Update app.py

app.py

@@ -79,49 +79,95 @@ def search_similar_chat(query: str, max_results: int = 100) -> List[Dict]:
         conn = get_db_conn()
         register_vector(conn)
         
-        # Java 방식: 문자열 포맷팅 사용한 SQL 쿼리
+        # 개선된 SQL 쿼리 - 가중 평균 방식으로 계산하고 각 임베딩 유형별 점수 포함
         sql = f"""
             WITH embeddings AS (
                 SELECT 
                     id, 
                     metadata, 
                     content,
-                    CASE WHEN full_embedding IS NOT NULL THEN 1 - (full_embedding <=> '{query_vector}'::vector) ELSE 0 END * {full_w} as full_sim,
-                    CASE WHEN topic_embedding IS NOT NULL THEN 1 - (topic_embedding <=> '{query_vector}'::vector) ELSE 0 END * {topic_w} as topic_sim,
-                    CASE WHEN customer_embedding IS NOT NULL THEN 1 - (customer_embedding <=> '{query_vector}'::vector) ELSE 0 END * {customer_w} as customer_sim,
-                    CASE WHEN agent_embedding IS NOT NULL THEN 1 - (agent_embedding <=> '{query_vector}'::vector) ELSE 0 END * {agent_w} as agent_sim
+                    -- 각 임베딩 유형별 유사도 계산 (NULL이 아닌 경우만)
+                    CASE WHEN full_embedding IS NOT NULL 
+                         THEN 1 - (full_embedding <=> '{query_vector}'::vector) 
+                         ELSE NULL END as full_sim,
+                    
+                    CASE WHEN topic_embedding IS NOT NULL 
+                         THEN 1 - (topic_embedding <=> '{query_vector}'::vector) 
+                         ELSE NULL END as topic_sim,
+                    
+                    CASE WHEN customer_embedding IS NOT NULL 
+                         THEN 1 - (customer_embedding <=> '{query_vector}'::vector) 
+                         ELSE NULL END as customer_sim,
+                    
+                    CASE WHEN agent_embedding IS NOT NULL 
+                         THEN 1 - (agent_embedding <=> '{query_vector}'::vector) 
+                         ELSE NULL END as agent_sim,
+                    
+                    -- 유효한 임베딩 카운트 (0으로 나누기 방지)
+                    (CASE WHEN full_embedding IS NOT NULL THEN 1 ELSE 0 END +
+                     CASE WHEN topic_embedding IS NOT NULL THEN 1 ELSE 0 END +
+                     CASE WHEN customer_embedding IS NOT NULL THEN 1 ELSE 0 END +
+                     CASE WHEN agent_embedding IS NOT NULL THEN 1 ELSE 0 END) as valid_count
                 FROM vector_store_multi_embeddings
-                WHERE full_embedding IS NOT NULL
-                OR topic_embedding IS NOT NULL
-                OR customer_embedding IS NOT NULL
-                OR agent_embedding IS NOT NULL
+                WHERE (full_embedding IS NOT NULL 
+                      OR topic_embedding IS NOT NULL 
+                      OR customer_embedding IS NOT NULL 
+                      OR agent_embedding IS NOT NULL)
+            ),
+            weighted_scores AS (
+                SELECT 
+                    id, 
+                    metadata, 
+                    content,
+                    full_sim,
+                    topic_sim,
+                    customer_sim,
+                    agent_sim,
+                    valid_count,
+                    
+                    -- 가중치를 적용한 총 유사도 점수 계산
+                    (COALESCE(full_sim, 0) * {full_w} +
+                     COALESCE(topic_sim, 0) * {topic_w} +
+                     COALESCE(customer_sim, 0) * {customer_w} +
+                     COALESCE(agent_sim, 0) * {agent_w}) as weighted_sum
+                FROM embeddings
             )
             SELECT 
                 id, 
                 metadata, 
                 content,
-                (full_sim + topic_sim + customer_sim + agent_sim) as combined_similarity
-            FROM embeddings
+                full_sim,
+                topic_sim,
+                customer_sim,
+                agent_sim,
+                weighted_sum as combined_similarity
+            FROM weighted_scores
+            WHERE weighted_sum >= {threshold}
             ORDER BY combined_similarity DESC
             LIMIT {limit}
         """
         
         with conn.cursor() as cur:
-            print(f"쿼리 실행 - Java 방식 포맷팅, 가중치 설정={full_w}, {topic_w}, {customer_w}, {agent_w}, 결과 제한={limit}")
+            print(f"쿼리 실행 - 개선된 방식, 가중치 설정={full_w}, {topic_w}, {customer_w}, {agent_w}, 결과 제한={limit}")
             # Java 방식: 매개변수 없이 직접 쿼리 실행
             cur.execute(sql)
             rows = cur.fetchall()
             
             print(f"검색 결과: 총 {len(rows)}개 데이터 조회됨")
             if len(rows) > 0:
-                print(f"첫 번째 결과 ID: {rows[0][0]}, 유사도: {float(rows[0][3])}")
+                print(f"첫 번째 결과 ID: {rows[0][0]}, 유사도: {float(rows[0][7])}")
+                print(f"첫 번째 결과 세부 유사도 - full: {rows[0][3]}, topic: {rows[0][4]}, customer: {rows[0][5]}, agent: {rows[0][6]}")
             
             results = []
             for row in rows:
                 id_val = row[0]
                 metadata_json = row[1]
                 content = row[2]
-                similarity_score = float(row[3])
+                full_similarity = None if row[3] is None else float(row[3])
+                topic_similarity = None if row[4] is None else float(row[4])
+                customer_similarity = None if row[5] is None else float(row[5])
+                agent_similarity = None if row[6] is None else float(row[6])
+                similarity_score = float(row[7])
                 
                 # 메타데이터 파싱
                 try:
@@ -132,7 +178,13 @@ def search_similar_chat(query: str, max_results: int = 100) -> List[Dict]:
                         "similarityScore": similarity_score,
                         "content": content,
                         "chatId": get_text_value(metadata, "chatId"),
-                        "topic": get_text_value(metadata, "topic")
+                        "topic": get_text_value(metadata, "topic"),
+                        "similarityDetails": {
+                            "full": full_similarity,
+                            "topic": topic_similarity,
+                            "customer": customer_similarity,
+                            "agent": agent_similarity
+                        }
                     }
                     
                     # 시간 필드 변환 없이 그대로 사용
@@ -148,15 +200,14 @@ def search_similar_chat(query: str, max_results: int = 100) -> List[Dict]:
                     print(f"문제가 발생한 메타데이터: {metadata_json[:200]}...")
                     continue
         
-        # 임계값 필터링
-        filtered_results = [r for r in results if r["similarityScore"] >= threshold]
-        print(f"임계값({threshold}) 이상 결과: {len(filtered_results)}개 / 전체 {len(results)}개")
+        print(f"임계값({threshold}) 이상 결과: {len(results)}개")
         
-        if len(filtered_results) > 0:
-            print(f"가장 높은 유사도 점수: {filtered_results[0]['similarityScore']}")
-            print(f"상위 결과 챗ID: {filtered_results[0].get('chatId')}, 주제: {filtered_results[0].get('topic', '')[:50]}...")
+        if len(results) > 0:
+            print(f"가장 높은 유사도 점수: {results[0]['similarityScore']}")
+            print(f"상위 결과 챗ID: {results[0].get('chatId')}, 주제: {results[0].get('topic', '')[:50]}...")
+            print(f"상위 결과 세부 유사도: {results[0]['similarityDetails']}")
         
-        return filtered_results
+        return results
     
     except Exception as e:
         print(f"다중 임베딩 검색 중 오류 발생: {str(e)}")