rag-memory-epf-mcp 3.1.4 → 3.2.0

package/dist/index.js

@@ -83,18 +83,10 @@ class RAGKnowledgeGraphManager {
             console.error(`✅ ${EMBEDDING_MODEL} model loaded successfully (fp16)`);
         }
         catch (error) {
-            console.error('❌ Failed to load embedding model (fp16):', error instanceof Error ? error.message : error);
-            const FALLBACK_MODEL = 'Xenova/all-MiniLM-L6-v2';
-            console.error(`🔄 Falling back to ${FALLBACK_MODEL}...`);
-            try {
-                this.embeddingModel = await pipeline('feature-extraction', FALLBACK_MODEL, { revision: 'main' });
-                this.modelInitialized = true;
-                console.error(`✅ ${FALLBACK_MODEL} loaded (fallback — 384-dim, English-focused)`);
-            }
-            catch (retryError) {
-                console.error('❌ All embedding models failed to load. Search will not work.', retryError instanceof Error ? retryError.message : retryError);
-                this.modelInitialized = false;
-            }
+            console.error('❌ Failed to load embedding model:', error instanceof Error ? error.message : error);
+            console.error('⚠️  Semantic search will be unavailable. Other tools (CRUD, graph queries) still work.');
+            console.error('💡 Fix: Ensure ONNX model cache is accessible. Try: rm -rf ~/.npm/_onnx_models && restart.');
+            this.modelInitialized = false;
         }
     }
     async runMigrations() {
@@ -1353,11 +1345,14 @@ class RAGKnowledgeGraphManager {
         if (!document) {
             throw new Error(`Document with ID ${documentId} not found`);
         }
-        // Get chunks for this document
+        // Get chunks for this document (with text for chunk-level matching)
         const chunks = this.db.prepare(`
-      SELECT rowid FROM chunk_metadata WHERE document_id = ?
+      SELECT rowid, text FROM chunk_metadata WHERE document_id = ?
     `).all(documentId);
         let linkedCount = 0;
+        const insertStmt = this.db.prepare(`
+      INSERT OR IGNORE INTO chunk_entities (chunk_rowid, entity_id) VALUES (?, ?)
+    `);
         for (const entityName of entityNames) {
             const entityId = `entity_${entityName.toLowerCase().replace(/[^a-z0-9]/g, '_')}`;
             // Verify entity exists
@@ -1368,14 +1363,17 @@ class RAGKnowledgeGraphManager {
                 console.warn(`Entity ${entityName} not found, skipping`);
                 continue;
             }
-            // Link entity to all chunks of the document
+            // Chunk-level filtering: only link to chunks where entity actually appears
+            const nameMatcher = this.buildEntityMatcher(entityName);
+            let entityLinked = false;
             for (const chunk of chunks) {
-                this.db.prepare(`
-          INSERT OR IGNORE INTO chunk_entities (chunk_rowid, entity_id)
-          VALUES (?, ?)
-        `).run(chunk.rowid, entityId);
+                if (nameMatcher(chunk.text)) {
+                    insertStmt.run(chunk.rowid, entityId);
+                    entityLinked = true;
+                }
             }
-            linkedCount++;
+            if (entityLinked)
+                linkedCount++;
         }
         console.error(`✅ Entities linked: ${linkedCount} entities linked to document`);
         return { documentId, linkedEntities: linkedCount };
@@ -1876,42 +1874,54 @@ class RAGKnowledgeGraphManager {
                     chunkEntities = [relEntities.source_name, relEntities.target_name];
                 }
             }
-            // Enhanced graph boost calculation
+            // Enhanced graph boost calculation with decay + cap
             let graphBoost = 0;
             if (useGraph) {
                 const queryEntities = this.extractTermsFromText(query);
                 // Base boost for knowledge graph chunks
                 if (result.chunk_type === 'entity') {
-                    graphBoost += 0.15; // Entities are inherently valuable
+                    graphBoost += 0.15;
                 }
                 else if (result.chunk_type === 'relationship') {
-                    graphBoost += 0.25; // Relationships show connections
+                    graphBoost += 0.25;
                 }
-                // Additional boost for entity matches
+                // Collect per-entity scores (instead of blind accumulation)
+                const entityScores = [];
                 const queryLower = query.toLowerCase();
                 for (const entity of chunkEntities) {
+                    let score = 0;
                     const entityLower = entity.toLowerCase();
                     // Vector-matched entity (cross-lingual: "할랄 인증" → "KMF")
                     if (queryMatchedEntities.has(entity)) {
-                        graphBoost += 0.3;
+                        score = 0.3;
                     }
                     // Exact text match with extracted terms
                     else if (queryEntities.some(qe => qe.toLowerCase() === entityLower)) {
-                        graphBoost += 0.3;
+                        score = 0.3;
                     }
                     // Partial match: entity name appears in query or vice versa
                     else if (queryLower.includes(entityLower) || entityLower.includes(queryLower)) {
-                        graphBoost += 0.2;
+                        score = 0.2;
                     }
                     // Word-level partial match
                     else if (entityLower.split(/\s+/).some(word => word.length >= 3 && queryLower.includes(word))) {
-                        graphBoost += 0.15;
+                        score = 0.15;
                     }
-                    // Connected to a vector-matched entity
+                    // Connected to a vector-matched entity (additive, independent)
                     if (connectedEntities.has(entity)) {
-                        graphBoost += 0.15;
+                        score += 0.15;
                     }
+                    if (score > 0)
+                        entityScores.push(score);
+                }
+                // Geometric decay: sort descending, apply 0.5^i decay per entity
+                entityScores.sort((a, b) => b - a);
+                let entityBoost = 0;
+                for (let i = 0; i < entityScores.length; i++) {
+                    entityBoost += entityScores[i] * Math.pow(0.5, i);
                 }
+                // Hard cap to prevent graph domination
+                graphBoost += Math.min(entityBoost, 0.4);
             }
             // Generate semantic summary
             const { summary, keyHighlight, relevanceScore } = await this.generateContentSummary(result.text, primaryQueryEmbedding, chunkEntities, result.chunk_type === 'relationship' ? 1 : 2 // Shorter summary for relationships

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "rag-memory-epf-mcp",
-  "version": "3.1.4",
+  "version": "3.2.0",
   "description": "MCP server for project-local RAG memory with knowledge graph and multilingual vector search",
   "license": "MIT",
   "author": "bripin123",