Deploy RAG Tourism API

app/main.py

@@ -813,6 +813,11 @@ async def search_locations(query: TourismQuery, background_tasks: BackgroundTask
                     # Boost PhoBERT confidence for primary usage
                     # Convert numpy.float32 to Python float to fix Pydantic serialization
                     similarity_score = float(loc.get('similarity_score', 0.5))
+                    # Ngưỡng liên quan: loại kết quả vector quá xa nghĩa
+                    # (tránh trưng card 35% lạc tỉnh). Chỉnh qua env
+                    # MIN_VECTOR_SIMILARITY (mặc định 0.35 ~ confidence 45%).
+                    if similarity_score < float(_os.getenv("MIN_VECTOR_SIMILARITY", "0.35")):
+                        continue
                     boosted_confidence = float(min(0.95, similarity_score + 0.1))
                     results.append(LocationResponse(
                         name=loc['name'],
@@ -836,6 +841,9 @@ async def search_locations(query: TourismQuery, background_tasks: BackgroundTask
             existing_names = {r.name.lower() for r in results}
             for loc in vector_results:  # ChromaDB already limits results
                 if loc['name'] and loc['name'].lower() not in existing_names:
+                    # Ngưỡng liên quan (đồng bộ với nhánh PhoBERT)
+                    if float(loc.get('similarity_score', 0.5)) < float(_os.getenv("MIN_VECTOR_SIMILARITY", "0.35")):
+                        continue
                     # Convert numpy types to Python types for Pydantic serialization
                     results.append(LocationResponse(
                         name=loc['name'],