@absolutejs/absolute 0.19.0-beta.491 → 0.19.0-beta.493

package/dist/ai/index.js

@@ -2161,10 +2161,66 @@ var collectMetadataStrings = (value) => {
   }
   return [];
 };
+var normalizeSourceForLexical = (source) => source.replace(/[#/_.-]+/g, " ").replace(/\bmd\b/g, "markdown").replace(/\bpptx\b/g, "presentation").replace(/\bxlsx\b/g, "spreadsheet workbook sheet").replace(/\bmp3\b/g, "audio transcript media").replace(/\bmp4\b/g, "video transcript media").replace(/\bzip\b/g, "archive bundle");
+var toFieldText = (value) => collectMetadataStrings(value).filter(Boolean).join(" ");
+var scoreTokenCoverage = (queryTokens, text) => {
+  const normalizedText = text.toLowerCase();
+  if (normalizedText.length === 0) {
+    return 0;
+  }
+  const tokens = tokenize(normalizedText);
+  if (tokens.length === 0) {
+    return 0;
+  }
+  const tokenSet = new Set(tokens);
+  const overlap = queryTokens.filter((token) => tokenSet.has(token)).length;
+  return overlap / Math.max(1, queryTokens.length);
+};
+var scorePhraseMatch = (query, text) => {
+  const normalizedQuery = tokenize(query).join(" ");
+  const normalizedText = tokenize(text).join(" ");
+  if (normalizedQuery.length === 0 || normalizedText.length === 0) {
+    return 0;
+  }
+  return normalizedText.includes(normalizedQuery) ? 1 : 0;
+};
+var scoreWeightedField = ({
+  coverageWeight,
+  phraseWeight,
+  query,
+  queryTokens,
+  text
+}) => scoreTokenCoverage(queryTokens, text) * coverageWeight + scorePhraseMatch(query, text) * phraseWeight;
+var extractWeightedLexicalFields = (result) => {
+  const metadata = result.metadata ?? {};
+  const source = result.source ?? "";
+  const archivePath = typeof metadata.archivePath === "string" ? metadata.archivePath : source.includes("#") ? source.split("#")[1] ?? "" : "";
+  const mediaSegments = Array.isArray(metadata.mediaSegments) ? metadata.mediaSegments.map((segment) => segment && typeof segment === "object" ? toFieldText(segment) : "").filter(Boolean).join(" ") : "";
+  const metadataFocus = [
+    metadata.sheetName,
+    metadata.sheetNames,
+    metadata.slideTitle,
+    metadata.slideTitles,
+    metadata.threadTopic,
+    metadata.speaker,
+    metadata.fileKind,
+    metadata.transcriptSource,
+    metadata.archiveType
+  ].flatMap((value) => collectMetadataStrings(value)).join(" ");
+  return {
+    archivePath,
+    chunkText: result.chunkText,
+    mediaSegments,
+    metadataFocus,
+    metadataText: toFieldText(metadata),
+    source: source ? normalizeSourceForLexical(source) : "",
+    title: result.title ?? ""
+  };
+};
 var buildRAGLexicalHaystack = (result) => [
   result.title,
   result.source,
-  typeof result.source === "string" ? result.source.replace(/[#/_.-]+/g, " ").replace(/\bmd\b/g, "markdown").replace(/\bpptx\b/g, "presentation").replace(/\bxlsx\b/g, "spreadsheet workbook sheet").replace(/\bmp3\b/g, "audio transcript media").replace(/\bmp4\b/g, "video transcript media").replace(/\bzip\b/g, "archive bundle") : undefined,
+  typeof result.source === "string" ? normalizeSourceForLexical(result.source) : undefined,
   result.chunkText,
   ...collectMetadataStrings(result.metadata)
 ].filter((value) => Boolean(value)).join(" ");
@@ -2173,20 +2229,67 @@ var scoreRAGLexicalMatch = (query, result) => {
   if (queryTokens.length === 0) {
     return 0;
   }
+  const fields = extractWeightedLexicalFields(result);
   const haystack = buildRAGLexicalHaystack(result).toLowerCase();
-  const haystackTokens = tokenize(haystack);
-  const haystackSet = new Set(haystackTokens);
-  const overlap = queryTokens.filter((token) => haystackSet.has(token)).length;
-  if (overlap === 0) {
+  const overallCoverage = scoreTokenCoverage(queryTokens, haystack);
+  if (overallCoverage === 0) {
     return 0;
   }
-  const exactPhraseBoost = haystack.includes(query.toLowerCase()) ? 1 : 0;
-  const sourceBoost = typeof result.source === "string" && queryTokens.some((token) => result.source?.toLowerCase().includes(token)) ? 0.5 : 0;
-  const coverageBoost = overlap / queryTokens.length;
+  const titleScore = scoreWeightedField({
+    coverageWeight: 1.8,
+    phraseWeight: 1.2,
+    query,
+    queryTokens,
+    text: fields.title
+  });
+  const sourceScore = scoreWeightedField({
+    coverageWeight: 2.6,
+    phraseWeight: 1.4,
+    query,
+    queryTokens,
+    text: fields.source
+  });
+  const metadataFocusScore = scoreWeightedField({
+    coverageWeight: 2.8,
+    phraseWeight: 1.6,
+    query,
+    queryTokens,
+    text: fields.metadataFocus
+  });
+  const archivePathScore = scoreWeightedField({
+    coverageWeight: 3.2,
+    phraseWeight: 2.2,
+    query,
+    queryTokens,
+    text: fields.archivePath
+  });
+  const mediaSegmentScore = scoreWeightedField({
+    coverageWeight: 3,
+    phraseWeight: 1.8,
+    query,
+    queryTokens,
+    text: fields.mediaSegments
+  });
+  const metadataScore = scoreWeightedField({
+    coverageWeight: 1.2,
+    phraseWeight: 0.8,
+    query,
+    queryTokens,
+    text: fields.metadataText
+  });
+  const chunkScore = scoreWeightedField({
+    coverageWeight: 0.9,
+    phraseWeight: 0.6,
+    query,
+    queryTokens,
+    text: fields.chunkText
+  });
+  const exactPhraseBoost = scorePhraseMatch(query, haystack);
+  const coverageBoost = overallCoverage;
   const fileKindBoost = resolveFileKindBoost(queryTokens, result.metadata);
   const transcriptBoost = resolveTranscriptBoost(queryTokens, result.metadata);
   const archiveBoost = resolveArchiveBoost(queryTokens, result);
-  return coverageBoost + exactPhraseBoost + sourceBoost + fileKindBoost + transcriptBoost + archiveBoost;
+  return titleScore + sourceScore + metadataFocusScore + archivePathScore + mediaSegmentScore + metadataScore + chunkScore + coverageBoost + exactPhraseBoost + fileKindBoost + transcriptBoost + archiveBoost;
 };
 var hasAnyToken = (tokens, values) => values.some((value) => tokens.includes(value));
 var resolveFileKindBoost = (queryTokens, metadata) => {
@@ -3877,8 +3980,10 @@ var createRAGCollection = (options) => {
     const model = input.model ?? options.defaultModel;
     const topK = input.topK ?? defaultTopK;
     const hasReranker = Boolean(input.rerank ?? options.rerank);
-    const candidateTopK = Math.max(topK, Math.floor(input.candidateTopK ?? (hasReranker ? topK * defaultCandidateMultiplier : topK)));
     const retrieval = resolveRAGHybridSearchOptions(input.retrieval);
+    const hasQueryTransform = Boolean(input.queryTransform ?? options.queryTransform);
+    const shouldExpandCandidates = hasReranker || hasQueryTransform || retrieval.mode !== "vector";
+    const candidateTopK = Math.max(topK, Math.floor(input.candidateTopK ?? (shouldExpandCandidates ? topK * defaultCandidateMultiplier : topK)));
     const transformed = await applyRAGQueryTransform({
       input: {
         candidateTopK,
@@ -8598,5 +8703,5 @@ export {
   aiChat
 };
 
-//# debugId=F84CC012821C275C64756E2164756E21
+//# debugId=36F6407CE8163A4F64756E2164756E21
 //# sourceMappingURL=index.js.map