hi

app.py

@@ -28,6 +28,7 @@ import time
 from typing import List, Dict, Any, Optional, Tuple, Union
 from functools import lru_cache
 from dataclasses import dataclass, field
+from contextlib import asynccontextmanager
 
 # ---------- 第三方函式庫 ----------
 import numpy as np
@@ -163,9 +164,8 @@ class RerankResult:
 
 # ---------- 核心 RAG 邏輯 ----------
 class RagPipeline:
-    def __init__(self, config):
-        self.config = config
-        self.state = type('state', (), {})()
+    def __init__(self):
+        # [MODIFIED] 不再傳入 AppConfig，直接引用
         if not LLM_API_CONFIG["api_key"] or not LLM_API_CONFIG["base_url"]:
              raise ValueError("LLM API Key or Base URL is not configured.")
         self.llm_client = OpenAI(api_key=LLM_API_CONFIG["api_key"], base_url=LLM_API_CONFIG["base_url"])
@@ -174,6 +174,7 @@ class RagPipeline:
         
         self.drug_name_to_ids: Dict[str, List[str]] = {}
         self.drug_vocab: Dict[str, set] = {"zh": set(), "en": set()}
+        self.state = type('state', (), {})()
 
     def _load_model(self, model_class, model_name: str, model_type: str):
         device = "cuda" if torch.cuda.is_available() else "cpu"
@@ -182,10 +183,19 @@ class RagPipeline:
             return model_class(model_name, device=device)
         except Exception as e:
             log.warning(f"載入模型至 {device} 失敗: {e}。嘗試切換至 CPU。")
-            return model_class(model_name, device="cpu")
+            try:
+                return model_class(model_name, device="cpu")
+            except Exception as e_cpu:
+                log.error(f"切換至 CPU 仍無法載入模型: {model_name}。請確認模型路徑或網路連線。錯誤訊息: {e_cpu}")
+                raise RuntimeError(f"模型載入失敗: {model_name}")
 
     def load_data(self):
         log.info("開始載入資料與模型...")
+        # [MODIFIED] 增加檔案存在性檢查
+        for path in [CSV_PATH, FAISS_INDEX, SENTENCES_PKL, BM25_PKL]:
+            if not pathlib.Path(path).exists():
+                raise FileNotFoundError(f"必要的資料檔案不存在: {path}")
+
         try:
             self.df_csv = pd.read_csv(CSV_PATH, dtype=str).fillna('')
             # [MODIFIED] 增加必要欄位檢查
@@ -236,22 +246,28 @@ class RagPipeline:
             for part in parts:
                 if re.search(r'[\u4e00-\u9fff]', part):
                     self.drug_vocab["zh"].add(part)
-                    try:
-                        jieba.add_word(part, freq=2_000_000)
-                    except Exception:
-                        pass
+                    # [MODIFIED] 檢查詞彙是否已存在
+                    if part not in jieba.dt.FREQ:
+                        try:
+                            jieba.add_word(part, freq=2_000_000)
+                        except Exception:
+                            pass
                 else:
                     self.drug_vocab["en"].add(part)
         for alias in DRUG_NAME_MAPPING:
             self.drug_vocab["en"].add(alias.lower())
             if re.search(r'[\u4e00-\u9fff]', alias):
-                try:
-                    jieba.add_word(alias, freq=2_000_000)
-                except Exception:
-                    pass
+                if alias not in jieba.dt.FREQ:
+                    try:
+                        jieba.add_word(alias, freq=2_000_000)
+                    except Exception:
+                        pass
 
     @retry(stop=stop_after_attempt(3), wait=wait_fixed(2))
     def _llm_call(self, messages, **kwargs) -> str:
+        start_time = time.time()
+        log.info(f"LLM 呼叫開始. 模型: {LLM_API_CONFIG['model']}, max_tokens: {kwargs.get('max_tokens', 'N/A')}, temperature: {kwargs.get('temperature', 'N/A')}")
+        
         try:
             config = {**LLM_MODEL_CONFIG, **kwargs}
             response = self.llm_client.chat.completions.create(
@@ -260,15 +276,25 @@ class RagPipeline:
                 temperature=config["temperature"],
                 max_tokens=config["max_tokens"],
             )
-            content = response.choices[0].message.content
-            # [MODIFIED] 確保回傳值為非空字串
+            
+            # [MODIFIED] 檢查回應結構並使用 getattr 安全地獲取內容
+            if not response or not response.choices or not response.choices[0].message:
+                log.error(f"LLM 呼叫成功 (200 OK)，但回傳的 JSON 結構不完整。回傳: {response.model_dump_json() if response else 'None'}")
+                raise ValueError("LLM response content is empty or not a string.")
+
+            content = getattr(response.choices[0].message, "content", None)
             if not isinstance(content, str) or not content.strip():
+                log.error(f"LLM 呼叫成功 (200 OK)，但回傳內容為空。Response: {content}")
                 raise ValueError("LLM response content is empty or not a string.")
+            
+            elapsed = time.time() - start_time
+            log.info(f"LLM 呼叫完成，耗時: {elapsed:.2f} 秒。��容長度: {len(content)} 字。")
             return content
         except Exception as e:
-            log.error(f"LLM API 呼叫失敗: {e}")
+            log.error(f"LLM API 呼叫失敗: {e}", exc_info=True)
             raise
 
+    # [MODIFIED] 實現動態流程，根據查詢複雜度決定是否使用 Reranker
     def answer_question(self, q_orig: str) -> str:
         start_time = time.time()
         log.info(f"===== 處理新查詢: '{q_orig}' =====")
@@ -277,17 +303,32 @@ class RagPipeline:
             if not drug_ids:
                 log.info("找不到藥品 ID，無法回答。")
                 return f"抱歉，資料庫中找不到該藥品。請確認藥品名稱，或直接諮詢醫師/藥師。{DISCLAIMER}"
-            log.info(f"步驟 1/5: 找到藥品 ID: {drug_ids}")
+            log.info(f"步驟 1/5: 找到藥品 ID: {drug_ids}，耗時: {time.time() - start_time:.2f} 秒")
+            step_start = time.time()
 
             analysis = self._analyze_query(q_orig)
             sub_queries, intents = analysis.get("sub_queries", [q_orig]), analysis.get("intents", [])
-            log.info(f"步驟 2/5: 意圖分析完成。子問題: {sub_queries}, 意圖: {intents}")
-
+            is_simple_query = self._is_simple_query(sub_queries, intents)
+            log.info(f"步驟 2/5: 意圖分析完成。子問題: {sub_queries}, 意圖: {intents}。判定為簡單查詢: {is_simple_query}。耗時: {time.time() - step_start:.2f} 秒")
+            step_start = time.time()
+            
             all_candidates = self._retrieve_candidates_for_all_queries(drug_ids, sub_queries, intents)
-            log.info(f"步驟 3/5: 檢索完成。所有子查詢共找到 {len(all_candidates)} 個不重複候選 chunks。")
-
-            reranked_results = self._rerank_with_crossencoder(q_orig, all_candidates)
-            log.info(f"步驟 4/5: Reranker 最終選出 {len(reranked_results)} 個高品質候選。")
+            log.info(f"步驟 3/5: 檢索完成。所有子查詢共找到 {len(all_candidates)} 個不重複候選 chunks。耗時: {time.time() - step_start:.2f} 秒")
+            step_start = time.time()
+
+            if is_simple_query:
+                log.info("偵測到簡單查詢，跳過 Reranker 步驟。")
+                final_candidates = all_candidates[:TOP_K_SENTENCES]
+                reranked_results = [
+                    RerankResult(idx=c.idx, rerank_score=c.fused_score, text=self.state.sentences[c.idx], meta=self.state.meta[c.idx])
+                    for c in final_candidates
+                ]
+            else:
+                log.info("偵測到複雜查詢，執行 Reranker。")
+                reranked_results = self._rerank_with_crossencoder(q_orig, all_candidates)
+            
+            log.info(f"步驟 4/5: 最終選出 {len(reranked_results)} 個高品質候選。耗時: {time.time() - step_start:.2f} 秒")
+            step_start = time.time()
 
             context = self._build_context(reranked_results)
             if not context:
@@ -298,7 +339,7 @@ class RagPipeline:
             answer = self._llm_call([{"role": "user", "content": prompt}])
             
             final_answer = f"{answer.strip()}\n\n{DISCLAIMER}"
-            log.info(f"步驟 5/5: 答案生成完成。答案長度: {len(answer.strip())} 字。")
+            log.info(f"步驟 5/5: 答案生成完成。答案長度: {len(answer.strip())} 字。耗時: {time.time() - step_start:.2f} 秒")
             log.info(f"===== 查詢處理完成，總耗時: {time.time() - start_time:.2f} 秒 =====")
             return final_answer
 
@@ -306,6 +347,10 @@ class RagPipeline:
             log.error(f"處理查詢 '{q_orig}' 時發生嚴重錯誤: {e}", exc_info=True)
             return f"處理您的問題時發生內部錯誤，請稍後再試。{DISCLAIMER}"
 
+    def _is_simple_query(self, sub_queries: List[str], intents: List[str]) -> bool:
+        # 如果意圖分析回傳的子查詢數量 <= 1，且意圖分類數量也 <= 1，則判定為簡單問題
+        return len(sub_queries) <= 1 and len(intents) <= 1
+
     @lru_cache(maxsize=128)
     def _find_drug_ids_from_name(self, query: str) -> List[str]:
         q = query.lower()
@@ -420,12 +465,13 @@ class RagPipeline:
         try:
             expanded_query = self._llm_call([{"role": "user", "content": prompt}])
             if expanded_query and expanded_query.strip():
+                log.info(f"查詢擴展成功。原始: '{query}', 擴展後: '{expanded_query}'")
                 return expanded_query
             else:
-                log.warning(f"Query expansion for '{query}' returned an empty result. Using original query.")
+                log.warning(f"查詢擴展回傳空內容。原始查詢: '{query}'。將使用原始查詢。")
                 return query
         except Exception as e:
-            log.error(f"Query expansion for '{query}' failed: {e}. Using original query.")
+            log.error(f"查詢擴展失敗: {e}。原始查詢: '{query}'。將使用原始查詢。")
             return query
 
     def _rerank_with_crossencoder(self, query: str, candidates: List[FusedCandidate]) -> List[RerankResult]:
@@ -460,31 +506,41 @@ class RagPipeline:
     
     # [MODIFIED] 增強 JSON 解析的穩健性，從字串中提取 JSON 物件
     def _safe_json_parse(self, s: str, default: Any = None) -> Any:
-        m = re.search(r'\{.*?\}', s, re.DOTALL)  # 非貪婪
-        if m:
-            s = m.group(0)
         try:
+            # 嘗試解析完整字串
             return json.loads(s)
-        except Exception:
-            log.warning(f"無法解析 LLM 回傳的 JSON: {s[:200]}...")
-            return default
+        except json.JSONDecodeError:
+            log.warning(f"無法解析完整 JSON。嘗試從字串中提取: {s[:200]}...")
+            # 如果失敗，嘗試用 regex 提取第一個 JSON 物件
+            m = re.search(r'\{.*?\}', s, re.DOTALL)
+            if m:
+                try:
+                    return json.loads(m.group(0))
+                except json.JSONDecodeError:
+                    log.warning(f"提取的 JSON 仍無法解析: {m.group(0)[:100]}...")
+        return default
 
 # ---------- FastAPI 事件與路由 ----------
-app = FastAPI()
-rag_pipeline: Optional[RagPipeline] = None
-
 # [MODIFIED] 將 LINE 配置集中管理並進行啟動時檢查
 class AppConfig:
     CHANNEL_ACCESS_TOKEN = _require_env("CHANNEL_ACCESS_TOKEN")
     CHANNEL_SECRET = _require_env("CHANNEL_SECRET")
 
-@app.on_event("startup")
-async def startup_event():
-    global rag_pipeline
+rag_pipeline: Optional[RagPipeline] = None
+
+# [MODIFIED] 使用 lifespan context manager
+@asynccontextmanager
+async def lifespan(app: FastAPI):
     _require_llm_config()
-    rag_pipeline = RagPipeline(AppConfig)
+    global rag_pipeline
+    rag_pipeline = RagPipeline()
     rag_pipeline.load_data()
     log.info("啟動完成，服務準備就緒。")
+    yield
+    # 若有資源需要關閉可在這裡實作
+    log.info("服務關閉中。")
+
+app = FastAPI(lifespan=lifespan)
 
 @app.post("/webhook")
 async def handle_webhook(request: Request, background_tasks: BackgroundTasks):