@retrivora-ai/rag-engine 1.9.2 → 1.9.3

package/dist/handlers/index.js

@@ -1090,12 +1090,15 @@ var init_MongoDBProvider = __esm({
           }
         );
         const results = await this.collection.aggregate(pipeline).toArray();
-        return results.map((res) => ({
-          id: res._id,
-          content: res[this.contentKey],
-          metadata: res[this.metadataKey] || {},
-          score: res.score
-        }));
+        return results.map((res) => {
+          const normalizedScore = res.score * 2 - 1;
+          return {
+            id: res._id,
+            content: res[this.contentKey],
+            metadata: res[this.metadataKey] || {},
+            score: normalizedScore
+          };
+        });
       }
       async delete(id, namespace) {
         await this.collection.deleteOne(__spreadValues({ _id: id }, namespace ? { namespace } : {}));
@@ -6387,7 +6390,8 @@ ${m.content}`).join("\n\n---\n\n") : "No relevant context found.";
         if (hasMetadataFilter) {
           displayCount = fullSources.length;
         } else {
-          const highlyRelevant = fullSources.filter((m) => m.score >= 0.4);
+          const baseThreshold = Math.max(scoreThreshold, 0.4);
+          const highlyRelevant = fullSources.filter((m) => m.score >= baseThreshold);
           displayCount = Math.max(highlyRelevant.length, topK);
           if (displayCount > 15) {
             displayCount = 15;

package/dist/handlers/index.mjs

@@ -1069,12 +1069,15 @@ var init_MongoDBProvider = __esm({
           }
         );
         const results = await this.collection.aggregate(pipeline).toArray();
-        return results.map((res) => ({
-          id: res._id,
-          content: res[this.contentKey],
-          metadata: res[this.metadataKey] || {},
-          score: res.score
-        }));
+        return results.map((res) => {
+          const normalizedScore = res.score * 2 - 1;
+          return {
+            id: res._id,
+            content: res[this.contentKey],
+            metadata: res[this.metadataKey] || {},
+            score: normalizedScore
+          };
+        });
       }
       async delete(id, namespace) {
         await this.collection.deleteOne(__spreadValues({ _id: id }, namespace ? { namespace } : {}));
@@ -6350,7 +6353,8 @@ ${m.content}`).join("\n\n---\n\n") : "No relevant context found.";
         if (hasMetadataFilter) {
           displayCount = fullSources.length;
         } else {
-          const highlyRelevant = fullSources.filter((m) => m.score >= 0.4);
+          const baseThreshold = Math.max(scoreThreshold, 0.4);
+          const highlyRelevant = fullSources.filter((m) => m.score >= baseThreshold);
           displayCount = Math.max(highlyRelevant.length, topK);
           if (displayCount > 15) {
             displayCount = 15;

package/dist/server.js

@@ -1090,12 +1090,15 @@ var init_MongoDBProvider = __esm({
           }
         );
         const results = await this.collection.aggregate(pipeline).toArray();
-        return results.map((res) => ({
-          id: res._id,
-          content: res[this.contentKey],
-          metadata: res[this.metadataKey] || {},
-          score: res.score
-        }));
+        return results.map((res) => {
+          const normalizedScore = res.score * 2 - 1;
+          return {
+            id: res._id,
+            content: res[this.contentKey],
+            metadata: res[this.metadataKey] || {},
+            score: normalizedScore
+          };
+        });
       }
       async delete(id, namespace) {
         await this.collection.deleteOne(__spreadValues({ _id: id }, namespace ? { namespace } : {}));
@@ -6464,7 +6467,8 @@ ${m.content}`).join("\n\n---\n\n") : "No relevant context found.";
         if (hasMetadataFilter) {
           displayCount = fullSources.length;
         } else {
-          const highlyRelevant = fullSources.filter((m) => m.score >= 0.4);
+          const baseThreshold = Math.max(scoreThreshold, 0.4);
+          const highlyRelevant = fullSources.filter((m) => m.score >= baseThreshold);
           displayCount = Math.max(highlyRelevant.length, topK);
           if (displayCount > 15) {
             displayCount = 15;

package/dist/server.mjs

@@ -1069,12 +1069,15 @@ var init_MongoDBProvider = __esm({
           }
         );
         const results = await this.collection.aggregate(pipeline).toArray();
-        return results.map((res) => ({
-          id: res._id,
-          content: res[this.contentKey],
-          metadata: res[this.metadataKey] || {},
-          score: res.score
-        }));
+        return results.map((res) => {
+          const normalizedScore = res.score * 2 - 1;
+          return {
+            id: res._id,
+            content: res[this.contentKey],
+            metadata: res[this.metadataKey] || {},
+            score: normalizedScore
+          };
+        });
       }
       async delete(id, namespace) {
         await this.collection.deleteOne(__spreadValues({ _id: id }, namespace ? { namespace } : {}));
@@ -6395,7 +6398,8 @@ ${m.content}`).join("\n\n---\n\n") : "No relevant context found.";
         if (hasMetadataFilter) {
           displayCount = fullSources.length;
         } else {
-          const highlyRelevant = fullSources.filter((m) => m.score >= 0.4);
+          const baseThreshold = Math.max(scoreThreshold, 0.4);
+          const highlyRelevant = fullSources.filter((m) => m.score >= baseThreshold);
           displayCount = Math.max(highlyRelevant.length, topK);
           if (displayCount > 15) {
             displayCount = 15;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@retrivora-ai/rag-engine",
-  "version": "1.9.2",
+  "version": "1.9.3",
   "description": "Retrivora AI is a plug-and-play AI engine for RAG chat experiences — generic vector DB + LLM provider, embeddable or standalone.",
   "author": "Abhinav Alkuchi",
   "license": "MIT",

package/src/core/Pipeline.ts

@@ -525,7 +525,8 @@ You are a helpful assistant. Use the provided context to answer questions accura
         displayCount = fullSources.length;
       } else {
         // For general semantic queries, count the number of matches meeting high relevance threshold
-        const highlyRelevant = fullSources.filter(m => m.score >= 0.4);
+        const baseThreshold = Math.max(scoreThreshold, 0.4);
+        const highlyRelevant = fullSources.filter(m => m.score >= baseThreshold);
         displayCount = Math.max(highlyRelevant.length, topK);
         // Cap non-metadata semantic searches to a maximum of 15 to keep the UI clean
         if (displayCount > 15) {

package/src/providers/vectordb/MongoDBProvider.ts

@@ -191,12 +191,19 @@ export class MongoDBProvider extends BaseVectorProvider {
 
     const results = await this.collection!.aggregate(pipeline).toArray();
 
-    return (results as unknown as Array<{ _id: string; score: number; [key: string]: unknown }>).map((res) => ({
-      id: res._id,
-      content: res[this.contentKey] as string,
-      metadata: (res[this.metadataKey] as Record<string, unknown>) || {},
-      score: res.score,
-    }));
+    return (results as unknown as Array<{ _id: string; score: number; [key: string]: unknown }>).map((res) => {
+      // Normalize MongoDB's vectorSearchScore (which is shifted to [0, 1] e.g. (1+cos)/2)
+      // back to the standard [-1, 1] cosine space. This ensures the global relevance 
+      // threshold (e.g. 0.4) properly filters out random orthogonal matches.
+      const normalizedScore = (res.score * 2) - 1;
+
+      return {
+        id: res._id,
+        content: res[this.contentKey] as string,
+        metadata: (res[this.metadataKey] as Record<string, unknown>) || {},
+        score: normalizedScore,
+      };
+    });
   }
 
   async delete(id: string | number, namespace?: string): Promise<void> {