Spaces:
Sleeping
Sleeping
Update app.py
Browse files
app.py
CHANGED
|
@@ -1,6 +1,3 @@
|
|
| 1 |
-
# 檔案名稱: app.py
|
| 2 |
-
# 部署在 Hugging Face Spaces (已修正 KeyError)
|
| 3 |
-
|
| 4 |
import gradio as gr
|
| 5 |
import requests
|
| 6 |
import json
|
|
@@ -19,25 +16,37 @@ SIMILARITY_THRESHOLD = 0.90
|
|
| 19 |
print("--- [1/5] 開始初始化應用 ---")
|
| 20 |
|
| 21 |
# --- 1. 載入知識庫 ---
|
|
|
|
|
|
|
| 22 |
try:
|
| 23 |
print(f"--- [2/5] 正在從 '{DATASET_REPO_ID}' 載入知識庫... ---")
|
| 24 |
-
|
| 25 |
-
dataset = load_dataset(DATASET_REPO_ID, token=HF_TOKEN)
|
| 26 |
-
raw_qa_dataset = dataset['train']
|
| 27 |
-
|
| 28 |
-
# *** 關鍵修正:解析被包裹在 'text' 欄位中的 JSON ***
|
| 29 |
-
parsed_qa_data = []
|
| 30 |
-
for item in raw_qa_dataset:
|
| 31 |
-
try:
|
| 32 |
-
# item 現在是 {'text': '{"question": "...", "sql": "..."}'}
|
| 33 |
-
line_dict = json.loads(item['text'])
|
| 34 |
-
parsed_qa_data.append(line_dict)
|
| 35 |
-
except (json.JSONDecodeError, KeyError) as e:
|
| 36 |
-
print(f"警告:跳過一行無法解析的數據: {item}, 錯誤: {e}")
|
| 37 |
-
|
| 38 |
-
# 使用解析後的數據創建一個新的、格式正確的 Dataset 對象
|
| 39 |
-
qa_dataset = Dataset.from_list(parsed_qa_data)
|
| 40 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 41 |
# 載入並解析 Schema JSON
|
| 42 |
schema_file_path = "sqlite_schema_FULL.json"
|
| 43 |
hf_hub_download(repo_id=DATASET_REPO_ID, filename=schema_file_path, repo_type='dataset', local_dir='.', token=HF_TOKEN)
|
|
@@ -45,15 +54,16 @@ try:
|
|
| 45 |
with open(schema_file_path, 'r', encoding='utf-8') as f:
|
| 46 |
schema_data = json.load(f)
|
| 47 |
|
| 48 |
-
print(f"--- >
|
|
|
|
| 49 |
except Exception as e:
|
| 50 |
print(f"!!! 致命錯誤: 無法載入或解析 Dataset '{DATASET_REPO_ID}'.")
|
| 51 |
print(f"詳細錯誤: {e}")
|
| 52 |
qa_dataset = Dataset.from_dict({"question": ["示例問題"], "sql": ["SELECT 'Dataset failed to load'"]})
|
| 53 |
-
schema_data = {}
|
| 54 |
|
| 55 |
# --- 2. 構建 DDL 和初始化檢索模型 ---
|
| 56 |
def load_schema_as_ddl(schema_dict: dict) -> str:
|
|
|
|
| 57 |
ddl_string = ""
|
| 58 |
for table_name, columns in schema_dict.items():
|
| 59 |
if not isinstance(columns, list): continue
|
|
@@ -67,33 +77,53 @@ SCHEMA_DDL = load_schema_as_ddl(schema_data)
|
|
| 67 |
print("--- [3/5] 正在載入句向量模型 (all-MiniLM-L6-v2)... ---")
|
| 68 |
embedder = SentenceTransformer('all-MiniLM-L6-v2', device='cpu')
|
| 69 |
|
| 70 |
-
# *** 關鍵修正:現在 qa_dataset 的結構是正確的了 ***
|
| 71 |
questions = [item['question'] for item in qa_dataset]
|
| 72 |
sql_answers = [item['sql'] for item in qa_dataset]
|
| 73 |
|
| 74 |
-
|
| 75 |
-
|
| 76 |
-
print("---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 77 |
|
| 78 |
# --- 3. 混合系統核心邏輯 ---
|
| 79 |
def get_sql_query(user_question: str):
|
| 80 |
-
# (
|
| 81 |
if not user_question:
|
| 82 |
return "請輸入您的問題。", "日誌:用戶未輸入問題。"
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 83 |
question_embedding = embedder.encode(user_question, convert_to_tensor=True)
|
| 84 |
hits = util.semantic_search(question_embedding, question_embeddings, top_k=5)
|
| 85 |
-
|
| 86 |
-
|
| 87 |
-
|
| 88 |
-
|
| 89 |
-
|
| 90 |
-
|
| 91 |
-
|
| 92 |
-
|
| 93 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 94 |
examples_context = ""
|
| 95 |
-
|
| 96 |
-
|
|
|
|
|
|
|
| 97 |
prompt = f"""### Task
|
| 98 |
Generate a SQLite SQL query that answers the following user question.
|
| 99 |
Your response must contain ONLY the SQL query. Do not add any explanation.
|
|
@@ -130,8 +160,9 @@ Your response must contain ONLY the SQL query. Do not add any explanation.
|
|
| 130 |
# --- 4. 創建 Gradio Web 界面 ---
|
| 131 |
print("--- [5/5] 正在創建 Gradio Web 界面... ---")
|
| 132 |
with gr.Blocks(theme=gr.themes.Soft()) as demo:
|
| 133 |
-
# (
|
| 134 |
gr.Markdown("# 智能 Text-to-SQL 系統 (混合模式)")
|
|
|
|
| 135 |
gr.Markdown("輸入您的自然語言問題,系統將首先嘗試從知識庫中快速檢索答案。如果問題較新穎,則會調用雲端大語言模型生成SQL。")
|
| 136 |
with gr.Row():
|
| 137 |
question_input = gr.Textbox(label="輸入您的問題", placeholder="例如:去年Nike的總業績是多少?", scale=4)
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
import gradio as gr
|
| 2 |
import requests
|
| 3 |
import json
|
|
|
|
| 16 |
print("--- [1/5] 開始初始化應用 ---")
|
| 17 |
|
| 18 |
# --- 1. 載入知識庫 ---
|
| 19 |
+
qa_dataset = None
|
| 20 |
+
schema_data = {}
|
| 21 |
try:
|
| 22 |
print(f"--- [2/5] 正在從 '{DATASET_REPO_ID}' 載入知識庫... ---")
|
| 23 |
+
raw_dataset = load_dataset(DATASET_REPO_ID, token=HF_TOKEN)['train']
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 24 |
|
| 25 |
+
# *** 關鍵修正:智能解析 Dataset ***
|
| 26 |
+
# 檢查第一條數據的結構來判斷格式
|
| 27 |
+
if raw_dataset and len(raw_dataset) > 0:
|
| 28 |
+
first_item = raw_dataset[0]
|
| 29 |
+
if 'text' in first_item and 'question' not in first_item:
|
| 30 |
+
# 這是舊的 {'text': '...'} 格式,需要解析
|
| 31 |
+
print("--- > 檢測到 'text' 格式,正在解析JSON...")
|
| 32 |
+
parsed_qa_data = []
|
| 33 |
+
for item in raw_dataset:
|
| 34 |
+
try:
|
| 35 |
+
line_dict = json.loads(item['text'])
|
| 36 |
+
parsed_qa_data.append(line_dict)
|
| 37 |
+
except (json.JSONDecodeError, KeyError):
|
| 38 |
+
continue # 跳過錯誤行
|
| 39 |
+
qa_dataset = Dataset.from_list(parsed_qa_data)
|
| 40 |
+
elif 'question' in first_item and 'sql' in first_item:
|
| 41 |
+
# 這已經是正確的 {'question': ..., 'sql': ...} 格式
|
| 42 |
+
print("--- > 檢測到已解析的 'question'/'sql' 格式,直接使用。")
|
| 43 |
+
qa_dataset = raw_dataset
|
| 44 |
+
else:
|
| 45 |
+
raise ValueError(f"未知的Dataset格式: {first_item}")
|
| 46 |
+
else:
|
| 47 |
+
# 數據集為空
|
| 48 |
+
raise ValueError("載入的Dataset為空。")
|
| 49 |
+
|
| 50 |
# 載入並解析 Schema JSON
|
| 51 |
schema_file_path = "sqlite_schema_FULL.json"
|
| 52 |
hf_hub_download(repo_id=DATASET_REPO_ID, filename=schema_file_path, repo_type='dataset', local_dir='.', token=HF_TOKEN)
|
|
|
|
| 54 |
with open(schema_file_path, 'r', encoding='utf-8') as f:
|
| 55 |
schema_data = json.load(f)
|
| 56 |
|
| 57 |
+
print(f"--- > 成功載入 {len(qa_dataset)} 條問答範例和 Schema。 ---")
|
| 58 |
+
|
| 59 |
except Exception as e:
|
| 60 |
print(f"!!! 致命錯誤: 無法載入或解析 Dataset '{DATASET_REPO_ID}'.")
|
| 61 |
print(f"詳細錯誤: {e}")
|
| 62 |
qa_dataset = Dataset.from_dict({"question": ["示例問題"], "sql": ["SELECT 'Dataset failed to load'"]})
|
|
|
|
| 63 |
|
| 64 |
# --- 2. 構建 DDL 和初始化檢索模型 ---
|
| 65 |
def load_schema_as_ddl(schema_dict: dict) -> str:
|
| 66 |
+
# (此函式無需修改)
|
| 67 |
ddl_string = ""
|
| 68 |
for table_name, columns in schema_dict.items():
|
| 69 |
if not isinstance(columns, list): continue
|
|
|
|
| 77 |
print("--- [3/5] 正在載入句向量模型 (all-MiniLM-L6-v2)... ---")
|
| 78 |
embedder = SentenceTransformer('all-MiniLM-L6-v2', device='cpu')
|
| 79 |
|
|
|
|
| 80 |
questions = [item['question'] for item in qa_dataset]
|
| 81 |
sql_answers = [item['sql'] for item in qa_dataset]
|
| 82 |
|
| 83 |
+
# 只有在 questions 列表不為空時才進行計算
|
| 84 |
+
if questions:
|
| 85 |
+
print(f"--- [4/5] 正在為 {len(questions)} 個問題計算向量... ---")
|
| 86 |
+
question_embeddings = embedder.encode(questions, convert_to_tensor=True, show_progress_bar=True)
|
| 87 |
+
print("--- > 向量計算完成! ---")
|
| 88 |
+
else:
|
| 89 |
+
print("--- [4/5] 警告:沒有可用的問題來計算向量。檢索功能將不可用。---")
|
| 90 |
+
question_embeddings = torch.Tensor([])
|
| 91 |
+
|
| 92 |
|
| 93 |
# --- 3. 混合系統核心邏輯 ---
|
| 94 |
def get_sql_query(user_question: str):
|
| 95 |
+
# (此函式剩餘部分幾乎無需修改)
|
| 96 |
if not user_question:
|
| 97 |
return "請輸入您的問題。", "日誌:用戶未輸入問題。"
|
| 98 |
+
|
| 99 |
+
# 增加一個檢查,確保知識庫不是空的
|
| 100 |
+
if len(questions) == 0:
|
| 101 |
+
log_message = "錯誤:知識庫為空,無法進行檢索。"
|
| 102 |
+
return "系統錯誤:知識庫未成功載入。", log_message
|
| 103 |
+
|
| 104 |
question_embedding = embedder.encode(user_question, convert_to_tensor=True)
|
| 105 |
hits = util.semantic_search(question_embedding, question_embeddings, top_k=5)
|
| 106 |
+
|
| 107 |
+
if not hits or not hits[0]:
|
| 108 |
+
log_message = "檢索失敗:找不到任何相似的問題。"
|
| 109 |
+
# 即使檢索失敗,也應該嘗試調用 LLM
|
| 110 |
+
else:
|
| 111 |
+
hits = hits[0]
|
| 112 |
+
most_similar_hit = hits[0]
|
| 113 |
+
similarity_score = most_similar_hit['score']
|
| 114 |
+
log_message = f"檢索到最相似問題: '{questions[most_similar_hit['corpus_id']]}' (相似度: {similarity_score:.4f})"
|
| 115 |
+
|
| 116 |
+
if similarity_score > SIMILARITY_THRESHOLD:
|
| 117 |
+
sql_result = sql_answers[most_similar_hit['corpus_id']]
|
| 118 |
+
log_message += f"\n相似度 > {SIMILARITY_THRESHOLD},[模式: 直接返回]。"
|
| 119 |
+
return sql_result, log_message
|
| 120 |
+
|
| 121 |
+
log_message += f"\n相似度低於閾值或檢索失敗,[模式: LLM生成]。正在構建 Prompt..."
|
| 122 |
examples_context = ""
|
| 123 |
+
if hits: # 只有在檢索到結果時才添加範例
|
| 124 |
+
for hit in hits[:3]:
|
| 125 |
+
examples_context += f"### A user asks: {questions[hit['corpus_id']]}\n{sql_answers[hit['corpus_id']]}\n\n"
|
| 126 |
+
|
| 127 |
prompt = f"""### Task
|
| 128 |
Generate a SQLite SQL query that answers the following user question.
|
| 129 |
Your response must contain ONLY the SQL query. Do not add any explanation.
|
|
|
|
| 160 |
# --- 4. 創建 Gradio Web 界面 ---
|
| 161 |
print("--- [5/5] 正在創建 Gradio Web 界面... ---")
|
| 162 |
with gr.Blocks(theme=gr.themes.Soft()) as demo:
|
| 163 |
+
# (此部分無需修改)
|
| 164 |
gr.Markdown("# 智能 Text-to-SQL 系統 (混合模式)")
|
| 165 |
+
# ... (Gradio界面代碼與之前相同)
|
| 166 |
gr.Markdown("輸入您的自然語言問題,系統將首先嘗試從知識庫中快速檢索答案。如果問題較新穎,則會調用雲端大語言模型生成SQL。")
|
| 167 |
with gr.Row():
|
| 168 |
question_input = gr.Textbox(label="輸入您的問題", placeholder="例如:去年Nike的總業績是多少?", scale=4)
|