Spaces:
Sleeping
Sleeping
Update app.py
Browse files
app.py
CHANGED
|
@@ -1,3 +1,6 @@
|
|
|
|
|
|
|
|
|
|
|
| 1 |
import gradio as gr
|
| 2 |
import requests
|
| 3 |
import json
|
|
@@ -8,14 +11,9 @@ import torch
|
|
| 8 |
from huggingface_hub import hf_hub_download
|
| 9 |
|
| 10 |
# --- 配置區 ---
|
| 11 |
-
# 從 Hugging Face Secrets 獲取 Token,這是最安全的方式
|
| 12 |
HF_TOKEN = os.environ.get("HF_TOKEN")
|
| 13 |
-
# 您的 Dataset 倉庫 ID
|
| 14 |
DATASET_REPO_ID = "Paul720810/Text-to-SQL-Softline"
|
| 15 |
-
# 雲端 LLM 模型的 API URL (推薦使用 CodeLlama-34b,它更強大)
|
| 16 |
LLM_API_URL = "https://api-inference.huggingface.co/models/codellama/CodeLlama-34b-Instruct-hf"
|
| 17 |
-
|
| 18 |
-
# 相似度閾值,高於此值則直接返回答案
|
| 19 |
SIMILARITY_THRESHOLD = 0.90
|
| 20 |
|
| 21 |
print("--- [1/5] 開始初始化應用 ---")
|
|
@@ -23,9 +21,22 @@ print("--- [1/5] 開始初始化應用 ---")
|
|
| 23 |
# --- 1. 載入知識庫 ---
|
| 24 |
try:
|
| 25 |
print(f"--- [2/5] 正在從 '{DATASET_REPO_ID}' 載入知識庫... ---")
|
| 26 |
-
#
|
| 27 |
-
dataset = load_dataset(DATASET_REPO_ID, token=HF_TOKEN
|
| 28 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 29 |
|
| 30 |
# 載入並解析 Schema JSON
|
| 31 |
schema_file_path = "sqlite_schema_FULL.json"
|
|
@@ -34,13 +45,10 @@ try:
|
|
| 34 |
with open(schema_file_path, 'r', encoding='utf-8') as f:
|
| 35 |
schema_data = json.load(f)
|
| 36 |
|
| 37 |
-
print(f"--- >
|
| 38 |
except Exception as e:
|
| 39 |
-
print(f"!!! 致命錯誤:
|
| 40 |
-
print("1. Dataset 倉庫是否設為 Public,或 HF_TOKEN 是否有讀取 Private 倉庫的權限。")
|
| 41 |
-
print("2. 倉庫中是否包含 training_data.jsonl 和 sqlite_schema_FULL.json。")
|
| 42 |
print(f"詳細錯誤: {e}")
|
| 43 |
-
# 如果載入失敗,則使用備用數據避免應用崩潰
|
| 44 |
qa_dataset = Dataset.from_dict({"question": ["示例問題"], "sql": ["SELECT 'Dataset failed to load'"]})
|
| 45 |
schema_data = {}
|
| 46 |
|
|
@@ -57,45 +65,35 @@ def load_schema_as_ddl(schema_dict: dict) -> str:
|
|
| 57 |
SCHEMA_DDL = load_schema_as_ddl(schema_data)
|
| 58 |
|
| 59 |
print("--- [3/5] 正在載入句向量模型 (all-MiniLM-L6-v2)... ---")
|
| 60 |
-
# 輕量級句向量模型,在 CPU 上運行極快
|
| 61 |
embedder = SentenceTransformer('all-MiniLM-L6-v2', device='cpu')
|
| 62 |
|
|
|
|
| 63 |
questions = [item['question'] for item in qa_dataset]
|
|
|
|
|
|
|
| 64 |
print(f"--- [4/5] 正在為 {len(questions)} 個問題計算向量 (這可能需要幾分鐘)... ---")
|
| 65 |
-
# 預先計算所有問題的向量,這是實現快速檢索的關鍵
|
| 66 |
question_embeddings = embedder.encode(questions, convert_to_tensor=True, show_progress_bar=True)
|
| 67 |
-
sql_answers = [item['sql'] for item in qa_dataset]
|
| 68 |
print("--- > 向量計算完成! ---")
|
| 69 |
|
| 70 |
# --- 3. 混合系統核心邏輯 ---
|
| 71 |
def get_sql_query(user_question: str):
|
|
|
|
| 72 |
if not user_question:
|
| 73 |
return "請輸入您的問題。", "日誌:用戶未輸入問題。"
|
| 74 |
-
|
| 75 |
-
# 1. 向量檢索
|
| 76 |
question_embedding = embedder.encode(user_question, convert_to_tensor=True)
|
| 77 |
hits = util.semantic_search(question_embedding, question_embeddings, top_k=5)
|
| 78 |
-
hits = hits[0]
|
| 79 |
-
|
| 80 |
most_similar_hit = hits[0]
|
| 81 |
similarity_score = most_similar_hit['score']
|
| 82 |
-
|
| 83 |
log_message = f"檢索到最相似問題: '{questions[most_similar_hit['corpus_id']]}' (相似度: {similarity_score:.4f})"
|
| 84 |
-
|
| 85 |
-
# 2. 如果相似度足夠高,直接返回預定義的 SQL
|
| 86 |
if similarity_score > SIMILARITY_THRESHOLD:
|
| 87 |
sql_result = sql_answers[most_similar_hit['corpus_id']]
|
| 88 |
log_message += f"\n相似度 > {SIMILARITY_THRESHOLD},[模式: 直接返回]。"
|
| 89 |
return sql_result, log_message
|
| 90 |
-
|
| 91 |
-
# 3. 否則,檢索幾個相關例子,用 LLM 生成新 SQL
|
| 92 |
log_message += f"\n相似度 < {SIMILARITY_THRESHOLD},[模式: LLM生成]。正在構建 Prompt..."
|
| 93 |
-
|
| 94 |
-
# 構建 Prompt
|
| 95 |
examples_context = ""
|
| 96 |
-
for hit in hits[:3]:
|
| 97 |
examples_context += f"### A user asks: {questions[hit['corpus_id']]}\n{sql_answers[hit['corpus_id']]}\n\n"
|
| 98 |
-
|
| 99 |
prompt = f"""### Task
|
| 100 |
Generate a SQLite SQL query that answers the following user question.
|
| 101 |
Your response must contain ONLY the SQL query. Do not add any explanation.
|
|
@@ -109,27 +107,19 @@ Your response must contain ONLY the SQL query. Do not add any explanation.
|
|
| 109 |
|
| 110 |
### SQL Query
|
| 111 |
"""
|
| 112 |
-
|
| 113 |
-
# 調用 Hugging Face Inference API
|
| 114 |
log_message += "\n正在請求雲端 LLM..."
|
| 115 |
headers = {"Authorization": f"Bearer {HF_TOKEN}"}
|
| 116 |
payload = {"inputs": prompt, "parameters": {"max_new_tokens": 512, "temperature": 0.1, "return_full_text": False}}
|
| 117 |
response_text = ""
|
| 118 |
-
|
| 119 |
try:
|
| 120 |
response = requests.post(LLM_API_URL, headers=headers, json=payload)
|
| 121 |
-
response_text = response.text
|
| 122 |
response.raise_for_status()
|
| 123 |
-
|
| 124 |
generated_text = response.json()[0]['generated_text'].strip()
|
| 125 |
-
|
| 126 |
-
# 清理常見的返回格式問題
|
| 127 |
if "```sql" in generated_text:
|
| 128 |
generated_text = generated_text.split("```sql")[1].split("```").strip()
|
| 129 |
if "```" in generated_text:
|
| 130 |
generated_text = generated_text.replace("```", "").strip()
|
| 131 |
-
|
| 132 |
-
|
| 133 |
log_message += f"\nLLM 生成成功!"
|
| 134 |
return generated_text, log_message
|
| 135 |
except Exception as e:
|
|
@@ -140,22 +130,19 @@ Your response must contain ONLY the SQL query. Do not add any explanation.
|
|
| 140 |
# --- 4. 創建 Gradio Web 界面 ---
|
| 141 |
print("--- [5/5] 正在創建 Gradio Web 界面... ---")
|
| 142 |
with gr.Blocks(theme=gr.themes.Soft()) as demo:
|
|
|
|
| 143 |
gr.Markdown("# 智能 Text-to-SQL 系統 (混合模式)")
|
| 144 |
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。")
|
| 145 |
-
|
| 146 |
with gr.Row():
|
| 147 |
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)
|
| 148 |
submit_button = gr.Button("生成SQL", variant="primary", scale=1)
|
| 149 |
-
|
| 150 |
sql_output = gr.Code(label="生成的 SQL 查詢", language="sql")
|
| 151 |
log_output = gr.Textbox(label="系統日誌 (執行過程)", lines=4, interactive=False)
|
| 152 |
-
|
| 153 |
submit_button.click(
|
| 154 |
fn=get_sql_query,
|
| 155 |
inputs=question_input,
|
| 156 |
outputs=[sql_output, log_output]
|
| 157 |
)
|
| 158 |
-
|
| 159 |
gr.Examples(
|
| 160 |
examples=[
|
| 161 |
"2024 最好的5個客人以及業績",
|
|
|
|
| 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
|
|
|
|
| 11 |
from huggingface_hub import hf_hub_download
|
| 12 |
|
| 13 |
# --- 配置區 ---
|
|
|
|
| 14 |
HF_TOKEN = os.environ.get("HF_TOKEN")
|
|
|
|
| 15 |
DATASET_REPO_ID = "Paul720810/Text-to-SQL-Softline"
|
|
|
|
| 16 |
LLM_API_URL = "https://api-inference.huggingface.co/models/codellama/CodeLlama-34b-Instruct-hf"
|
|
|
|
|
|
|
| 17 |
SIMILARITY_THRESHOLD = 0.90
|
| 18 |
|
| 19 |
print("--- [1/5] 開始初始化應用 ---")
|
|
|
|
| 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 |
+
# 載入問答範例, 移除已過時的 trust_remote_code 參數
|
| 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"
|
|
|
|
| 45 |
with open(schema_file_path, 'r', encoding='utf-8') as f:
|
| 46 |
schema_data = json.load(f)
|
| 47 |
|
| 48 |
+
print(f"--- > 成功解析 {len(qa_dataset)} 條問答範例和 Schema。 ---")
|
| 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 |
|
|
|
|
| 65 |
SCHEMA_DDL = load_schema_as_ddl(schema_data)
|
| 66 |
|
| 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 |
print(f"--- [4/5] 正在為 {len(questions)} 個問題計算向量 (這可能需要幾分鐘)... ---")
|
|
|
|
| 75 |
question_embeddings = embedder.encode(questions, convert_to_tensor=True, show_progress_bar=True)
|
|
|
|
| 76 |
print("--- > 向量計算完成! ---")
|
| 77 |
|
| 78 |
# --- 3. 混合系統核心邏輯 ---
|
| 79 |
def get_sql_query(user_question: str):
|
| 80 |
+
# (此函式剩餘部分無需修改,保持原樣)
|
| 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 |
+
hits = hits[0]
|
|
|
|
| 86 |
most_similar_hit = hits[0]
|
| 87 |
similarity_score = most_similar_hit['score']
|
|
|
|
| 88 |
log_message = f"檢索到最相似問題: '{questions[most_similar_hit['corpus_id']]}' (相似度: {similarity_score:.4f})"
|
|
|
|
|
|
|
| 89 |
if similarity_score > SIMILARITY_THRESHOLD:
|
| 90 |
sql_result = sql_answers[most_similar_hit['corpus_id']]
|
| 91 |
log_message += f"\n相似度 > {SIMILARITY_THRESHOLD},[模式: 直接返回]。"
|
| 92 |
return sql_result, log_message
|
|
|
|
|
|
|
| 93 |
log_message += f"\n相似度 < {SIMILARITY_THRESHOLD},[模式: LLM生成]。正在構建 Prompt..."
|
|
|
|
|
|
|
| 94 |
examples_context = ""
|
| 95 |
+
for hit in hits[:3]:
|
| 96 |
examples_context += f"### A user asks: {questions[hit['corpus_id']]}\n{sql_answers[hit['corpus_id']]}\n\n"
|
|
|
|
| 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.
|
|
|
|
| 107 |
|
| 108 |
### SQL Query
|
| 109 |
"""
|
|
|
|
|
|
|
| 110 |
log_message += "\n正在請求雲端 LLM..."
|
| 111 |
headers = {"Authorization": f"Bearer {HF_TOKEN}"}
|
| 112 |
payload = {"inputs": prompt, "parameters": {"max_new_tokens": 512, "temperature": 0.1, "return_full_text": False}}
|
| 113 |
response_text = ""
|
|
|
|
| 114 |
try:
|
| 115 |
response = requests.post(LLM_API_URL, headers=headers, json=payload)
|
| 116 |
+
response_text = response.text
|
| 117 |
response.raise_for_status()
|
|
|
|
| 118 |
generated_text = response.json()[0]['generated_text'].strip()
|
|
|
|
|
|
|
| 119 |
if "```sql" in generated_text:
|
| 120 |
generated_text = generated_text.split("```sql")[1].split("```").strip()
|
| 121 |
if "```" in generated_text:
|
| 122 |
generated_text = generated_text.replace("```", "").strip()
|
|
|
|
|
|
|
| 123 |
log_message += f"\nLLM 生成成功!"
|
| 124 |
return generated_text, log_message
|
| 125 |
except Exception as e:
|
|
|
|
| 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 |
+
# (此部分無需修改,保持原樣)
|
| 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)
|
| 138 |
submit_button = gr.Button("生成SQL", variant="primary", scale=1)
|
|
|
|
| 139 |
sql_output = gr.Code(label="生成的 SQL 查詢", language="sql")
|
| 140 |
log_output = gr.Textbox(label="系統日誌 (執行過程)", lines=4, interactive=False)
|
|
|
|
| 141 |
submit_button.click(
|
| 142 |
fn=get_sql_query,
|
| 143 |
inputs=question_input,
|
| 144 |
outputs=[sql_output, log_output]
|
| 145 |
)
|
|
|
|
| 146 |
gr.Examples(
|
| 147 |
examples=[
|
| 148 |
"2024 最好的5個客人以及業績",
|