@absolutejs/absolute 0.19.0-beta.554 → 0.19.0-beta.556

package/dist/ai/index.js

@@ -2551,7 +2551,9 @@ var resolveRAGHybridSearchOptions = (retrieval) => {
       fusionConstant: DEFAULT_FUSION_CONSTANT,
       lexicalTopK: undefined,
       lexicalWeight: 2,
+      mmrLambda: 0.7,
       sourceBalanceStrategy: "cap",
+      diversityStrategy: "none",
       mode: "vector",
       vectorWeight: 1
     };
@@ -2562,7 +2564,9 @@ var resolveRAGHybridSearchOptions = (retrieval) => {
       fusionConstant: DEFAULT_FUSION_CONSTANT,
       lexicalTopK: undefined,
       lexicalWeight: 2,
+      mmrLambda: 0.7,
       sourceBalanceStrategy: "cap",
+      diversityStrategy: "none",
       mode: retrieval,
       vectorWeight: 1
     };
@@ -2573,7 +2577,9 @@ var resolveRAGHybridSearchOptions = (retrieval) => {
     lexicalTopK: retrieval.lexicalTopK,
     maxResultsPerSource: typeof retrieval.maxResultsPerSource === "number" ? Math.max(1, Math.floor(retrieval.maxResultsPerSource)) : undefined,
     lexicalWeight: Math.max(0, retrieval.lexicalWeight ?? 2),
+    mmrLambda: Math.min(1, Math.max(0, retrieval.mmrLambda ?? 0.7)),
     mode: retrieval.mode ?? "vector",
+    diversityStrategy: retrieval.diversityStrategy === "mmr" ? "mmr" : "none",
     sourceBalanceStrategy: retrieval.sourceBalanceStrategy === "round_robin" ? "round_robin" : "cap",
     vectorWeight: Math.max(0, retrieval.vectorWeight ?? 1)
   };
@@ -4376,6 +4382,47 @@ var applyRAGSourceDiversity = (results, maxResultsPerSource, strategy = "cap") =
   }
   return limited;
 };
+var dotProduct = (left, right) => {
+  const length = Math.min(left.length, right.length);
+  let total = 0;
+  for (let index = 0;index < length; index += 1) {
+    total += (left[index] ?? 0) * (right[index] ?? 0);
+  }
+  return total;
+};
+var applyRAGMMRDiversity = (results, queryVector, lambda) => {
+  const withEmbeddings = results.filter((result) => Array.isArray(result.embedding) && result.embedding.length > 0);
+  if (withEmbeddings.length < 2 || queryVector.length === 0) {
+    return results;
+  }
+  const selected = [];
+  const remaining = [...withEmbeddings];
+  while (remaining.length > 0) {
+    let bestIndex = 0;
+    let bestScore = Number.NEGATIVE_INFINITY;
+    for (let index = 0;index < remaining.length; index += 1) {
+      const candidate = remaining[index];
+      if (!candidate) {
+        continue;
+      }
+      const relevance = dotProduct(queryVector, candidate.embedding);
+      const redundancy = selected.length > 0 ? Math.max(...selected.map((entry) => dotProduct(entry.embedding, candidate.embedding))) : 0;
+      const mmrScore = lambda * relevance - (1 - lambda) * redundancy;
+      if (mmrScore > bestScore) {
+        bestScore = mmrScore;
+        bestIndex = index;
+      }
+    }
+    const next = remaining.splice(bestIndex, 1)[0];
+    if (!next) {
+      break;
+    }
+    selected.push(next);
+  }
+  const selectedIds = new Set(selected.map((entry) => entry.chunkId));
+  const tail = results.filter((entry) => !selectedIds.has(entry.chunkId));
+  return [...selected, ...tail];
+};
 var weightQueryResults = (results, queryIndex) => {
   if (queryIndex === 0) {
     return results;
@@ -4432,7 +4479,9 @@ var createRAGCollection = (options) => {
           hasQueryTransform,
           hasReranker,
           maxResultsPerSource: retrieval.maxResultsPerSource ?? null,
+          mmrLambda: retrieval.mmrLambda,
           mode: retrieval.mode,
+          diversityStrategy: retrieval.diversityStrategy,
           runLexical,
           runVector,
           sourceBalanceStrategy: retrieval.sourceBalanceStrategy,
@@ -4550,7 +4599,19 @@ var createRAGCollection = (options) => {
       },
       stage: "rerank"
     });
-    const diversified = applyRAGSourceDiversity(reranked, retrieval.maxResultsPerSource, retrieval.sourceBalanceStrategy);
+    const diversityAdjusted = retrieval.diversityStrategy === "mmr" ? applyRAGMMRDiversity(reranked, queryVector, retrieval.mmrLambda) : reranked;
+    if (retrieval.diversityStrategy === "mmr") {
+      steps.push({
+        count: diversityAdjusted.length,
+        label: "Applied MMR diversity reordering",
+        metadata: {
+          applied: queryVector.length > 0 && diversityAdjusted.some((entry) => Array.isArray(entry.embedding)),
+          mmrLambda: retrieval.mmrLambda
+        },
+        stage: "diversity"
+      });
+    }
+    const diversified = applyRAGSourceDiversity(diversityAdjusted, retrieval.maxResultsPerSource, retrieval.sourceBalanceStrategy);
     if (typeof retrieval.maxResultsPerSource === "number") {
       steps.push({
         count: diversified.length,
@@ -4578,7 +4639,9 @@ var createRAGCollection = (options) => {
           candidateTopK,
           lexicalTopK,
           maxResultsPerSource: retrieval.maxResultsPerSource,
+          mmrLambda: retrieval.mmrLambda,
           mode: retrieval.mode,
+          diversityStrategy: retrieval.diversityStrategy,
           query: input.query,
           resultCounts: {
             final: limited.length,
@@ -4621,7 +4684,9 @@ var createRAGCollection = (options) => {
         candidateTopK,
         lexicalTopK,
         maxResultsPerSource: retrieval.maxResultsPerSource,
+        mmrLambda: retrieval.mmrLambda,
         mode: retrieval.mode,
+        diversityStrategy: retrieval.diversityStrategy,
         query: input.query,
         resultCounts: {
           final: filtered.length,
@@ -10704,6 +10769,7 @@ var createInMemoryRAGStore = (options = {}) => {
     return results.slice(0, input.topK).map((entry) => ({
       chunkId: entry.chunk.chunkId,
       chunkText: entry.chunk.text,
+      embedding: entry.chunk.vector,
       metadata: entry.chunk.metadata,
       score: entry.score,
       source: entry.chunk.source,
@@ -10856,7 +10922,7 @@ var IDENTIFIER_RE = /^[a-zA-Z_][a-zA-Z0-9_]*$/;
 var isParsedMetadata = (value) => Boolean(value) && typeof value === "object";
 var isObjectRecord2 = (value) => Boolean(value) && typeof value === "object";
 var isStoredRow = (value) => isObjectRecord2(value) && typeof value.chunk_id === "string" && typeof value.text === "string" && (typeof value.title === "string" || value.title === null) && (typeof value.source === "string" || value.source === null) && (typeof value.metadata === "string" || value.metadata === null) && typeof value.embedding === "string";
-var isNativeStoredRow = (value) => isObjectRecord2(value) && typeof value.chunk_id === "string" && typeof value.chunk_text === "string" && (typeof value.title === "string" || value.title === null) && (typeof value.source === "string" || value.source === null) && (typeof value.metadata === "string" || value.metadata === null) && typeof value.distance === "number";
+var isNativeStoredRow = (value) => isObjectRecord2(value) && typeof value.chunk_id === "string" && typeof value.chunk_text === "string" && (typeof value.title === "string" || value.title === null) && (typeof value.source === "string" || value.source === null) && (typeof value.metadata === "string" || value.metadata === null) && typeof value.embedding === "string" && typeof value.distance === "number";
 var toStoredRows = (value) => Array.isArray(value) ? value.filter((row) => isStoredRow(row)) : [];
 var toNativeStoredRows = (value) => Array.isArray(value) ? value.filter((row) => isNativeStoredRow(row)) : [];
 var createSQLiteStatus = (dimensions, nativeDiagnostics, useNative) => ({
@@ -11020,6 +11086,7 @@ var createNativeVec0Statements = (db, tableName) => {
   const querySql = `
 		SELECT
 			chunk_id,
+			embedding,
 			chunk_text,
 			title,
 			source,
@@ -11297,6 +11364,7 @@ var createSQLiteRAGStore = (options = {}) => {
     return filtered.slice(0, input.topK).map(({ chunk, score }) => ({
       chunkId: chunk.chunkId,
       chunkText: chunk.text,
+      embedding: chunk.vector,
       metadata: chunk.metadata,
       score,
       source: chunk.source,
@@ -11318,7 +11386,7 @@ var createSQLiteRAGStore = (options = {}) => {
         source: row.source ?? undefined,
         text: row.chunk_text,
         title: row.title ?? undefined,
-        vector: []
+        vector: parseVector(row.embedding)
       };
       return {
         chunk,
@@ -11327,6 +11395,7 @@ var createSQLiteRAGStore = (options = {}) => {
     }).filter(({ chunk }) => matchesFilter(chunk, input.filter)).map((entry) => ({
       chunkId: entry.chunk.chunkId,
       chunkText: entry.chunk.text,
+      embedding: entry.chunk.vector,
       metadata: entry.chunk.metadata,
       score: entry.score,
       source: entry.chunk.source,
@@ -12644,5 +12713,5 @@ export {
   aiChat
 };
 
-//# debugId=952E4DA5A5D3C13F64756E2164756E21
+//# debugId=9DF63A32267D961C64756E2164756E21
 //# sourceMappingURL=index.js.map