Update app.py

app.py

@@ -63,6 +63,9 @@ print(f"N_BATCH={N_BATCH}")
 print(f"暫存目錄: {TEMP_DIR}")
 print("=" * 60)
 
+# 關閉 Gradio 分析上報，減少不必要的請求與雜訊
+os.environ.setdefault("GRADIO_ANALYTICS_ENABLED", "False")
+
 # ==================== 工具函數 ====================
 def get_current_time():
     return datetime.now().strftime('%Y-%m-%d %H:%M:%S')
@@ -136,6 +139,45 @@ def parse_sql_from_response(response_text: str) -> Optional[str]:
 
     return None
 
+def sanitize_sql(sql_text: str) -> str:
+    """
+    將模型輸出清理為更可執行的 SQL：
+    - 全形標點轉半形（（ ） ； ： ， 。 等）
+    - 過濾清單符號（- 開頭）
+    - 僅保留第一個 SELECT 片段直到分號或字串結尾
+    - 簡易平衡多餘的右括號
+    - 補齊分號
+    """
+    if not sql_text:
+        return sql_text
+    s = sql_text.strip()
+    trans = str.maketrans({'（': '(', '）': ')', '；': ';', '：': ':', '，': ',', '。': '.', '【': '(', '】': ')'})
+    s = s.translate(trans)
+    cleaned_lines = []
+    for line in s.splitlines():
+        line = line.strip()
+        if line.startswith('- '):
+            continue
+        cleaned_lines.append(line)
+    s = ' '.join(cleaned_lines)
+    m = re.search(r"(SELECT\s+.*?)(;|$)", s, flags=re.IGNORECASE | re.DOTALL)
+    if m:
+        s = m.group(1)
+    open_cnt, close_cnt = s.count('('), s.count(')')
+    if close_cnt > open_cnt:
+        excess = close_cnt - open_cnt
+        out = []
+        for ch in s[::-1]:
+            if ch == ')' and excess > 0:
+                excess -= 1
+                continue
+            out.append(ch)
+        s = ''.join(out[::-1])
+    s = s.rstrip(' .)')
+    if s and not s.endswith(';'):
+        s += ';'
+    return s
+
 # ==================== Text-to-SQL 核心類 ====================
 class TextToSQLSystem:
     def __init__(self, embed_model_name=EMBED_MODEL_NAME):
@@ -756,7 +798,10 @@ SELECT
             return None, f"無法解析SQL。原始回應:\n{raw_response}"
 
         self._log(f"📊 解析出的原始 SQL: {parsed_sql}", "DEBUG")
-        fixed_sql = " " + parsed_sql.strip() + " "
+        normalized_sql = sanitize_sql(parsed_sql)
+        if normalized_sql != parsed_sql:
+            self._log(f"🧹 清理後 SQL: {normalized_sql}", "DEBUG")
+        fixed_sql = " " + normalized_sql.strip() + " "
         fixes_applied = []
 
         # 方言修正
@@ -783,6 +828,15 @@ SELECT
                 fixed_sql = re.sub(pat, correct, fixed_sql, flags=re.IGNORECASE)
                 fixes_applied.append(f"映射 Schema: '{wrong}' -> '{correct}'")
 
+        # 若沒有 FROM，補上預設資料來源
+        if re.search(r"\bSELECT\b", fixed_sql, re.IGNORECASE) and not re.search(r"\bFROM\b", fixed_sql, re.IGNORECASE):
+            if re.search(r"COUNT\s*\(\s*\*\s*\)", fixed_sql, re.IGNORECASE):
+                fixed_sql = " SELECT COUNT(DISTINCT jt.JobNo) FROM JobTimeline AS jt WHERE jt.ReportAuthorization IS NOT NULL "
+                fixes_applied.append("補上預設 FROM JobTimeline (COUNT 專用)")
+            else:
+                fixed_sql = " SELECT * FROM JobTimeline AS jt WHERE jt.ReportAuthorization IS NOT NULL "
+                fixes_applied.append("補上預設 FROM JobTimeline")
+
         status = "AI 生成並成功修正" if fixes_applied else "AI 生成且無需修正"
         return self._finalize_sql(fixed_sql, status)
 
@@ -876,7 +930,12 @@ def process_query(q: str, prompt_override: str = ""):
             return po, "override", logs
         # 否則當作完整 prompt 丟給 LLM
         text_to_sql_system._log("使用 prompt_override 直接調用 LLM")
-        response = text_to_sql_system.huggingface_api_call(po)
+        constrained_po = (
+            po.rstrip()
+            + "\n\nReturn only the final SQL query in a fenced code block (```sql ... ```). "
+            + "Do not output narration, bullets, or explanations. The SQL must start with SELECT and end with a semicolon."
+        )
+        response = text_to_sql_system.huggingface_api_call(constrained_po)
         fixed_sql, status_message = text_to_sql_system._validate_and_fix_sql(q or "", response)
         if not fixed_sql:
             fixed_sql = text_to_sql_system._generate_fallback_sql(po)
@@ -927,6 +986,55 @@ with gr.Blocks(theme=gr.themes.Soft(), title="Text-to-SQL 智能助手 (HF Space
     btn.click(process_query, inputs=[inp, prompt_override], outputs=[sql_out, status, logs], api_name="/predict")
     inp.submit(process_query, inputs=[inp, prompt_override], outputs=[sql_out, status, logs])
 
+    # ========== 健康檢查端點 /health ==========
+    @demo.add_server_route("/health", methods=["GET"])  # type: ignore[attr-defined]
+    def health_endpoint():
+        endpoints = []
+        try:
+            cfg = getattr(demo, "config", None)
+            if isinstance(cfg, dict):
+                deps = cfg.get("dependencies") or []
+                for dep in deps:
+                    endpoints.append({
+                        "api_name": dep.get("api_name"),
+                        "fn_index": dep.get("fn_index"),
+                        "inputs_count": len(dep.get("inputs") or []),
+                        "outputs_count": len(dep.get("outputs") or []),
+                    })
+        except Exception:
+            pass
+
+        if not endpoints:
+            endpoints.append({
+                "api_name": "/predict",
+                "fn_index": None,
+                "inputs_count": 2,
+                "outputs_count": 3,
+            })
+
+        env_info = {
+            "USE_GPU": USE_GPU,
+            "DEVICE": DEVICE,
+            "N_GPU_LAYERS": N_GPU_LAYERS,
+            "THREADS": THREADS,
+            "CTX": CTX,
+            "MAX_TOKENS": MAX_TOKENS,
+            "FEW_SHOT_EXAMPLES_COUNT": FEW_SHOT_EXAMPLES_COUNT,
+            "ENABLE_INDEX": ENABLE_INDEX,
+            "EMBED_BATCH": EMBED_BATCH,
+            "N_BATCH": N_BATCH,
+            "GGUF_REPO_ID": GGUF_REPO_ID,
+            "GGUF_FILENAME": GGUF_FILENAME,
+        }
+
+        server_info = {
+            "time": get_current_time(),
+            "gradio_version": getattr(gr, "__version__", "unknown"),
+            "pid": os.getpid(),
+        }
+
+        return {"status": "ok", "endpoints": endpoints, "env": env_info, "server": server_info}
+
 if __name__ == "__main__":
     demo.launch(
         server_name="0.0.0.0",