAM,IAAA,CAAK,QAAA,IAAY,WAAW,CAAA;AAC3E,EAAA,MAAM,OAAA,GACJ,KAAK,OAAA,IACL,yBAAA;AAAA,IACE,KAAK,aAAA,KAAkB,MAAA,GACnB,EAAE,KAAA,EAAO,KAAK,KAAA,IAAS,aAAA,EAAe,aAAA,EAAe,IAAA,CAAK,eAAc,GACxE,EAAE,KAAA,EAAO,IAAA,CAAK,SAAS,aAAA;AAAc,GAC3C;AAEF,EAAA,MAAM,MAAA,GAAS,MAAM,OAAA,CAAQ,KAAA,CAAM,MAAA,CAAO,IAAI,CAAC,CAAA,KAAM,CAAA,CAAE,IAAI,CAAC,CAAA;AAC5D,EAAA,IAAI,MAAA,CAAO,OAAA,CAAQ,MAAA,KAAW,MAAA,CAAO,MAAA,EAAQ;AAC3C,IAAA,MAAM,IAAI,MAAM,CAAA,iBAAA,EAAoB,MAAA,CAAO,QAAQ,MAAM,CAAA,aAAA,EAAgB,MAAA,CAAO,MAAM,CAAA,OAAA,CAAS,CAAA;AAAA,EACjG;AAEA,EAAA,MAAM,eAAA,GAAoC,MAAA,CAAO,GAAA,CAAI,CAAC,GAAG,CAAA,MAAO;AAAA,IAC9D,IAAI,CAAA,CAAE,EAAA;AAAA,IACN,YAAY,CAAA,CAAE,UAAA;AAAA,IACd,YAAY,CAAA,CAAE,UAAA;AAAA,IACd,MAAA,EAAQ,MAAA,CAAO,OAAA,CAAQ,CAAC;AAAA,GAC1B,CAAE,CAAA;AAEF,EAAA,MAAM,KAAA,GAAwB;AAAA,IAC5B,OAAO,OAAA,CAAQ,KAAA;AAAA,IACf,eAAe,OAAA,CAAQ,aAAA;AAAA,IACvB,WAAW,MAAA,CAAO,SAAA;AAAA,IAClB,QAAQ,OAAA,CAAQ,MAAA;AAAA,IAChB,YAAY,OAAA,CAAQ,UAAA;AAAA,IACpB,QAAA,EAAU,KAAK,QAAA,IAAY,SAAA;AAAA,IAC3B,MAAA,EAAQ;AAAA,GACV;AAEA,EAAA,OAAO,EAAE,aAAA,EAAe,CAAA,EAAG,MAAA,EAAQ,CAAC,KAAK,CAAA,EAAE;AAC7C;;;AC/BO,SAAS,cAAA,CACd,OAAA,EACA,WAAA,EACA,IAAA,GAAsB,EAAC,EACV;AACb,EAAA,MAAM,KAAA,GAAQ,SAAA,CAAU,OAAA,EAAS,IAAA,CAAK,KAAK,CAAA;AAC3C,EAAA,IAAI,CAAC,KAAA,EAAO,MAAM,IAAI,MAAM,mCAAmC,CAAA;AAC/D,EAAA,IAAI,WAAA,CAAY,MAAA,KAAW,KAAA,CAAM,SAAA,EAAW;AAC1C,IAAA,MAAM,IAAI,KAAA,CAAM,CAAA,uBAAA,EAA0B,WAAA,CAAY,MAAM,CAAA,sBAAA,EAAyB,KAAA,CAAM,KAAK,CAAA,EAAA,EAAK,KAAA,CAAM,SAAS,CAAA,CAAA,CAAG,CAAA;AAAA,EACzH;AAEA,EAAA,MAAM,CAAA,GAAI,KAAK,CAAA,IAAK,CAAA;AACpB,EAAA,MAAM,MAAA,GAAsB,KAAA,CAAM,MAAA,CAAO,GAAA,CAAI,CAAC,CAAA,MAAO;AAAA,IACnD,YAAY,KAAA,CAAM,KAAA;AAAA,IAClB,SAAS,CAAA,CAAE,EAAA;AAAA,IACX,YAAY,CAAA,CAAE,UAAA;AAAA,IACd,YAAY,CAAA,CAAE,UAAA;AAAA,IACd,OAAO,UAAA,CAAW,WAAA,EAAa,CAAA,CAAE,MAAA,EAAQ,MAAM,MAAM;AAAA,GACvD,CAAE,CAAA;AAGF,EAAA,MAAM,KAAA,GAAQ,KAAA,CAAM,MAAA,KAAW,WAAA,GAAc,CAAA,GAAI,EAAA;AACjD,EAAA,MAAA,CAAO,IAAA,CAAK,CAAC,CAAA,EAAG,CAAA,KAAM,SAAS,CAAA,CAAE,KAAA,GAAQ,EAAE,KAAA,CAAM,CAAA;AACjD,EAAA,OAAO,MAAA,CAAO,KAAA,CAAM,CAAA,EAAG,CAAC,CAAA;AAC1B;AAEA,SAAS,SAAA,CAAU,SAA4B,SAAA,EAAgD;AAC7F,EAAA,IAAI,SAAA,SAAkB,OAAA,CAAQ,MAAA,CAAO,KAAK,CAAC,CAAA,KAAM,CAAA,CAAE,KAAA,KAAU,SAAS,CAAA;AACtE,EAAA,OAAO,OAAA,CAAQ,OAAO,CAAC,CAAA;AACzB;AAEA,SAAS,UAAA,CAAW,CAAA,EAAiB,CAAA,EAAiB,MAAA,EAAiC;AACrF,EAAA,IAAI,WAAW,WAAA,EAAa;AAC1B,IAAA,IAAI,GAAA,GAAM,CAAA;AACV,IAAA,KAAA,IAAS,IAAI,CAAA,EAAG,CAAA,GAAI,CAAA,CAAE,MAAA,EAAQ,KAAK,CAAA,EAAG;AACpC,MAAA,MAAM,CAAA,GAAI,CAAA,CAAE,CAAC,CAAA,GAAK,EAAE,CAAC,CAAA;AACrB,MAAA,GAAA,IAAO,CAAA,GAAI,CAAA;AAAA,IACb;AACA,IAAA,OAAO,IAAA,CAAK,KAAK,GAAG,CAAA;AAAA,EACtB;AACA,EAAA,IAAI,GAAA,GAAM,CAAA;AACV,EAAA,KAAA,IAAS,CAAA,GAAI,CAAA,EAAG,CAAA,GAAI,CAAA,CAAE,MAAA,EAAQ,CAAA,IAAK,CAAA,EAAG,GAAA,IAAO,CAAA,CAAE,CAAC,CAAA,GAAK,CAAA,CAAE,CAAC,CAAA;AACxD,EAAA,IAAI,MAAA,KAAW,OAAO,OAAO,GAAA;AAE7B,EAAA,IAAI,EAAA,GAAK,CAAA;AACT,EAAA,IAAI,EAAA,GAAK,CAAA;AACT,EAAA,KAAA,IAAS,IAAI,CAAA,EAAG,CAAA,GAAI,CAAA,CAAE,MAAA,EAAQ,KAAK,CAAA,EAAG;AACpC,IAAA,EAAA,IAAM,CAAA,CAAE,CAAC,CAAA,GAAK,CAAA,CAAE,CAAC,CAAA;AACjB,IAAA,EAAA,IAAM,CAAA,CAAE,CAAC,CAAA,GAAK,CAAA,CAAE,CAAC,CAAA;AAAA,EACnB;AACA,EAAA,MAAM,QAAQ,IAAA,CAAK,IAAA,CAAK,EAAE,CAAA,GAAI,IAAA,CAAK,KAAK,EAAE,CAAA;AAC1C,EAAA,OAAO,KAAA,KAAU,CAAA,GAAI,CAAA,GAAI,GAAA,GAAM,KAAA;AACjC;;;AClDA,IAAM,gBAAA,GAAmB,mDAAA;AAElB,SAAS,yBAAyB,IAAA,EAAmD;AAC1F,EAAA,MAAM,QAAQ,IAAA,CAAK,KAAA,IAAS,QAAQ,GAAA,CAAI,QAAA,IAAY,QAAQ,GAAA,CAAI,iBAAA;AAChE,EAAA,IAAI,CAAC,KAAA,EAAO;AACV,IAAA,MAAM,IAAI,MAAM,0EAA0E,CAAA;AAAA,EAC5F;AACA,EAAA,MAAM,OAAA,GAAU,KAAK,OAAA,IAAW,gBAAA;AAChC,EAAA,MAAM,MAAM,CAAA,EAAG,OAAO,IAAI,SAAA,CAAU,IAAA,CAAK,KAAK,CAAC,CAAA,4BAAA,CAAA;AAE/C,EAAA,MAAM,OAAA,GAA4B;AAAA,IAChC,OAAO,IAAA,CAAK,KAAA;AAAA,IACZ,aAAA,EAAe,KAAK,aAAA,IAAiB,MAAA;AAAA,IACrC,MAAA,EAAQ,KAAK,MAAA,IAAU,QAAA;AAAA,IACvB,UAAA,EAAY,IAAA;AAAA,IACZ,MAAM,MAAM,KAAA,EAA2C;AACrD,MAAA,IAAI,KAAA,CAAM,WAAW,CAAA,EAAG;AACtB,QAAA,OAAO,EAAE,OAAA,EAAS,IAAI,SAAA,EAAW,IAAA,CAAK,aAAa,CAAA,EAAE;AAAA,MACvD;AACA,MAAA,MAAM,IAAA,GAAO,IAAA,CAAK,SAAA,CAAU,EAAE,MAAA,EAAQ,KAAA,EAAO,OAAA,EAAS,EAAE,cAAA,EAAgB,IAAA,EAAK,EAAG,CAAA;AAChF,MAAA,MAAM,GAAA,GAAM,MAAM,KAAA,CAAM,GAAA,EAAK;AAAA,QAC3B,MAAA,EAAQ,MAAA;AAAA,QACR,OAAA,EAAS;AAAA,UACP,cAAA,EAAgB,kBAAA;AAAA,UAChB,aAAA,EAAe,UAAU,KAAK,CAAA;AAAA,SAChC;AAAA,QACA;AAAA,OACD,CAAA;AACD,MAAA,IAAI,CAAC,IAAI,EAAA,EAAI;AACX,QAAA,MAAM,SAAS,MAAM,GAAA,CAAI,MAAK,CAAE,KAAA,CAAM,MAAM,EAAE,CAAA;AAC9C,QAAA,MAAM,IAAI,KAAA,CAAM,CAAA,iBAAA,EAAoB,GAAA,CAAI,MAAM,CAAA,KAAA,EAAQ,IAAA,CAAK,KAAK,CAAA,EAAA,EAAK,MAAA,IAAU,GAAA,CAAI,UAAU,CAAA,CAAE,CAAA;AAAA,MACjG;AACA,MAAA,MAAM,GAAA,GAAO,MAAM,GAAA,CAAI,IAAA,EAAK;AAC5B,MAAA,MAAM,MAAA,GAAS,eAAA,CAAgB,GAAA,EAAK,KAAA,CAAM,MAAM,CAAA;AAChD,MAAA,MAAM,YAAY,MAAA,CAAO,CAAC,CAAA,EAAG,MAAA,IAAU,KAAK,SAAA,IAAa,CAAA;AACzD,MAAA,MAAM,OAAA,GAAU,MAAA,CAAO,GAAA,CAAI,CAAC,CAAA,KAAM,UAAU,YAAA,CAAa,IAAA,CAAK,CAAC,CAAC,CAAC,CAAA;AACjE,MAAA,OAAO,EAAE,SAAS,SAAA,EAAU;AAAA,IAC9B;AAAA,GACF;AACA,EAAA,IAAI,IAAA,CAAK,SAAA,KAAc,MAAA,EAAW,OAAA,CAAQ,YAAY,IAAA,CAAK,SAAA;AAC3D,EAAA,OAAO,OAAA;AACT;AAWA,SAAS,eAAA,CAAgB,KAAc,aAAA,EAAmC;AACxE,EAAA,IAAI,CAAC,KAAA,CAAM,OAAA,CAAQ,GAAG,CAAA,EAAG;AACvB,IAAA,MAAM,IAAI,MAAM,2CAA2C,CAAA;AAAA,EAC7D;AACA,EAAA,IAAI,GAAA,CAAI,MAAA,KAAW,CAAA,EAAG,OAAO,EAAC;AAE9B,EAAA,MAAM,KAAA,GAAQ,IAAI,CAAC,CAAA;AACnB,EAAA,IAAI,OAAO,UAAU,QAAA,EAAU;AAC7B,IAAA,IAAI,kBAAkB,CAAA,EAAG;AACvB,MAAA,MAAM,IAAI,KAAA,CAAM,CAAA,yCAAA,EAA4C,aAAa,CAAA,CAAE,CAAA;AAAA,IAC7E;AACA,IAAA,OAAO,CAAC,GAAe,CAAA;AAAA,EACzB;AACA,EAAA,IAAI,KAAA,CAAM,OAAA,CAAQ,KAAK,CAAA,KAAM,KAAA,CAAM,MAAA,KAAW,CAAA,IAAK,OAAO,KAAA,CAAM,CAAC,CAAA,KAAM,QAAA,CAAA,EAAW;AAChF,IAAA,OAAO,GAAA;AAAA,EACT;AACA,EAAA,IAAI,KAAA,CAAM,QAAQ,KAAK,CAAA,IAAK,MAAM,OAAA,CAAQ,KAAA,CAAM,CAAC,CAAC,CAAA,EAAG;AAEnD,IAAA,OAAQ,GAAA,CAAqB,IAAI,QAAQ,CAAA;AAAA,EAC3C;AACA,EAAA,MAAM,IAAI,MAAM,+CAA+C,CAAA;AACjE;AAEA,SAAS,SAAS,MAAA,EAA8B;AAC9C,EAAA,IAAI,MAAA,CAAO,MAAA,KAAW,CAAA,EAAG,OAAO,EAAC;AACjC,EAAA,MAAM,GAAA,GAAM,MAAA,CAAO,CAAC,CAAA,CAAG,MAAA;AACvB,EAAA,MAAM,MAAM,IAAI,KAAA,CAAc,GAAG,CAAA,CAAE,KAAK,CAAC,CAAA;AACzC,EAAA,KAAA,MAAW,KAAK,MAAA,EAAQ;AACtB,IAAA,KAAA,IAAS,CAAA,GAAI,CAAA,EAAG,CAAA,GAAI,GAAA,EAAK,CAAA,IAAK,GAAG,GAAA,CAAI,CAAC,CAAA,IAAM,CAAA,CAAE,CAAC,CAAA;AAAA,EACjD;AACA,EAAA,KAAA,IAAS,CAAA,GAAI,GAAG,CAAA,GAAI,GAAA,EAAK,KAAK,CAAA,EAAG,GAAA,CAAI,CAAC,CAAA,IAAM,MAAA,CAAO,MAAA;AACnD,EAAA,OAAO,GAAA;AACT;AAEA,SAAS,UAAU,CAAA,EAA+B;AAChD,EAAA,IAAI,GAAA,GAAM,CAAA;AACV,EAAA,KAAA,IAAS,CAAA,GAAI,CAAA,EAAG,CAAA,GAAI,CAAA,CAAE,MAAA,EAAQ,CAAA,IAAK,CAAA,EAAG,GAAA,IAAO,CAAA,CAAE,CAAC,CAAA,GAAK,CAAA,CAAE,CAAC,CAAA;AACxD,EAAA,MAAM,IAAA,GAAO,IAAA,CAAK,IAAA,CAAK,GAAG,CAAA;AAC1B,EAAA,IAAI,IAAA,KAAS,GAAG,OAAO,CAAA;AACvB,EAAA,KAAA,IAAS,CAAA,GAAI,CAAA,EAAG,CAAA,GAAI,CAAA,CAAE,MAAA,EAAQ,CAAA,IAAK,CAAA,EAAG,CAAA,CAAE,CAAC,CAAA,GAAI,CAAA,CAAE,CAAC,CAAA,GAAK,IAAA;AACrD,EAAA,OAAO,CAAA;AACT","file":"index.js","sourcesContent":["/**\n * Section-based markdown chunker.\n *\n * Splits a markdown document on ATX headings (`#`, `##`, ...). Each chunk\n * carries the byte offset and length into the original UTF-8 source so a\n * downstream consumer can map a vector hit back to the exact substring\n * without re-tokenising. Pre-heading content becomes a \"preamble\" chunk.\n *\n * Per spec §5.1, `textOffset`/`textLength` are UTF-8 *byte* offsets into the\n * markdown source. We encode the document once with `TextEncoder`, track a\n * byte cursor while iterating lines (counting the trailing `\\n` byte), and\n * derive each chunk's `text` by decoding the corresponding byte slice. This\n * keeps the offsets in agreement with the Go and Python SDKs for any\n * non-ASCII résumé.\n */\n\nconst encoder = new TextEncoder();\nconst decoder = new TextDecoder();\n\nexport type ChunkingMode = 'document' | 'section' | 'paragraph';\n\nexport interface MarkdownChunk {\n  id: string;\n  textOffset: number;\n  textLength: number;\n  text: string;\n}\n\nexport interface ChunkOptions {\n  mode?: ChunkingMode;\n}\n\nconst HEADING = /^(#{1,6})\\s+(.+?)\\s*$/;\n\n/** A source line plus its UTF-8 byte offset and byte length (including any trailing `\\n`). */\ninterface ByteLine {\n  text: string;\n  offset: number;\n  byteLength: number;\n}\n\nexport function chunkMarkdown(markdown: string, opts: ChunkOptions = {}): MarkdownChunk[] {\n  const mode = opts.mode ?? 'section';\n  const bytes = encoder.encode(markdown);\n  if (mode === 'document') {\n    return [documentChunk(bytes)];\n  }\n  if (mode === 'paragraph') {\n    return paragraphChunks(bytes);\n  }\n  return sectionChunks(bytes);\n}\n\nfunction documentChunk(bytes: Uint8Array): MarkdownChunk {\n  return { id: 'document', textOffset: 0, textLength: bytes.byteLength, text: sliceText(bytes, 0, bytes.byteLength) };\n}\n\nfunction sectionChunks(bytes: Uint8Array): MarkdownChunk[] {\n  const lines = splitWithByteOffsets(bytes);\n  const sections: MarkdownChunk[] = [];\n  let current: { id: string; start: number; end: number } | null = null;\n  const ids = new Set<string>();\n\n  function flush(end: number): void {\n    if (!current) return;\n    const text = sliceText(bytes, current.start, end);\n    if (text.trim().length === 0) {\n      current = null;\n      return;\n    }\n    sections.push({ id: current.id, textOffset: current.start, textLength: end - current.start, text });\n    current = null;\n  }\n\n  for (const line of lines) {\n    const match = HEADING.exec(line.text);\n    const lineEnd = line.offset + line.byteLength;\n    if (match) {\n      flush(line.offset);\n      const id = uniqueId(slugify(match[2] ?? `section-${sections.length + 1}`), ids);\n      ids.add(id);\n      current = { id, start: line.offset, end: lineEnd };\n      continue;\n    }\n    if (current === null) {\n      const id = uniqueId('preamble', ids);\n      ids.add(id);\n      current = { id, start: line.offset, end: lineEnd };\n    } else {\n      current.end = lineEnd;\n    }\n  }\n  flush(bytes.byteLength);\n\n  if (sections.length === 0) {\n    return [documentChunk(bytes)];\n  }\n  return sections;\n}\n\nfunction paragraphChunks(bytes: Uint8Array): MarkdownChunk[] {\n  const out: MarkdownChunk[] = [];\n  const ids = new Set<string>();\n  const separator = encoder.encode('\\n\\n');\n  let cursor = 0;\n  let i = 0;\n  while (cursor < bytes.byteLength) {\n    let end = indexOfBytes(bytes, separator, cursor);\n    if (end === -1) end = bytes.byteLength;\n    const text = sliceText(bytes, cursor, end);\n    if (text.trim().length > 0) {\n      const id = uniqueId(slugify(text.split('\\n')[0] ?? `p-${i}`), ids);\n      ids.add(id);\n      out.push({ id, textOffset: cursor, textLength: end - cursor, text });\n      i += 1;\n    }\n    cursor = end + separator.byteLength;\n  }\n  if (out.length === 0) {\n    return [documentChunk(bytes)];\n  }\n  return out;\n}\n\n/** Decode the UTF-8 byte slice `[start, end)` back into a string. */\nfunction sliceText(bytes: Uint8Array, start: number, end: number): string {\n  return decoder.decode(bytes.subarray(start, end));\n}\n\n/** Split UTF-8 bytes into lines, each tagged with its byte offset and byte length (newline included). */\nfunction splitWithByteOffsets(bytes: Uint8Array): ByteLine[] {\n  const newline = 0x0a; // '\\n'\n  const lines: ByteLine[] = [];\n  let start = 0;\n  for (let i = 0; i < bytes.byteLength; i += 1) {\n    if (bytes[i] === newline) {\n      const byteLength = i - start + 1;\n      lines.push({ text: sliceText(bytes, start, i + 1), offset: start, byteLength });\n      start = i + 1;\n    }\n  }\n  if (start < bytes.byteLength) {\n    lines.push({ text: sliceText(bytes, start, bytes.byteLength), offset: start, byteLength: bytes.byteLength - start });\n  }\n  return lines;\n}\n\n/** Find the byte index of `needle` in `haystack` at or after `from`, or -1. */\nfunction indexOfBytes(haystack: Uint8Array, needle: Uint8Array, from: number): number {\n  const last = haystack.byteLength - needle.byteLength;\n  for (let i = from; i <= last; i += 1) {\n    let matched = true;\n    for (let j = 0; j < needle.byteLength; j += 1) {\n      if (haystack[i + j] !== needle[j]) {\n        matched = false;\n        break;\n      }\n    }\n    if (matched) return i;\n  }\n  return -1;\n}\n\nfunction slugify(s: string): string {\n  return (\n    s\n      .toLowerCase()\n      .replace(/[^a-z0-9]+/g, '-')\n      .replace(/^-+|-+$/g, '')\n      .slice(0, 64) || 'section'\n  );\n}\n\nfunction uniqueId(base: string, taken: Set<string>): string {\n  if (!taken.has(base)) return base;\n  let n = 2;\n  while (taken.has(`${base}-${n}`)) n += 1;\n  return `${base}-${n}`;\n}\n","/**\n * transformers.js backend (browser + Node + Bun via @huggingface/transformers).\n *\n * Loads the model lazily and reuses the pipeline across calls. Models are\n * cached on disk by transformers.js itself (HF cache layout).\n */\n\nimport { pipeline, env } from '@huggingface/transformers';\nimport type { EmbeddingBackend, EmbeddingBackendOptions, EmbeddingMatrix } from './types.js';\n\ninterface TransformersBackendOptions extends EmbeddingBackendOptions {\n  /** Override Xenova quantisation. Defaults to fp32 for closest fidelity. */\n  dtype?: 'fp32' | 'fp16' | 'q8' | 'q4';\n  /** Force a backend device (e.g. 'cpu', 'gpu', 'wasm'). Defaults to auto. */\n  device?: 'cpu' | 'gpu' | 'wasm' | 'webgpu';\n  /** Allow remote model downloads. true by default. */\n  allowRemoteModels?: boolean;\n}\n\nexport function createTransformersBackend(opts: TransformersBackendOptions): EmbeddingBackend {\n  if (opts.allowRemoteModels === false) {\n    env.allowRemoteModels = false;\n  }\n\n  let pipelinePromise: Promise<unknown> | null = null;\n  let resolvedDimension: number | null = null;\n\n  function getPipeline(): Promise<unknown> {\n    if (!pipelinePromise) {\n      const modelOpts: Record<string, unknown> = { dtype: opts.dtype ?? 'fp32' };\n      if (opts.device !== undefined) modelOpts.device = opts.device;\n      pipelinePromise = pipeline('feature-extraction', opts.model, modelOpts as never);\n    }\n    return pipelinePromise;\n  }\n\n  const backend: EmbeddingBackend = {\n    model: opts.model,\n    modelRevision: opts.modelRevision ?? 'main',\n    metric: opts.metric ?? 'cosine',\n    normalized: true,\n    async embed(texts: string[]): Promise<EmbeddingMatrix> {\n      if (texts.length === 0) {\n        return { vectors: [], dimension: opts.dimension ?? 0 };\n      }\n      const pipe = (await getPipeline()) as (\n        texts: string[],\n        opts: { pooling: 'mean'; normalize: boolean },\n      ) => Promise<{ data: Float32Array; dims: number[] }>;\n      const tensor = await pipe(texts, { pooling: 'mean', normalize: true });\n      const data = tensor.data;\n      const dimension = tensor.dims[tensor.dims.length - 1] as number;\n      if (resolvedDimension !== null && resolvedDimension !== dimension) {\n        throw new Error(`Model emitted inconsistent dimension: had ${resolvedDimension}, now ${dimension}`);\n      }\n      resolvedDimension = dimension;\n      const vectors: Float32Array[] = [];\n      for (let i = 0; i < texts.length; i += 1) {\n        vectors.push(new Float32Array(data.buffer, data.byteOffset + i * dimension * 4, dimension).slice());\n      }\n      return { vectors, dimension };\n    },\n  };\n  if (opts.dimension !== undefined) backend.dimension = opts.dimension;\n  return backend;\n}\n","import type { EmbeddingMetric } from '@cvfile/sdk';\n\nexport interface EmbeddingMatrix {\n  vectors: Float32Array[];\n  dimension: number;\n}\n\nexport interface EmbeddingBackend {\n  model: string;\n  modelRevision: string;\n  /** Optional pre-declared dimension; the backend may override after first call. */\n  dimension?: number;\n  metric: EmbeddingMetric;\n  normalized: boolean;\n  embed(texts: string[]): Promise<EmbeddingMatrix>;\n}\n\nexport interface EmbeddingBackendOptions {\n  model: string;\n  modelRevision?: string;\n  /** Optional pre-known dimension (e.g. 1024 for bge-m3). */\n  dimension?: number;\n  metric?: EmbeddingMetric;\n}\n\n/** Recommended default per spec §5: BAAI BGE-M3, MIT, multilingual, 1024-dim. */\nexport const DEFAULT_MODEL = 'Xenova/bge-m3';\nexport const DEFAULT_MODEL_DIMENSION = 1024;\n","/**\n * High-level embed() API: markdown in, EmbeddingsPayload out, ready to drop\n * into pack({ embeddings: ... }).\n */\n\nimport type { EmbeddingChunk, EmbeddingSpace, EmbeddingsPayload } from '@cvfile/sdk';\nimport { chunkMarkdown, type ChunkingMode } from './chunk.js';\nimport { createTransformersBackend } from './transformers-backend.js';\nimport { DEFAULT_MODEL, type EmbeddingBackend } from './types.js';\n\nexport interface EmbedOptions {\n  /** HF model id; defaults to BGE-M3 (Xenova/bge-m3). */\n  model?: string;\n  /** Pinned model revision; recorded in the payload for reproducibility. */\n  modelRevision?: string;\n  /** Chunking strategy; default 'section'. */\n  chunking?: ChunkingMode;\n  /** Bring-your-own backend (e.g. an OpenAI/Voyage adapter). */\n  backend?: EmbeddingBackend;\n}\n\nexport async function embed(markdown: string, opts: EmbedOptions = {}): Promise<EmbeddingsPayload> {\n  const chunks = chunkMarkdown(markdown, { mode: opts.chunking ?? 'section' });\n  const backend =\n    opts.backend ??\n    createTransformersBackend(\n      opts.modelRevision !== undefined\n        ? { model: opts.model ?? DEFAULT_MODEL, modelRevision: opts.modelRevision }\n        : { model: opts.model ?? DEFAULT_MODEL },\n    );\n\n  const matrix = await backend.embed(chunks.map((c) => c.text));\n  if (matrix.vectors.length !== chunks.length) {\n    throw new Error(`Backend returned ${matrix.vectors.length} vectors for ${chunks.length} chunks`);\n  }\n\n  const embeddingChunks: EmbeddingChunk[] = chunks.map((c, i) => ({\n    id: c.id,\n    textOffset: c.textOffset,\n    textLength: c.textLength,\n    vector: matrix.vectors[i]!,\n  }));\n\n  const space: EmbeddingSpace = {\n    model: backend.model,\n    modelRevision: backend.modelRevision,\n    dimension: matrix.dimension,\n    metric: backend.metric,\n    normalized: backend.normalized,\n    chunking: opts.chunking ?? 'section',\n    chunks: embeddingChunks,\n  };\n\n  return { formatVersion: 1, spaces: [space] };\n}\n","/**\n * Pure similarity search over an EmbeddingsPayload. Vector input is the\n * caller's responsibility (encode the query with the same model that was\n * used to populate the space, ideally pulled from space.model).\n */\n\nimport type { EmbeddingMetric, EmbeddingSpace, EmbeddingsPayload } from '@cvfile/sdk';\n\nexport interface SearchHit {\n  spaceModel: string;\n  chunkId: string;\n  textOffset: number;\n  textLength: number;\n  score: number;\n}\n\nexport interface SearchOptions {\n  /** Restrict to a specific embedding space; default = first space. */\n  model?: string;\n  /** Top-k results; default 5. */\n  k?: number;\n}\n\nexport function searchSemantic(\n  payload: EmbeddingsPayload,\n  queryVector: Float32Array,\n  opts: SearchOptions = {},\n): SearchHit[] {\n  const space = pickSpace(payload, opts.model);\n  if (!space) throw new Error('No matching embedding space found');\n  if (queryVector.length !== space.dimension) {\n    throw new Error(`Query vector dimension ${queryVector.length} does not match space ${space.model} (${space.dimension})`);\n  }\n\n  const k = opts.k ?? 5;\n  const scored: SearchHit[] = space.chunks.map((c) => ({\n    spaceModel: space.model,\n    chunkId: c.id,\n    textOffset: c.textOffset,\n    textLength: c.textLength,\n    score: similarity(queryVector, c.vector, space.metric),\n  }));\n\n  // Higher score = better for cosine/dot; lower for euclidean.\n  const order = space.metric === 'euclidean' ? 1 : -1;\n  scored.sort((a, b) => order * (a.score - b.score));\n  return scored.slice(0, k);\n}\n\nfunction pickSpace(payload: EmbeddingsPayload, modelHint?: string): EmbeddingSpace | undefined {\n  if (modelHint) return payload.spaces.find((s) => s.model === modelHint);\n  return payload.spaces[0];\n}\n\nfunction similarity(a: Float32Array, b: Float32Array, metric: EmbeddingMetric): number {\n  if (metric === 'euclidean') {\n    let sum = 0;\n    for (let i = 0; i < a.length; i += 1) {\n      const d = a[i]! - b[i]!;\n      sum += d * d;\n    }\n    return Math.sqrt(sum);\n  }\n  let dot = 0;\n  for (let i = 0; i < a.length; i += 1) dot += a[i]! * b[i]!;\n  if (metric === 'dot') return dot;\n  // cosine: assume normalized vectors when produced by our backend\n  let na = 0;\n  let nb = 0;\n  for (let i = 0; i < a.length; i += 1) {\n    na += a[i]! * a[i]!;\n    nb += b[i]! * b[i]!;\n  }\n  const denom = Math.sqrt(na) * Math.sqrt(nb);\n  return denom === 0 ? 0 : dot / denom;\n}\n","/**\n * Hugging Face Inference API backend.\n *\n * Calls https://router.huggingface.co/hf-inference/models/<model>/pipeline/feature-extraction\n * with HF_TOKEN. For sentence-transformers models (BGE-M3, MiniLM, etc.) the\n * response is already mean-pooled per input — one vector per text. We\n * normalise client-side so cosine math is consistent across backends.\n */\n\nimport type { EmbeddingBackend, EmbeddingMatrix } from './types.js';\nimport type { EmbeddingMetric } from '@cvfile/sdk';\n\nexport interface HuggingFaceBackendOptions {\n  model: string;\n  /** HF token. Defaults to `process.env.HF_TOKEN`. */\n  token?: string;\n  /** Pinned revision; recorded in the payload. Default 'main'. */\n  modelRevision?: string;\n  /** Pre-known dimension. Optional; inferred from first response otherwise. */\n  dimension?: number;\n  metric?: EmbeddingMetric;\n  /** Override base URL (e.g. for self-hosted TEI). */\n  baseUrl?: string;\n}\n\nconst DEFAULT_BASE_URL = 'https://router.huggingface.co/hf-inference/models';\n\nexport function createHuggingFaceBackend(opts: HuggingFaceBackendOptions): EmbeddingBackend {\n  const token = opts.token ?? process.env.HF_TOKEN ?? process.env.HUGGINGFACE_TOKEN;\n  if (!token) {\n    throw new Error('HF_TOKEN (or HUGGINGFACE_TOKEN) is required for the Hugging Face backend');\n  }\n  const baseUrl = opts.baseUrl ?? DEFAULT_BASE_URL;\n  const url = `${baseUrl}/${encodeURI(opts.model)}/pipeline/feature-extraction`;\n\n  const backend: EmbeddingBackend = {\n    model: opts.model,\n    modelRevision: opts.modelRevision ?? 'main',\n    metric: opts.metric ?? 'cosine',\n    normalized: true,\n    async embed(texts: string[]): Promise<EmbeddingMatrix> {\n      if (texts.length === 0) {\n        return { vectors: [], dimension: opts.dimension ?? 0 };\n      }\n      const body = JSON.stringify({ inputs: texts, options: { wait_for_model: true } });\n      const res = await fetch(url, {\n        method: 'POST',\n        headers: {\n          'content-type': 'application/json',\n          authorization: `Bearer ${token}`,\n        },\n        body,\n      });\n      if (!res.ok) {\n        const detail = await res.text().catch(() => '');\n        throw new Error(`HF Inference API ${res.status} for ${opts.model}: ${detail || res.statusText}`);\n      }\n      const raw = (await res.json()) as unknown;\n      const matrix = parseHfResponse(raw, texts.length);\n      const dimension = matrix[0]?.length ?? opts.dimension ?? 0;\n      const vectors = matrix.map((v) => normalize(Float32Array.from(v)));\n      return { vectors, dimension };\n    },\n  };\n  if (opts.dimension !== undefined) backend.dimension = opts.dimension;\n  return backend;\n}\n\n/**\n * Coerce the variety of shapes the HF Inference API returns into a flat\n * `number[][]`: one mean-pooled vector per input.\n *\n * Observed shapes:\n *   - sentence-transformers (BGE-M3, MiniLM): `[[...vec], [...vec]]`\n *   - feature-extraction without pooling: `[[[...token0], [...token1], ...], ...]`\n *   - single-input convenience form: `[...vec]`\n */\nfunction parseHfResponse(raw: unknown, expectedCount: number): number[][] {\n  if (!Array.isArray(raw)) {\n    throw new Error('HF Inference API: expected array response');\n  }\n  if (raw.length === 0) return [];\n\n  const first = raw[0];\n  if (typeof first === 'number') {\n    if (expectedCount !== 1) {\n      throw new Error(`HF Inference API: got 1 vector, expected ${expectedCount}`);\n    }\n    return [raw as number[]];\n  }\n  if (Array.isArray(first) && (first.length === 0 || typeof first[0] === 'number')) {\n    return raw as number[][];\n  }\n  if (Array.isArray(first) && Array.isArray(first[0])) {\n    // token-level embeddings: mean-pool per input\n    return (raw as number[][][]).map(meanPool);\n  }\n  throw new Error('HF Inference API: unrecognised response shape');\n}\n\nfunction meanPool(tokens: number[][]): number[] {\n  if (tokens.length === 0) return [];\n  const dim = tokens[0]!.length;\n  const out = new Array<number>(dim).fill(0);\n  for (const t of tokens) {\n    for (let i = 0; i < dim; i += 1) out[i]! += t[i]!;\n  }\n  for (let i = 0; i < dim; i += 1) out[i]! /= tokens.length;\n  return out;\n}\n\nfunction normalize(v: Float32Array): Float32Array {\n  let sum = 0;\n  for (let i = 0; i < v.length; i += 1) sum += v[i]! * v[i]!;\n  const norm = Math.sqrt(sum);\n  if (norm === 0) return v;\n  for (let i = 0; i < v.length; i += 1) v[i] = v[i]! / norm;\n  return v;\n}\n"]}

package/package.json

@@ -1,23 +1,31 @@
 {
   "name": "@cvfile/embed",
-  "version": "0.1.0",
+  "version": "0.3.1",
   "description": "Open-source embedding generation for the .cv format. BGE-M3 by default; pluggable.",
   "license": "Apache-2.0",
   "homepage": "https://cvfile.org",
   "repository": {
     "type": "git",
-    "url": "https://github.com/cvfile/cv",
+    "url": "git+https://github.com/cvfile/cv.git",
     "directory": "packages/embed-js"
   },
+  "publishConfig": {
+    "access": "public"
+  },
   "type": "module",
   "main": "./dist/index.cjs",
   "module": "./dist/index.js",
-  "types": "./dist/index.d.ts",
+  "types": "./dist/index.d.cts",
   "exports": {
     ".": {
-      "types": "./dist/index.d.ts",
-      "import": "./dist/index.js",
-      "require": "./dist/index.cjs"
+      "import": {
+        "types": "./dist/index.d.ts",
+        "default": "./dist/index.js"
+      },
+      "require": {
+        "types": "./dist/index.d.cts",
+        "default": "./dist/index.cjs"
+      }
     }
   },
   "files": [
@@ -28,7 +36,7 @@
   "sideEffects": false,
   "dependencies": {
     "@huggingface/transformers": "^3.0.0",
-    "@cvfile/sdk": "0.1.0"
+    "@cvfile/sdk": "^0.3.1"
   },
   "devDependencies": {
     "@types/node": "^22.0.0",
@@ -41,6 +49,7 @@
     "test": "vitest run",
     "test:watch": "vitest",
     "typecheck": "tsc --noEmit",
+    "lint": "eslint .",
     "clean": "rm -rf dist .turbo"
   }
 }