Spaces:
Sleeping
Sleeping
Update app.py
Browse files
app.py
CHANGED
|
@@ -2,10 +2,11 @@ import gradio as gr
|
|
| 2 |
import os
|
| 3 |
import json
|
| 4 |
import time
|
|
|
|
| 5 |
from datetime import datetime
|
| 6 |
import numpy as np
|
| 7 |
-
from datasets import load_dataset
|
| 8 |
-
from huggingface_hub import HfApi
|
| 9 |
import torch
|
| 10 |
from transformers import AutoTokenizer, AutoModel
|
| 11 |
from openai import OpenAI
|
|
@@ -14,22 +15,34 @@ import io
|
|
| 14 |
from typing import List, Dict, Any, Optional, Tuple
|
| 15 |
from dataclasses import dataclass, field
|
| 16 |
from enum import Enum
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 17 |
|
| 18 |
# ==========================================
|
| 19 |
# 環境變數設定
|
| 20 |
# ==========================================
|
| 21 |
HF_TOKEN = os.environ.get("HF_TOKEN", "")
|
| 22 |
OPENAI_API_KEY = os.environ.get("OPENAI_API_KEY", "")
|
|
|
|
| 23 |
|
| 24 |
# 資料集配置
|
| 25 |
DATASET_NAME = "s880453/interview-transcripts-vectorized"
|
|
|
|
| 26 |
EMBEDDING_MODEL = "intfloat/multilingual-e5-large"
|
| 27 |
|
| 28 |
# 採訪者名單(需要排除)
|
| 29 |
INTERVIEWERS = ["徐美苓", "許弘諺", "郭禹彤"]
|
| 30 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 31 |
# ==========================================
|
| 32 |
-
#
|
| 33 |
# ==========================================
|
| 34 |
@dataclass
|
| 35 |
class SearchResult:
|
|
@@ -44,23 +57,15 @@ class SearchResult:
|
|
| 44 |
relevance_reasoning: str = ""
|
| 45 |
|
| 46 |
@dataclass
|
| 47 |
-
class
|
| 48 |
-
"""
|
| 49 |
-
|
| 50 |
-
|
| 51 |
-
|
| 52 |
-
|
| 53 |
-
|
| 54 |
-
|
| 55 |
-
|
| 56 |
-
class QuestionAnswerPair:
|
| 57 |
-
"""問答對結構"""
|
| 58 |
-
question: str
|
| 59 |
-
answers: List[str]
|
| 60 |
-
raw_contexts: List[str] # 原始 RAG 內容
|
| 61 |
-
relevant_turn_indexes: List[int]
|
| 62 |
-
confidence_scores: List[float]
|
| 63 |
-
search_results: List[SearchResult]
|
| 64 |
|
| 65 |
# ==========================================
|
| 66 |
# 全域變數
|
|
@@ -70,15 +75,18 @@ embeddings = None
|
|
| 70 |
tokenizer = None
|
| 71 |
model = None
|
| 72 |
openai_client = None
|
|
|
|
| 73 |
all_speakers = []
|
| 74 |
init_success = False
|
|
|
|
|
|
|
| 75 |
|
| 76 |
# ==========================================
|
| 77 |
# 初始化函數
|
| 78 |
# ==========================================
|
| 79 |
def initialize_system():
|
| 80 |
"""初始化系統"""
|
| 81 |
-
global dataset, embeddings, tokenizer, model, openai_client, all_speakers, init_success
|
| 82 |
|
| 83 |
try:
|
| 84 |
print("🔄 正在初始化系統...")
|
|
@@ -91,6 +99,22 @@ def initialize_system():
|
|
| 91 |
print("⚠️ 未設定 OpenAI API Key")
|
| 92 |
return False, "請設定 OPENAI_API_KEY"
|
| 93 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 94 |
# 載入資料集
|
| 95 |
print(f"📊 正在載入資料集: {DATASET_NAME}")
|
| 96 |
dataset = load_dataset(DATASET_NAME, split="train", token=HF_TOKEN)
|
|
@@ -126,7 +150,60 @@ def initialize_system():
|
|
| 126 |
return False, error_msg
|
| 127 |
|
| 128 |
# ==========================================
|
| 129 |
-
#
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 130 |
# ==========================================
|
| 131 |
def average_pool(last_hidden_states, attention_mask):
|
| 132 |
"""Average pooling for embeddings"""
|
|
@@ -134,12 +211,10 @@ def average_pool(last_hidden_states, attention_mask):
|
|
| 134 |
return last_hidden.sum(dim=1) / attention_mask.sum(dim=1)[..., None]
|
| 135 |
|
| 136 |
def generate_query_embedding(query_text):
|
| 137 |
-
"""生成查詢向量
|
| 138 |
try:
|
| 139 |
-
# 添加查詢前綴(按照 multilingual-e5-large 的要求)
|
| 140 |
query_with_prefix = f"query: {query_text}"
|
| 141 |
|
| 142 |
-
# Tokenize
|
| 143 |
inputs = tokenizer(
|
| 144 |
[query_with_prefix],
|
| 145 |
max_length=512,
|
|
@@ -148,11 +223,9 @@ def generate_query_embedding(query_text):
|
|
| 148 |
return_tensors='pt'
|
| 149 |
)
|
| 150 |
|
| 151 |
-
# 生成嵌入
|
| 152 |
with torch.no_grad():
|
| 153 |
outputs = model(**inputs)
|
| 154 |
query_embedding = average_pool(outputs.last_hidden_state, inputs['attention_mask'])
|
| 155 |
-
# L2 正規化
|
| 156 |
query_embedding = torch.nn.functional.normalize(query_embedding, p=2, dim=1)
|
| 157 |
|
| 158 |
return query_embedding.cpu().numpy()[0]
|
|
@@ -169,35 +242,14 @@ def build_reranking_prompt(query: str, search_results: List[Dict]) -> str:
|
|
| 169 |
你將收到一個查詢和幾個檢索到的訪談片段。你的任務是根據片段與查詢的相關性來評估和評分每個片段。
|
| 170 |
|
| 171 |
評分指南:
|
| 172 |
-
1.
|
| 173 |
-
|
| 174 |
-
|
| 175 |
-
|
| 176 |
-
|
| 177 |
-
- 0.7 = 相關
|
| 178 |
-
- 0.9 = 高度相關
|
| 179 |
-
- 1 = 完全相關
|
| 180 |
-
|
| 181 |
-
特別注意:
|
| 182 |
-
- 必須排除採訪者(徐美苓、許弘諺、郭禹彤)的一般回覆
|
| 183 |
-
- 檢查上下文相關性(turn_index前後範圍)
|
| 184 |
-
- 評估多重主題匹配的可能性
|
| 185 |
|
| 186 |
-
請為每個搜尋結果提供JSON
|
| 187 |
-
{
|
| 188 |
-
"results": [
|
| 189 |
-
{
|
| 190 |
-
"index": 0,
|
| 191 |
-
"reasoning": "分析原因",
|
| 192 |
-
"speaker_relevance": "發言人相關性",
|
| 193 |
-
"content_relevance": "內容相關性",
|
| 194 |
-
"context_relevance": "上下文相關性",
|
| 195 |
-
"relevance_score": 0.8
|
| 196 |
-
}
|
| 197 |
-
]
|
| 198 |
-
}"""
|
| 199 |
|
| 200 |
-
# 構建搜尋結果文本
|
| 201 |
results_text = f"查詢:{query}\n\n檢索結果:\n"
|
| 202 |
for i, result in enumerate(search_results):
|
| 203 |
results_text += f"\n結果 {i}:\n"
|
|
@@ -213,7 +265,7 @@ def intelligent_routing_and_reranking(query: str, selected_speakers: List[str],
|
|
| 213 |
return []
|
| 214 |
|
| 215 |
try:
|
| 216 |
-
# Step 1: 向量檢索
|
| 217 |
query_vector = generate_query_embedding(query)
|
| 218 |
if query_vector is None:
|
| 219 |
return []
|
|
@@ -221,15 +273,14 @@ def intelligent_routing_and_reranking(query: str, selected_speakers: List[str],
|
|
| 221 |
# Step 2: 計算相似度並過濾
|
| 222 |
initial_results = []
|
| 223 |
for i, item in enumerate(dataset):
|
| 224 |
-
#
|
| 225 |
if item['speaker'] in INTERVIEWERS:
|
| 226 |
continue
|
| 227 |
|
| 228 |
-
#
|
| 229 |
-
# 如果有選擇特定受訪者,只使用該受訪者的資料
|
| 230 |
if selected_speakers and len(selected_speakers) > 0:
|
| 231 |
if item['speaker'] not in selected_speakers:
|
| 232 |
-
continue
|
| 233 |
|
| 234 |
# 計算向量相似度
|
| 235 |
item_vector = np.array(item['embedding'])
|
|
@@ -251,10 +302,10 @@ def intelligent_routing_and_reranking(query: str, selected_speakers: List[str],
|
|
| 251 |
if not candidates:
|
| 252 |
return []
|
| 253 |
|
| 254 |
-
# Step 3: LLM
|
| 255 |
-
rerank_prompt = build_reranking_prompt(query, candidates[:10]) # 只重排序前10個
|
| 256 |
-
|
| 257 |
try:
|
|
|
|
|
|
|
| 258 |
response = openai_client.chat.completions.create(
|
| 259 |
model="gpt-4o-mini",
|
| 260 |
messages=[
|
|
@@ -267,19 +318,17 @@ def intelligent_routing_and_reranking(query: str, selected_speakers: List[str],
|
|
| 267 |
|
| 268 |
rerank_results = json.loads(response.choices[0].message.content)
|
| 269 |
|
| 270 |
-
# Step 4: 加權計分
|
| 271 |
final_results = []
|
| 272 |
for i, candidate in enumerate(candidates[:10]):
|
| 273 |
-
llm_score = 0.5
|
| 274 |
|
| 275 |
-
# 從 LLM 結果中找到對應的分數
|
| 276 |
if 'results' in rerank_results:
|
| 277 |
for r in rerank_results['results']:
|
| 278 |
if r.get('index') == i:
|
| 279 |
llm_score = r.get('relevance_score', 0.5)
|
| 280 |
break
|
| 281 |
|
| 282 |
-
# 計算加權分數
|
| 283 |
weighted_score = 0.3 * candidate['vector_score'] + 0.7 * llm_score
|
| 284 |
|
| 285 |
final_results.append(SearchResult(
|
|
@@ -292,7 +341,7 @@ def intelligent_routing_and_reranking(query: str, selected_speakers: List[str],
|
|
| 292 |
weighted_score=weighted_score
|
| 293 |
))
|
| 294 |
|
| 295 |
-
#
|
| 296 |
for candidate in candidates[10:]:
|
| 297 |
final_results.append(SearchResult(
|
| 298 |
text=candidate['text'],
|
|
@@ -307,14 +356,14 @@ def intelligent_routing_and_reranking(query: str, selected_speakers: List[str],
|
|
| 307 |
# 按加權分數排序
|
| 308 |
final_results.sort(key=lambda x: x.weighted_score, reverse=True)
|
| 309 |
|
| 310 |
-
#
|
| 311 |
-
|
| 312 |
|
| 313 |
-
return
|
| 314 |
|
| 315 |
except Exception as e:
|
| 316 |
-
print(f"LLM
|
| 317 |
-
#
|
| 318 |
return [SearchResult(
|
| 319 |
text=c['text'],
|
| 320 |
speaker=c['speaker'],
|
|
@@ -323,110 +372,274 @@ def intelligent_routing_and_reranking(query: str, selected_speakers: List[str],
|
|
| 323 |
vector_score=c['vector_score'],
|
| 324 |
llm_score=0.0,
|
| 325 |
weighted_score=c['vector_score']
|
| 326 |
-
) for c in candidates[
|
| 327 |
|
| 328 |
except Exception as e:
|
| 329 |
print(f"智慧路由失敗: {str(e)}")
|
| 330 |
return []
|
| 331 |
|
| 332 |
-
|
| 333 |
-
|
| 334 |
-
|
| 335 |
-
|
| 336 |
-
|
| 337 |
-
|
| 338 |
-
|
| 339 |
-
|
| 340 |
-
|
| 341 |
-
|
| 342 |
-
|
| 343 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 344 |
|
| 345 |
-
|
| 346 |
-
|
|
|
|
|
|
|
|
|
|
| 347 |
|
| 348 |
-
|
| 349 |
-
|
| 350 |
-
|
| 351 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 352 |
|
| 353 |
-
|
|
|
|
|
|
|
| 354 |
|
| 355 |
# ==========================================
|
| 356 |
-
# RAG
|
| 357 |
# ==========================================
|
| 358 |
def rag_chat(question, selected_speakers, history):
|
| 359 |
-
"""RAG 對話處理
|
| 360 |
if not init_success:
|
| 361 |
return history + [[question, "系統尚未初始化,請稍後再試。"]]
|
| 362 |
|
| 363 |
try:
|
| 364 |
# 執行智慧路由與重排序
|
| 365 |
-
search_results = intelligent_routing_and_reranking(question, selected_speakers, top_k=
|
|
|
|
|
|
|
|
|
|
| 366 |
|
| 367 |
-
if not
|
| 368 |
return history + [[question, "未找到相關內容,請嘗試其他問題。"]]
|
| 369 |
|
| 370 |
-
#
|
| 371 |
context = "相關訪談內容:\n\n"
|
| 372 |
raw_contexts = []
|
| 373 |
|
| 374 |
-
for i, result in enumerate(
|
| 375 |
-
context += f"[片段 {i}]\n"
|
| 376 |
context += f"發言人:{result.speaker}\n"
|
| 377 |
context += f"內容:{result.text}\n"
|
| 378 |
-
context += f"
|
| 379 |
|
| 380 |
-
# 保存原始內容 - 確保文本存在
|
| 381 |
if result.text:
|
| 382 |
raw_context_text = f"[{result.speaker} - Turn {result.turn_index}]: {result.text}"
|
| 383 |
raw_contexts.append(raw_context_text)
|
| 384 |
|
| 385 |
-
# 確保有原始內容
|
| 386 |
if not raw_contexts:
|
| 387 |
raw_contexts = ["未能提取原始內容"]
|
| 388 |
|
| 389 |
-
# 構建 GPT prompt
|
| 390 |
prompt = f"""基於以下訪談內容回答問題。請提供準確、完整的回答。
|
| 391 |
|
| 392 |
{context}
|
| 393 |
|
| 394 |
-
問題:{question}
|
| 395 |
-
|
| 396 |
-
要求:
|
| 397 |
-
1. 基於提供的訪談內容回答
|
| 398 |
-
2. 引用具體的發言人和內容
|
| 399 |
-
3. 如果內容不足以回答,請明確說明"""
|
| 400 |
-
|
| 401 |
-
# 調用 GPT(每次獨立調用)
|
| 402 |
-
response = openai_client.chat.completions.create(
|
| 403 |
-
model="gpt-4o-mini",
|
| 404 |
-
messages=[
|
| 405 |
-
{"role": "system", "content": "你是一個專業的訪談分析助手。只基於提供的內容回答,不要添加額外信息。"},
|
| 406 |
-
{"role": "user", "content": prompt}
|
| 407 |
-
],
|
| 408 |
-
temperature=0.1
|
| 409 |
-
)
|
| 410 |
|
| 411 |
-
|
|
|
|
| 412 |
|
| 413 |
-
# 添加原始 RAG
|
| 414 |
-
answer_with_sources = f"{answer}\n\n---\n📚 **原始 RAG
|
| 415 |
-
for i, raw_context in enumerate(raw_contexts
|
| 416 |
-
# 確保 raw_context 是字串且有內容
|
| 417 |
if raw_context and raw_context != "未能提取原始內容":
|
| 418 |
-
|
| 419 |
-
answer_with_sources += f"\n**來源 {i}:**\n{raw_context}\n"
|
| 420 |
else:
|
| 421 |
-
answer_with_sources += f"\n**來源 {i}:** 無內容\n"
|
| 422 |
|
| 423 |
return history + [[question, answer_with_sources]]
|
| 424 |
|
| 425 |
except Exception as e:
|
| 426 |
return history + [[question, f"處理過程中發生錯誤:{str(e)}"]]
|
| 427 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 428 |
# ==========================================
|
| 429 |
-
#
|
| 430 |
# ==========================================
|
| 431 |
def parse_word_document(file_path):
|
| 432 |
"""解析 Word 文檔中的問題"""
|
|
@@ -454,20 +667,19 @@ def extract_speaker_from_filename(filename, available_speakers):
|
|
| 454 |
base_name = os.path.basename(filename)
|
| 455 |
base_name_no_ext = os.path.splitext(base_name)[0]
|
| 456 |
|
| 457 |
-
# 檢查檔名中是否包含任何受訪者名稱
|
| 458 |
for speaker in available_speakers:
|
| 459 |
if speaker in base_name_no_ext:
|
| 460 |
return [speaker]
|
| 461 |
|
| 462 |
-
return None
|
| 463 |
|
| 464 |
-
def single_interviewee_guide_filling(file_path, selected_speakers, file_name=None):
|
| 465 |
-
"""單一受訪者訪綱填答 -
|
| 466 |
if not init_success:
|
| 467 |
return None, "系統尚未初始化"
|
| 468 |
|
| 469 |
try:
|
| 470 |
-
#
|
| 471 |
if file_name:
|
| 472 |
detected_speakers = extract_speaker_from_filename(file_name, all_speakers)
|
| 473 |
if detected_speakers:
|
|
@@ -480,90 +692,47 @@ def single_interviewee_guide_filling(file_path, selected_speakers, file_name=Non
|
|
| 480 |
if not questions:
|
| 481 |
return None, "未能從文檔中提取問題"
|
| 482 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 483 |
# 創建新的 Word 文檔
|
| 484 |
output_doc = Document()
|
| 485 |
output_doc.add_heading('訪談訪綱 - AI 智慧填答', 0)
|
| 486 |
output_doc.add_paragraph(f'處理時間:{datetime.now().strftime("%Y-%m-%d %H:%M:%S")}')
|
| 487 |
output_doc.add_paragraph(f'原始檔案:{file_name if file_name else "未知"}')
|
| 488 |
output_doc.add_paragraph(f'選擇的受訪者:{", ".join(selected_speakers) if selected_speakers else "全部"}')
|
| 489 |
-
output_doc.add_paragraph(f'
|
|
|
|
| 490 |
output_doc.add_paragraph('')
|
| 491 |
|
| 492 |
-
#
|
| 493 |
-
for i,
|
| 494 |
output_doc.add_heading(f'問題 {i}', level=1)
|
| 495 |
-
output_doc.add_paragraph(question)
|
| 496 |
-
|
| 497 |
-
# 使用智慧路由與重排序檢索
|
| 498 |
-
search_results = intelligent_routing_and_reranking(question, selected_speakers, top_k=15)
|
| 499 |
-
|
| 500 |
-
# 過濾掉低相關性的結果(相關性分數低於 0.4 的視為雜訊)
|
| 501 |
-
filtered_results = [r for r in search_results if r.weighted_score >= 0.4]
|
| 502 |
|
| 503 |
-
if
|
| 504 |
-
#
|
| 505 |
-
context = ""
|
| 506 |
-
raw_contexts = []
|
| 507 |
-
|
| 508 |
-
for j, result in enumerate(filtered_results[:5]):
|
| 509 |
-
context += f"[片段 {j+1}]\n"
|
| 510 |
-
context += f"發言人:{result.speaker}\n"
|
| 511 |
-
context += f"內容:{result.text}\n"
|
| 512 |
-
context += f"相關性:向量={result.vector_score:.3f}, LLM={result.llm_score:.3f}\n\n"
|
| 513 |
-
|
| 514 |
-
raw_contexts.append({
|
| 515 |
-
'speaker': result.speaker,
|
| 516 |
-
'text': result.text,
|
| 517 |
-
'turn_index': result.turn_index,
|
| 518 |
-
'score': result.weighted_score
|
| 519 |
-
})
|
| 520 |
-
|
| 521 |
-
# 獨立調用 GPT 生成回答
|
| 522 |
-
prompt = f"""基於以下訪談內容回答訪綱問題:
|
| 523 |
-
|
| 524 |
-
{context}
|
| 525 |
-
|
| 526 |
-
問題:{question}
|
| 527 |
-
|
| 528 |
-
請提供:
|
| 529 |
-
1. 主要回答
|
| 530 |
-
2. 不同受訪者的觀點(如果有)
|
| 531 |
-
3. 具體引述"""
|
| 532 |
-
|
| 533 |
-
response = openai_client.chat.completions.create(
|
| 534 |
-
model="gpt-4o-mini",
|
| 535 |
-
messages=[
|
| 536 |
-
{"role": "system", "content": "你是訪談分析專家。基於提供的內容準確回答。"},
|
| 537 |
-
{"role": "user", "content": prompt}
|
| 538 |
-
],
|
| 539 |
-
temperature=0.1
|
| 540 |
-
)
|
| 541 |
-
|
| 542 |
-
answer = response.choices[0].message.content
|
| 543 |
-
|
| 544 |
-
# 添加 AI 回答
|
| 545 |
output_doc.add_heading('AI 分析回答:', level=2)
|
| 546 |
-
for line in answer.split('\n'):
|
| 547 |
if line.strip():
|
| 548 |
output_doc.add_paragraph(line)
|
| 549 |
|
| 550 |
-
#
|
| 551 |
-
output_doc.add_heading('原始 RAG
|
| 552 |
-
for j, raw in enumerate(raw_contexts
|
| 553 |
p = output_doc.add_paragraph()
|
| 554 |
-
p.add_run(f"{j}. [{raw['speaker']} - Turn {raw['turn_index']}] ").bold = True
|
| 555 |
p.add_run(f"(相關性: {raw['score']:.3f})\n")
|
| 556 |
-
#
|
| 557 |
-
p.add_run(raw['text'])
|
| 558 |
else:
|
| 559 |
-
output_doc.add_paragraph(
|
| 560 |
|
| 561 |
-
output_doc.add_page_break()
|
| 562 |
-
|
| 563 |
-
# 添加未使用內容分析(如果是單一受訪者)
|
| 564 |
-
if len(selected_speakers) == 1:
|
| 565 |
-
output_doc.add_heading('補充:可能相關但未被問及的內容', level=1)
|
| 566 |
-
# 這裡可以加入額外的分析邏輯
|
| 567 |
|
| 568 |
# 保存文檔
|
| 569 |
output_buffer = io.BytesIO()
|
|
@@ -577,64 +746,146 @@ def single_interviewee_guide_filling(file_path, selected_speakers, file_name=Non
|
|
| 577 |
with open(output_filename, 'wb') as f:
|
| 578 |
f.write(output_buffer.getvalue())
|
| 579 |
|
| 580 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 581 |
|
| 582 |
except Exception as e:
|
| 583 |
return None, f"處理失敗:{str(e)}"
|
| 584 |
|
| 585 |
-
def batch_process_guides(files, default_speakers):
|
| 586 |
-
"""
|
| 587 |
if not init_success:
|
| 588 |
return [], "系統尚未初始化"
|
| 589 |
|
| 590 |
results = []
|
| 591 |
processed_files = []
|
|
|
|
| 592 |
|
| 593 |
try:
|
| 594 |
-
total_files = len(files)
|
| 595 |
print(f"開始批量處理 {total_files} 個檔案")
|
| 596 |
|
| 597 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 598 |
try:
|
| 599 |
file_name = file.name if hasattr(file, 'name') else str(file)
|
| 600 |
-
print(f"\n處理檔案 {idx}/{total_files}: {file_name}")
|
| 601 |
|
| 602 |
-
#
|
| 603 |
detected_speakers = extract_speaker_from_filename(file_name, all_speakers)
|
| 604 |
|
| 605 |
if detected_speakers:
|
| 606 |
speakers_to_use = detected_speakers
|
| 607 |
-
status_msg = f"檔案 {idx}:
|
| 608 |
else:
|
| 609 |
speakers_to_use = default_speakers
|
| 610 |
-
status_msg = f"檔案 {idx}: 使用預設受訪者"
|
| 611 |
|
| 612 |
print(status_msg)
|
| 613 |
-
results.append(status_msg)
|
| 614 |
|
| 615 |
-
#
|
| 616 |
output_file, process_status = single_interviewee_guide_filling(
|
| 617 |
file.name if hasattr(file, 'name') else file,
|
| 618 |
speakers_to_use,
|
| 619 |
-
file_name
|
|
|
|
| 620 |
)
|
| 621 |
|
| 622 |
if output_file:
|
| 623 |
-
|
| 624 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 625 |
else:
|
| 626 |
-
|
| 627 |
-
|
| 628 |
-
|
| 629 |
-
|
| 630 |
-
|
|
|
|
| 631 |
except Exception as e:
|
| 632 |
-
|
| 633 |
-
|
| 634 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 635 |
|
| 636 |
# 彙總結果
|
| 637 |
-
summary = f"\n處理完成!\n成功: {
|
|
|
|
|
|
|
| 638 |
summary += "\n詳細結果:\n" + "\n".join(results)
|
| 639 |
|
| 640 |
return processed_files, summary
|
|
@@ -655,22 +906,23 @@ def create_interface():
|
|
| 655 |
.gradio-container {
|
| 656 |
font-family: 'Microsoft JhengHei', sans-serif;
|
| 657 |
}
|
| 658 |
-
.
|
| 659 |
-
|
| 660 |
}
|
| 661 |
"""
|
| 662 |
) as app:
|
| 663 |
# 標題
|
| 664 |
gr.Markdown("""
|
| 665 |
-
# 🎙️ 訪談轉錄稿智慧分析系統
|
| 666 |
|
| 667 |
-
**技術架構:** Multilingual-E5-Large + GPT-4o-mini + 冠軍級 RAG
|
| 668 |
|
| 669 |
**核心功能:**
|
| 670 |
-
- 🔍
|
| 671 |
-
- 💬 AI
|
| 672 |
-
- 📝
|
| 673 |
-
- 📊
|
|
|
|
| 674 |
""")
|
| 675 |
|
| 676 |
# 系統狀態
|
|
@@ -686,7 +938,7 @@ def create_interface():
|
|
| 686 |
with gr.Tabs():
|
| 687 |
# Tab 1: AI 對話
|
| 688 |
with gr.Tab("💬 AI 對話"):
|
| 689 |
-
gr.Markdown("###
|
| 690 |
|
| 691 |
with gr.Row():
|
| 692 |
with gr.Column(scale=1):
|
|
@@ -714,7 +966,17 @@ def create_interface():
|
|
| 714 |
|
| 715 |
with gr.Row():
|
| 716 |
clear_btn = gr.Button("清除對話")
|
| 717 |
-
gr.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 718 |
|
| 719 |
# Tab 2: 訪綱填答
|
| 720 |
with gr.Tab("📝 訪綱填答"):
|
|
@@ -722,15 +984,11 @@ def create_interface():
|
|
| 722 |
### 智慧訪綱填答系統
|
| 723 |
|
| 724 |
**特色功能:**
|
| 725 |
-
-
|
| 726 |
-
-
|
| 727 |
-
-
|
| 728 |
-
-
|
| 729 |
-
-
|
| 730 |
-
|
| 731 |
-
**檔名識別規則:**
|
| 732 |
-
- 如果檔名包含受訪者姓名(如:`訪綱_陳美玲.docx`),系統會自動使用該受訪者的資料
|
| 733 |
-
- 如果檔名未包含受訪者姓名,則使用您勾選的受訪者
|
| 734 |
""")
|
| 735 |
|
| 736 |
with gr.Row():
|
|
@@ -741,7 +999,7 @@ def create_interface():
|
|
| 741 |
info="檔名中有受訪者名稱時會自動覆蓋此選擇"
|
| 742 |
)
|
| 743 |
|
| 744 |
-
#
|
| 745 |
with gr.Accordion("單檔處理", open=False):
|
| 746 |
single_file_input = gr.File(
|
| 747 |
label="上傳單個訪綱 (Word 格式)",
|
|
@@ -749,29 +1007,38 @@ def create_interface():
|
|
| 749 |
)
|
| 750 |
single_process_btn = gr.Button("處理單檔", variant="secondary")
|
| 751 |
|
| 752 |
-
#
|
| 753 |
batch_file_input = gr.File(
|
| 754 |
label="批量上傳訪綱(最多15個 Word 檔案)",
|
| 755 |
file_types=[".docx"],
|
| 756 |
file_count="multiple"
|
| 757 |
)
|
| 758 |
|
| 759 |
-
batch_process_btn = gr.Button("🚀
|
| 760 |
|
| 761 |
with gr.Column():
|
|
|
|
|
|
|
|
|
|
| 762 |
process_status = gr.Textbox(
|
| 763 |
label="處理狀態",
|
| 764 |
interactive=False,
|
| 765 |
lines=10
|
| 766 |
)
|
| 767 |
|
| 768 |
-
#
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 769 |
single_download_file = gr.File(
|
| 770 |
label="下載單檔結果",
|
| 771 |
visible=False
|
| 772 |
)
|
| 773 |
|
| 774 |
-
# 批量下載
|
| 775 |
batch_download_files = gr.File(
|
| 776 |
label="下載所有結果",
|
| 777 |
visible=False,
|
|
@@ -781,23 +1048,32 @@ def create_interface():
|
|
| 781 |
# 技術細節
|
| 782 |
with gr.Accordion("🔧 技術細節", open=False):
|
| 783 |
gr.Markdown("""
|
| 784 |
-
###
|
| 785 |
-
|
| 786 |
-
**1.
|
| 787 |
-
-
|
| 788 |
-
-
|
| 789 |
-
-
|
| 790 |
-
|
| 791 |
-
**2.
|
| 792 |
-
-
|
| 793 |
-
-
|
| 794 |
-
-
|
| 795 |
-
|
| 796 |
-
|
| 797 |
-
|
| 798 |
-
-
|
| 799 |
-
-
|
| 800 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 801 |
""")
|
| 802 |
|
| 803 |
# 事件處理
|
|
@@ -810,34 +1086,97 @@ def create_interface():
|
|
| 810 |
def clear_chat():
|
| 811 |
return []
|
| 812 |
|
| 813 |
-
def
|
| 814 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 815 |
if not file:
|
| 816 |
-
return
|
|
|
|
|
|
|
|
|
|
|
|
|
| 817 |
|
| 818 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 819 |
|
| 820 |
if result_file:
|
| 821 |
-
return
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 822 |
else:
|
| 823 |
-
return
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 824 |
|
| 825 |
-
def
|
| 826 |
-
"""
|
| 827 |
if not files:
|
| 828 |
-
return
|
|
|
|
|
|
|
|
|
|
|
|
|
| 829 |
|
| 830 |
-
# 限制最多15個檔案
|
| 831 |
if len(files) > 15:
|
| 832 |
-
return
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 833 |
|
| 834 |
# 批量處理
|
| 835 |
-
processed_files, status = batch_process_guides(files, speakers)
|
| 836 |
|
| 837 |
if processed_files:
|
| 838 |
-
return
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 839 |
else:
|
| 840 |
-
return
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 841 |
|
| 842 |
def update_status():
|
| 843 |
success, message = initialize_system()
|
|
@@ -864,18 +1203,38 @@ def create_interface():
|
|
| 864 |
|
| 865 |
clear_btn.click(clear_chat, outputs=[chatbot])
|
| 866 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 867 |
# 單檔處理
|
| 868 |
single_process_btn.click(
|
| 869 |
-
|
| 870 |
inputs=[single_file_input, guide_speakers],
|
| 871 |
-
outputs=[
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 872 |
)
|
| 873 |
|
| 874 |
# 批量處理
|
| 875 |
batch_process_btn.click(
|
| 876 |
-
|
| 877 |
inputs=[batch_file_input, guide_speakers],
|
| 878 |
-
outputs=[
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 879 |
)
|
| 880 |
|
| 881 |
init_btn.click(
|
|
@@ -888,6 +1247,12 @@ def create_interface():
|
|
| 888 |
update_status,
|
| 889 |
outputs=[status_text, speaker_selector, guide_speakers]
|
| 890 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 891 |
|
| 892 |
return app
|
| 893 |
|
|
@@ -895,6 +1260,32 @@ def create_interface():
|
|
| 895 |
# 主程式入口
|
| 896 |
# ==========================================
|
| 897 |
if __name__ == "__main__":
|
| 898 |
-
#
|
| 899 |
app = create_interface()
|
| 900 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 2 |
import os
|
| 3 |
import json
|
| 4 |
import time
|
| 5 |
+
import asyncio
|
| 6 |
from datetime import datetime
|
| 7 |
import numpy as np
|
| 8 |
+
from datasets import load_dataset, Dataset as HFDataset
|
| 9 |
+
from huggingface_hub import HfApi, upload_file, create_repo
|
| 10 |
import torch
|
| 11 |
from transformers import AutoTokenizer, AutoModel
|
| 12 |
from openai import OpenAI
|
|
|
|
| 15 |
from typing import List, Dict, Any, Optional, Tuple
|
| 16 |
from dataclasses import dataclass, field
|
| 17 |
from enum import Enum
|
| 18 |
+
from concurrent.futures import ThreadPoolExecutor, as_completed
|
| 19 |
+
import threading
|
| 20 |
+
from tqdm import tqdm
|
| 21 |
+
import traceback
|
| 22 |
+
import queue
|
| 23 |
|
| 24 |
# ==========================================
|
| 25 |
# 環境變數設定
|
| 26 |
# ==========================================
|
| 27 |
HF_TOKEN = os.environ.get("HF_TOKEN", "")
|
| 28 |
OPENAI_API_KEY = os.environ.get("OPENAI_API_KEY", "")
|
| 29 |
+
ACCESS_PASSWORD = os.environ.get("ACCESS_PASSWORD", "netzero2025") # 密碼保護
|
| 30 |
|
| 31 |
# 資料集配置
|
| 32 |
DATASET_NAME = "s880453/interview-transcripts-vectorized"
|
| 33 |
+
OUTPUT_DATASET = "s880453/interview-outputs" # 輸出資料集
|
| 34 |
EMBEDDING_MODEL = "intfloat/multilingual-e5-large"
|
| 35 |
|
| 36 |
# 採訪者名單(需要排除)
|
| 37 |
INTERVIEWERS = ["徐美苓", "許弘諺", "郭禹彤"]
|
| 38 |
|
| 39 |
+
# 並行處理設定
|
| 40 |
+
MAX_WORKERS = 5 # 最多 5 個並行處理
|
| 41 |
+
MAX_RETRIES = 3 # 最多重試 3 次
|
| 42 |
+
RELEVANCE_THRESHOLD = 0.4 # 相關性門檻
|
| 43 |
+
|
| 44 |
# ==========================================
|
| 45 |
+
# 結構化數據模型
|
| 46 |
# ==========================================
|
| 47 |
@dataclass
|
| 48 |
class SearchResult:
|
|
|
|
| 57 |
relevance_reasoning: str = ""
|
| 58 |
|
| 59 |
@dataclass
|
| 60 |
+
class ProcessingStatus:
|
| 61 |
+
"""處理狀態追蹤"""
|
| 62 |
+
total_items: int = 0
|
| 63 |
+
completed_items: int = 0
|
| 64 |
+
failed_items: int = 0
|
| 65 |
+
current_item: str = ""
|
| 66 |
+
start_time: float = 0.0
|
| 67 |
+
estimated_time: float = 0.0
|
| 68 |
+
errors: List[str] = field(default_factory=list)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 69 |
|
| 70 |
# ==========================================
|
| 71 |
# 全域變數
|
|
|
|
| 75 |
tokenizer = None
|
| 76 |
model = None
|
| 77 |
openai_client = None
|
| 78 |
+
hf_api = None
|
| 79 |
all_speakers = []
|
| 80 |
init_success = False
|
| 81 |
+
processing_status = ProcessingStatus()
|
| 82 |
+
status_lock = threading.Lock()
|
| 83 |
|
| 84 |
# ==========================================
|
| 85 |
# 初始化函數
|
| 86 |
# ==========================================
|
| 87 |
def initialize_system():
|
| 88 |
"""初始化系統"""
|
| 89 |
+
global dataset, embeddings, tokenizer, model, openai_client, all_speakers, init_success, hf_api
|
| 90 |
|
| 91 |
try:
|
| 92 |
print("🔄 正在初始化系統...")
|
|
|
|
| 99 |
print("⚠️ 未設定 OpenAI API Key")
|
| 100 |
return False, "請設定 OPENAI_API_KEY"
|
| 101 |
|
| 102 |
+
# 初始化 Hugging Face API
|
| 103 |
+
hf_api = HfApi(token=HF_TOKEN)
|
| 104 |
+
|
| 105 |
+
# 確保輸出資料集存在
|
| 106 |
+
try:
|
| 107 |
+
create_repo(
|
| 108 |
+
repo_id=OUTPUT_DATASET,
|
| 109 |
+
repo_type="dataset",
|
| 110 |
+
private=True,
|
| 111 |
+
exist_ok=True,
|
| 112 |
+
token=HF_TOKEN
|
| 113 |
+
)
|
| 114 |
+
print(f"✅ 輸出資料集 {OUTPUT_DATASET} 已準備")
|
| 115 |
+
except Exception as e:
|
| 116 |
+
print(f"⚠️ 創建輸出資料集時出現問題: {e}")
|
| 117 |
+
|
| 118 |
# 載入資料集
|
| 119 |
print(f"📊 正在載入資料集: {DATASET_NAME}")
|
| 120 |
dataset = load_dataset(DATASET_NAME, split="train", token=HF_TOKEN)
|
|
|
|
| 150 |
return False, error_msg
|
| 151 |
|
| 152 |
# ==========================================
|
| 153 |
+
# HF Dataset 上傳函數
|
| 154 |
+
# ==========================================
|
| 155 |
+
def upload_to_hf_dataset(file_path, file_type="guide", metadata=None):
|
| 156 |
+
"""上傳檔案到 HF Dataset"""
|
| 157 |
+
try:
|
| 158 |
+
timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
|
| 159 |
+
|
| 160 |
+
# 決定路徑
|
| 161 |
+
if file_type == "guide":
|
| 162 |
+
repo_path = f"guide_outputs/{os.path.basename(file_path)}"
|
| 163 |
+
elif file_type == "chat":
|
| 164 |
+
repo_path = f"chat_exports/{os.path.basename(file_path)}"
|
| 165 |
+
else:
|
| 166 |
+
repo_path = f"others/{os.path.basename(file_path)}"
|
| 167 |
+
|
| 168 |
+
# 上傳檔案
|
| 169 |
+
upload_file(
|
| 170 |
+
path_or_fileobj=file_path,
|
| 171 |
+
path_in_repo=repo_path,
|
| 172 |
+
repo_id=OUTPUT_DATASET,
|
| 173 |
+
repo_type="dataset",
|
| 174 |
+
token=HF_TOKEN
|
| 175 |
+
)
|
| 176 |
+
|
| 177 |
+
print(f"✅ 檔案已上傳到 HF Dataset: {repo_path}")
|
| 178 |
+
|
| 179 |
+
# 更新 metadata
|
| 180 |
+
if metadata:
|
| 181 |
+
metadata_path = f"metadata/{timestamp}.json"
|
| 182 |
+
metadata_content = json.dumps(metadata, ensure_ascii=False, indent=2)
|
| 183 |
+
|
| 184 |
+
# 創建臨時檔案並上傳
|
| 185 |
+
with open(f"temp_metadata_{timestamp}.json", "w", encoding="utf-8") as f:
|
| 186 |
+
f.write(metadata_content)
|
| 187 |
+
|
| 188 |
+
upload_file(
|
| 189 |
+
path_or_fileobj=f"temp_metadata_{timestamp}.json",
|
| 190 |
+
path_in_repo=metadata_path,
|
| 191 |
+
repo_id=OUTPUT_DATASET,
|
| 192 |
+
repo_type="dataset",
|
| 193 |
+
token=HF_TOKEN
|
| 194 |
+
)
|
| 195 |
+
|
| 196 |
+
# 刪除臨時檔案
|
| 197 |
+
os.remove(f"temp_metadata_{timestamp}.json")
|
| 198 |
+
|
| 199 |
+
return True, repo_path
|
| 200 |
+
|
| 201 |
+
except Exception as e:
|
| 202 |
+
print(f"❌ 上傳到 HF Dataset 失敗: {str(e)}")
|
| 203 |
+
return False, str(e)
|
| 204 |
+
|
| 205 |
+
# ==========================================
|
| 206 |
+
# 向量搜尋函數
|
| 207 |
# ==========================================
|
| 208 |
def average_pool(last_hidden_states, attention_mask):
|
| 209 |
"""Average pooling for embeddings"""
|
|
|
|
| 211 |
return last_hidden.sum(dim=1) / attention_mask.sum(dim=1)[..., None]
|
| 212 |
|
| 213 |
def generate_query_embedding(query_text):
|
| 214 |
+
"""生成查詢向量"""
|
| 215 |
try:
|
|
|
|
| 216 |
query_with_prefix = f"query: {query_text}"
|
| 217 |
|
|
|
|
| 218 |
inputs = tokenizer(
|
| 219 |
[query_with_prefix],
|
| 220 |
max_length=512,
|
|
|
|
| 223 |
return_tensors='pt'
|
| 224 |
)
|
| 225 |
|
|
|
|
| 226 |
with torch.no_grad():
|
| 227 |
outputs = model(**inputs)
|
| 228 |
query_embedding = average_pool(outputs.last_hidden_state, inputs['attention_mask'])
|
|
|
|
| 229 |
query_embedding = torch.nn.functional.normalize(query_embedding, p=2, dim=1)
|
| 230 |
|
| 231 |
return query_embedding.cpu().numpy()[0]
|
|
|
|
| 242 |
你將收到一個查詢和幾個檢索到的訪談片段。你的任務是根據片段與查詢的相關性來評估和評分每個片段。
|
| 243 |
|
| 244 |
評分指南:
|
| 245 |
+
- 1.0 = 完全相關
|
| 246 |
+
- 0.7-0.9 = 高度相關
|
| 247 |
+
- 0.5-0.7 = 中等相關
|
| 248 |
+
- 0.3-0.5 = 輕微相關
|
| 249 |
+
- 0-0.3 = 幾乎無關
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 250 |
|
| 251 |
+
請為每個搜尋結果提供JSON格式的評分。"""
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 252 |
|
|
|
|
| 253 |
results_text = f"查詢:{query}\n\n檢索結果:\n"
|
| 254 |
for i, result in enumerate(search_results):
|
| 255 |
results_text += f"\n結果 {i}:\n"
|
|
|
|
| 265 |
return []
|
| 266 |
|
| 267 |
try:
|
| 268 |
+
# Step 1: 向量檢索
|
| 269 |
query_vector = generate_query_embedding(query)
|
| 270 |
if query_vector is None:
|
| 271 |
return []
|
|
|
|
| 273 |
# Step 2: 計算相似度並過濾
|
| 274 |
initial_results = []
|
| 275 |
for i, item in enumerate(dataset):
|
| 276 |
+
# 排除採訪者
|
| 277 |
if item['speaker'] in INTERVIEWERS:
|
| 278 |
continue
|
| 279 |
|
| 280 |
+
# 嚴格的受訪者過濾
|
|
|
|
| 281 |
if selected_speakers and len(selected_speakers) > 0:
|
| 282 |
if item['speaker'] not in selected_speakers:
|
| 283 |
+
continue
|
| 284 |
|
| 285 |
# 計算向量相似度
|
| 286 |
item_vector = np.array(item['embedding'])
|
|
|
|
| 302 |
if not candidates:
|
| 303 |
return []
|
| 304 |
|
| 305 |
+
# Step 3: LLM 重排序(只對前10個)
|
|
|
|
|
|
|
| 306 |
try:
|
| 307 |
+
rerank_prompt = build_reranking_prompt(query, candidates[:10])
|
| 308 |
+
|
| 309 |
response = openai_client.chat.completions.create(
|
| 310 |
model="gpt-4o-mini",
|
| 311 |
messages=[
|
|
|
|
| 318 |
|
| 319 |
rerank_results = json.loads(response.choices[0].message.content)
|
| 320 |
|
| 321 |
+
# Step 4: 加權計分
|
| 322 |
final_results = []
|
| 323 |
for i, candidate in enumerate(candidates[:10]):
|
| 324 |
+
llm_score = 0.5
|
| 325 |
|
|
|
|
| 326 |
if 'results' in rerank_results:
|
| 327 |
for r in rerank_results['results']:
|
| 328 |
if r.get('index') == i:
|
| 329 |
llm_score = r.get('relevance_score', 0.5)
|
| 330 |
break
|
| 331 |
|
|
|
|
| 332 |
weighted_score = 0.3 * candidate['vector_score'] + 0.7 * llm_score
|
| 333 |
|
| 334 |
final_results.append(SearchResult(
|
|
|
|
| 341 |
weighted_score=weighted_score
|
| 342 |
))
|
| 343 |
|
| 344 |
+
# 加入剩餘的候選
|
| 345 |
for candidate in candidates[10:]:
|
| 346 |
final_results.append(SearchResult(
|
| 347 |
text=candidate['text'],
|
|
|
|
| 356 |
# 按加權分數排序
|
| 357 |
final_results.sort(key=lambda x: x.weighted_score, reverse=True)
|
| 358 |
|
| 359 |
+
# 只返回高於門檻的結果
|
| 360 |
+
filtered_results = [r for r in final_results if r.weighted_score >= RELEVANCE_THRESHOLD]
|
| 361 |
|
| 362 |
+
return filtered_results
|
| 363 |
|
| 364 |
except Exception as e:
|
| 365 |
+
print(f"LLM 重排序失敗: {str(e)}")
|
| 366 |
+
# 降級處理
|
| 367 |
return [SearchResult(
|
| 368 |
text=c['text'],
|
| 369 |
speaker=c['speaker'],
|
|
|
|
| 372 |
vector_score=c['vector_score'],
|
| 373 |
llm_score=0.0,
|
| 374 |
weighted_score=c['vector_score']
|
| 375 |
+
) for c in candidates if c['vector_score'] >= RELEVANCE_THRESHOLD]
|
| 376 |
|
| 377 |
except Exception as e:
|
| 378 |
print(f"智慧路由失敗: {str(e)}")
|
| 379 |
return []
|
| 380 |
|
| 381 |
+
# ==========================================
|
| 382 |
+
# 並行處理與錯誤重試
|
| 383 |
+
# ==========================================
|
| 384 |
+
def call_gpt_with_retry(prompt, max_retries=MAX_RETRIES):
|
| 385 |
+
"""調用 GPT 並實現錯誤重試機制"""
|
| 386 |
+
for attempt in range(max_retries):
|
| 387 |
+
try:
|
| 388 |
+
response = openai_client.chat.completions.create(
|
| 389 |
+
model="gpt-4o-mini",
|
| 390 |
+
messages=[
|
| 391 |
+
{"role": "system", "content": "你是訪談分析專家。基於提供的內容準確回答。"},
|
| 392 |
+
{"role": "user", "content": prompt}
|
| 393 |
+
],
|
| 394 |
+
temperature=0.1,
|
| 395 |
+
timeout=30 # 30秒超時
|
| 396 |
+
)
|
| 397 |
+
return response.choices[0].message.content
|
| 398 |
+
except Exception as e:
|
| 399 |
+
if attempt < max_retries - 1:
|
| 400 |
+
wait_time = 2 ** attempt # 指數退避
|
| 401 |
+
print(f"API 調用失敗,{wait_time}秒後重試... (嘗試 {attempt + 1}/{max_retries})")
|
| 402 |
+
time.sleep(wait_time)
|
| 403 |
+
else:
|
| 404 |
+
print(f"API 調用最終失敗: {str(e)}")
|
| 405 |
+
return f"處理失敗: {str(e)}"
|
| 406 |
+
|
| 407 |
+
def process_single_question(question, selected_speakers, question_index=0, total_questions=1):
|
| 408 |
+
"""處理單個問題(用於並行處理)"""
|
| 409 |
+
try:
|
| 410 |
+
# 更新狀態
|
| 411 |
+
with status_lock:
|
| 412 |
+
processing_status.current_item = f"問題 {question_index + 1}/{total_questions}"
|
| 413 |
+
|
| 414 |
+
# 使用智慧路由與重排序檢索
|
| 415 |
+
search_results = intelligent_routing_and_reranking(question, selected_speakers, top_k=30)
|
| 416 |
+
|
| 417 |
+
# 過濾結果
|
| 418 |
+
filtered_results = [r for r in search_results if r.weighted_score >= RELEVANCE_THRESHOLD]
|
| 419 |
+
|
| 420 |
+
if not filtered_results:
|
| 421 |
+
return {
|
| 422 |
+
'question': question,
|
| 423 |
+
'answer': "未找到相關內容",
|
| 424 |
+
'raw_contexts': [],
|
| 425 |
+
'success': False
|
| 426 |
+
}
|
| 427 |
+
|
| 428 |
+
# 構建上下文(使用所有符合條件的結果)
|
| 429 |
+
context = ""
|
| 430 |
+
raw_contexts = []
|
| 431 |
+
|
| 432 |
+
for j, result in enumerate(filtered_results):
|
| 433 |
+
context += f"[片段 {j+1}]\n"
|
| 434 |
+
context += f"發言人:{result.speaker}\n"
|
| 435 |
+
context += f"內容:{result.text}\n"
|
| 436 |
+
context += f"相關性:{result.weighted_score:.3f}\n\n"
|
| 437 |
+
|
| 438 |
+
raw_contexts.append({
|
| 439 |
+
'speaker': result.speaker,
|
| 440 |
+
'text': result.text,
|
| 441 |
+
'turn_index': result.turn_index,
|
| 442 |
+
'score': result.weighted_score
|
| 443 |
+
})
|
| 444 |
+
|
| 445 |
+
# 調用 GPT(含重試機制)
|
| 446 |
+
prompt = f"""基於以下訪談內容回答訪綱問題:
|
| 447 |
+
|
| 448 |
+
{context}
|
| 449 |
+
|
| 450 |
+
問題:{question}
|
| 451 |
+
|
| 452 |
+
請提供:
|
| 453 |
+
1. 主要回答
|
| 454 |
+
2. 不同受訪者的觀點(如果有)
|
| 455 |
+
3. 具體引述"""
|
| 456 |
+
|
| 457 |
+
answer = call_gpt_with_retry(prompt)
|
| 458 |
+
|
| 459 |
+
return {
|
| 460 |
+
'question': question,
|
| 461 |
+
'answer': answer,
|
| 462 |
+
'raw_contexts': raw_contexts,
|
| 463 |
+
'success': True
|
| 464 |
+
}
|
| 465 |
+
|
| 466 |
+
except Exception as e:
|
| 467 |
+
print(f"處理問題失敗: {str(e)}")
|
| 468 |
+
return {
|
| 469 |
+
'question': question,
|
| 470 |
+
'answer': f"處理失敗: {str(e)}",
|
| 471 |
+
'raw_contexts': [],
|
| 472 |
+
'success': False
|
| 473 |
+
}
|
| 474 |
+
|
| 475 |
+
def parallel_process_questions(questions, selected_speakers, progress_callback=None):
|
| 476 |
+
"""並行處理多個問題"""
|
| 477 |
+
results = []
|
| 478 |
+
total = len(questions)
|
| 479 |
|
| 480 |
+
with status_lock:
|
| 481 |
+
processing_status.total_items = total
|
| 482 |
+
processing_status.completed_items = 0
|
| 483 |
+
processing_status.failed_items = 0
|
| 484 |
+
processing_status.start_time = time.time()
|
| 485 |
|
| 486 |
+
with ThreadPoolExecutor(max_workers=MAX_WORKERS) as executor:
|
| 487 |
+
# 提交所有任務
|
| 488 |
+
future_to_question = {
|
| 489 |
+
executor.submit(
|
| 490 |
+
process_single_question,
|
| 491 |
+
question,
|
| 492 |
+
selected_speakers,
|
| 493 |
+
i,
|
| 494 |
+
total
|
| 495 |
+
): (i, question)
|
| 496 |
+
for i, question in enumerate(questions)
|
| 497 |
+
}
|
| 498 |
+
|
| 499 |
+
# 處理完成的任務
|
| 500 |
+
for future in as_completed(future_to_question):
|
| 501 |
+
i, question = future_to_question[future]
|
| 502 |
+
try:
|
| 503 |
+
result = future.result(timeout=60) # 60秒超時
|
| 504 |
+
results.append((i, result))
|
| 505 |
+
|
| 506 |
+
with status_lock:
|
| 507 |
+
if result['success']:
|
| 508 |
+
processing_status.completed_items += 1
|
| 509 |
+
else:
|
| 510 |
+
processing_status.failed_items += 1
|
| 511 |
+
|
| 512 |
+
# 計算預估時間
|
| 513 |
+
elapsed = time.time() - processing_status.start_time
|
| 514 |
+
if processing_status.completed_items > 0:
|
| 515 |
+
avg_time = elapsed / processing_status.completed_items
|
| 516 |
+
remaining = total - processing_status.completed_items
|
| 517 |
+
processing_status.estimated_time = avg_time * remaining
|
| 518 |
+
|
| 519 |
+
# 進度回調
|
| 520 |
+
if progress_callback:
|
| 521 |
+
progress = (processing_status.completed_items + processing_status.failed_items) / total
|
| 522 |
+
progress_callback(progress, f"已處理 {processing_status.completed_items}/{total} 個問題")
|
| 523 |
+
|
| 524 |
+
except Exception as e:
|
| 525 |
+
print(f"任務執行失敗: {str(e)}")
|
| 526 |
+
results.append((i, {
|
| 527 |
+
'question': question,
|
| 528 |
+
'answer': f"處理失敗: {str(e)}",
|
| 529 |
+
'raw_contexts': [],
|
| 530 |
+
'success': False
|
| 531 |
+
}))
|
| 532 |
+
|
| 533 |
+
with status_lock:
|
| 534 |
+
processing_status.failed_items += 1
|
| 535 |
+
processing_status.errors.append(str(e))
|
| 536 |
|
| 537 |
+
# 按原始順序排序結果
|
| 538 |
+
results.sort(key=lambda x: x[0])
|
| 539 |
+
return [r[1] for r in results]
|
| 540 |
|
| 541 |
# ==========================================
|
| 542 |
+
# RAG 對話函數
|
| 543 |
# ==========================================
|
| 544 |
def rag_chat(question, selected_speakers, history):
|
| 545 |
+
"""RAG 對話處理"""
|
| 546 |
if not init_success:
|
| 547 |
return history + [[question, "系統尚未初始化,請稍後再試。"]]
|
| 548 |
|
| 549 |
try:
|
| 550 |
# 執行智慧路由與重排序
|
| 551 |
+
search_results = intelligent_routing_and_reranking(question, selected_speakers, top_k=30)
|
| 552 |
+
|
| 553 |
+
# 過濾結果(使用所有 >= 0.4 的結果)
|
| 554 |
+
filtered_results = [r for r in search_results if r.weighted_score >= RELEVANCE_THRESHOLD]
|
| 555 |
|
| 556 |
+
if not filtered_results:
|
| 557 |
return history + [[question, "未找到相關內容,請嘗試其他問題。"]]
|
| 558 |
|
| 559 |
+
# 構建上下文(使用所有符合條件的結果)
|
| 560 |
context = "相關訪談內容:\n\n"
|
| 561 |
raw_contexts = []
|
| 562 |
|
| 563 |
+
for i, result in enumerate(filtered_results):
|
| 564 |
+
context += f"[片段 {i+1}]\n"
|
| 565 |
context += f"發言人:{result.speaker}\n"
|
| 566 |
context += f"內容:{result.text}\n"
|
| 567 |
+
context += f"相關性分數:{result.weighted_score:.3f}\n\n"
|
| 568 |
|
|
|
|
| 569 |
if result.text:
|
| 570 |
raw_context_text = f"[{result.speaker} - Turn {result.turn_index}]: {result.text}"
|
| 571 |
raw_contexts.append(raw_context_text)
|
| 572 |
|
|
|
|
| 573 |
if not raw_contexts:
|
| 574 |
raw_contexts = ["未能提取原始內容"]
|
| 575 |
|
| 576 |
+
# 構建 GPT prompt
|
| 577 |
prompt = f"""基於以下訪談內容回答問題。請提供準確、完整的回答。
|
| 578 |
|
| 579 |
{context}
|
| 580 |
|
| 581 |
+
問題:{question}"""
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 582 |
|
| 583 |
+
# 調用 GPT(含重試機制)
|
| 584 |
+
answer = call_gpt_with_retry(prompt)
|
| 585 |
|
| 586 |
+
# 添加原始 RAG 內容(完整顯示所有符合條件的)
|
| 587 |
+
answer_with_sources = f"{answer}\n\n---\n📚 **原始 RAG 來源(共 {len(raw_contexts)} 個):**\n"
|
| 588 |
+
for i, raw_context in enumerate(raw_contexts):
|
|
|
|
| 589 |
if raw_context and raw_context != "未能提取原始內容":
|
| 590 |
+
answer_with_sources += f"\n**來源 {i+1}:**\n{raw_context}\n"
|
|
|
|
| 591 |
else:
|
| 592 |
+
answer_with_sources += f"\n**來源 {i+1}:** 無內容\n"
|
| 593 |
|
| 594 |
return history + [[question, answer_with_sources]]
|
| 595 |
|
| 596 |
except Exception as e:
|
| 597 |
return history + [[question, f"處理過程中發生錯誤:{str(e)}"]]
|
| 598 |
|
| 599 |
+
def export_chat_to_word(chat_history):
|
| 600 |
+
"""將對話匯出為 Word 檔案"""
|
| 601 |
+
try:
|
| 602 |
+
doc = Document()
|
| 603 |
+
doc.add_heading('AI 對話記錄', 0)
|
| 604 |
+
doc.add_paragraph(f'匯出時間:{datetime.now().strftime("%Y-%m-%d %H:%M:%S")}')
|
| 605 |
+
doc.add_paragraph('')
|
| 606 |
+
|
| 607 |
+
for i, (question, answer) in enumerate(chat_history, 1):
|
| 608 |
+
doc.add_heading(f'對話 {i}', level=1)
|
| 609 |
+
|
| 610 |
+
doc.add_heading('問題:', level=2)
|
| 611 |
+
doc.add_paragraph(question)
|
| 612 |
+
|
| 613 |
+
doc.add_heading('回答:', level=2)
|
| 614 |
+
for line in answer.split('\n'):
|
| 615 |
+
if line.strip():
|
| 616 |
+
doc.add_paragraph(line)
|
| 617 |
+
|
| 618 |
+
doc.add_page_break()
|
| 619 |
+
|
| 620 |
+
# 保存到記憶體
|
| 621 |
+
output_buffer = io.BytesIO()
|
| 622 |
+
doc.save(output_buffer)
|
| 623 |
+
output_buffer.seek(0)
|
| 624 |
+
|
| 625 |
+
# 保存到檔案
|
| 626 |
+
output_filename = f"chat_export_{datetime.now().strftime('%Y%m%d_%H%M%S')}.docx"
|
| 627 |
+
with open(output_filename, 'wb') as f:
|
| 628 |
+
f.write(output_buffer.getvalue())
|
| 629 |
+
|
| 630 |
+
# 上傳到 HF Dataset
|
| 631 |
+
upload_to_hf_dataset(output_filename, file_type="chat", metadata={
|
| 632 |
+
'export_time': datetime.now().isoformat(),
|
| 633 |
+
'total_conversations': len(chat_history)
|
| 634 |
+
})
|
| 635 |
+
|
| 636 |
+
return output_filename, "對話已匯出並上傳到資料集"
|
| 637 |
+
|
| 638 |
+
except Exception as e:
|
| 639 |
+
return None, f"匯出失敗:{str(e)}"
|
| 640 |
+
|
| 641 |
# ==========================================
|
| 642 |
+
# 訪綱填答函數
|
| 643 |
# ==========================================
|
| 644 |
def parse_word_document(file_path):
|
| 645 |
"""解析 Word 文檔中的問題"""
|
|
|
|
| 667 |
base_name = os.path.basename(filename)
|
| 668 |
base_name_no_ext = os.path.splitext(base_name)[0]
|
| 669 |
|
|
|
|
| 670 |
for speaker in available_speakers:
|
| 671 |
if speaker in base_name_no_ext:
|
| 672 |
return [speaker]
|
| 673 |
|
| 674 |
+
return None
|
| 675 |
|
| 676 |
+
def single_interviewee_guide_filling(file_path, selected_speakers, file_name=None, progress_callback=None):
|
| 677 |
+
"""單一受訪者訪綱填答 - 使用並行處理"""
|
| 678 |
if not init_success:
|
| 679 |
return None, "系統尚未初始化"
|
| 680 |
|
| 681 |
try:
|
| 682 |
+
# 從檔名檢測受訪者
|
| 683 |
if file_name:
|
| 684 |
detected_speakers = extract_speaker_from_filename(file_name, all_speakers)
|
| 685 |
if detected_speakers:
|
|
|
|
| 692 |
if not questions:
|
| 693 |
return None, "未能從文檔中提取問題"
|
| 694 |
|
| 695 |
+
# 進度更新
|
| 696 |
+
if progress_callback:
|
| 697 |
+
progress_callback(0.1, f"開始處理 {len(questions)} 個問題...")
|
| 698 |
+
|
| 699 |
+
# 並行處理所有問題
|
| 700 |
+
print(f"開始並行處理 {len(questions)} 個問題")
|
| 701 |
+
results = parallel_process_questions(questions, selected_speakers, progress_callback)
|
| 702 |
+
|
| 703 |
# 創建新的 Word 文檔
|
| 704 |
output_doc = Document()
|
| 705 |
output_doc.add_heading('訪談訪綱 - AI 智慧填答', 0)
|
| 706 |
output_doc.add_paragraph(f'處理時間:{datetime.now().strftime("%Y-%m-%d %H:%M:%S")}')
|
| 707 |
output_doc.add_paragraph(f'原始檔案:{file_name if file_name else "未知"}')
|
| 708 |
output_doc.add_paragraph(f'選擇的受訪者:{", ".join(selected_speakers) if selected_speakers else "全部"}')
|
| 709 |
+
output_doc.add_paragraph(f'使用技術:並行處理 ({MAX_WORKERS} 組) + 冠軍級 RAG')
|
| 710 |
+
output_doc.add_paragraph(f'處理統計:成功 {processing_status.completed_items}/{processing_status.total_items},失敗 {processing_status.failed_items}')
|
| 711 |
output_doc.add_paragraph('')
|
| 712 |
|
| 713 |
+
# 添加處理結果
|
| 714 |
+
for i, result in enumerate(results, 1):
|
| 715 |
output_doc.add_heading(f'問題 {i}', level=1)
|
| 716 |
+
output_doc.add_paragraph(result['question'])
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 717 |
|
| 718 |
+
if result['success'] and result['raw_contexts']:
|
| 719 |
+
# AI 回答
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 720 |
output_doc.add_heading('AI 分析回答:', level=2)
|
| 721 |
+
for line in result['answer'].split('\n'):
|
| 722 |
if line.strip():
|
| 723 |
output_doc.add_paragraph(line)
|
| 724 |
|
| 725 |
+
# 原始 RAG 內容(顯示所有符合條件的)
|
| 726 |
+
output_doc.add_heading(f'原始 RAG 向量檢索內容(共 {len(result["raw_contexts"])} 個):', level=2)
|
| 727 |
+
for j, raw in enumerate(result['raw_contexts']):
|
| 728 |
p = output_doc.add_paragraph()
|
| 729 |
+
p.add_run(f"{j+1}. [{raw['speaker']} - Turn {raw['turn_index']}] ").bold = True
|
| 730 |
p.add_run(f"(相關性: {raw['score']:.3f})\n")
|
| 731 |
+
p.add_run(raw['text']) # 完整顯示
|
|
|
|
| 732 |
else:
|
| 733 |
+
output_doc.add_paragraph(result['answer'])
|
| 734 |
|
| 735 |
+
output_doc.add_page_break()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 736 |
|
| 737 |
# 保存文檔
|
| 738 |
output_buffer = io.BytesIO()
|
|
|
|
| 746 |
with open(output_filename, 'wb') as f:
|
| 747 |
f.write(output_buffer.getvalue())
|
| 748 |
|
| 749 |
+
# 自動上傳到 HF Dataset
|
| 750 |
+
if progress_callback:
|
| 751 |
+
progress_callback(0.9, "正在上傳到資料集...")
|
| 752 |
+
|
| 753 |
+
success, path = upload_to_hf_dataset(output_filename, file_type="guide", metadata={
|
| 754 |
+
'original_file': file_name,
|
| 755 |
+
'speakers': selected_speakers,
|
| 756 |
+
'total_questions': len(questions),
|
| 757 |
+
'successful_answers': processing_status.completed_items,
|
| 758 |
+
'processing_time': time.time() - processing_status.start_time
|
| 759 |
+
})
|
| 760 |
+
|
| 761 |
+
if success:
|
| 762 |
+
message = f"訪綱填答完成!已上傳到:{path}"
|
| 763 |
+
else:
|
| 764 |
+
message = f"訪綱填答完成!本地檔案:{output_filename}"
|
| 765 |
+
|
| 766 |
+
if progress_callback:
|
| 767 |
+
progress_callback(1.0, message)
|
| 768 |
+
|
| 769 |
+
return output_filename, message
|
| 770 |
|
| 771 |
except Exception as e:
|
| 772 |
return None, f"處理失敗:{str(e)}"
|
| 773 |
|
| 774 |
+
def batch_process_guides(files, default_speakers, progress_callback=None):
|
| 775 |
+
"""批量處理多個訪綱檔案(並行處理)"""
|
| 776 |
if not init_success:
|
| 777 |
return [], "系統尚未初始化"
|
| 778 |
|
| 779 |
results = []
|
| 780 |
processed_files = []
|
| 781 |
+
total_files = len(files)
|
| 782 |
|
| 783 |
try:
|
|
|
|
| 784 |
print(f"開始批量處理 {total_files} 個檔案")
|
| 785 |
|
| 786 |
+
# 重置狀態
|
| 787 |
+
with status_lock:
|
| 788 |
+
processing_status.total_items = total_files
|
| 789 |
+
processing_status.completed_items = 0
|
| 790 |
+
processing_status.failed_items = 0
|
| 791 |
+
processing_status.errors.clear()
|
| 792 |
+
|
| 793 |
+
def process_file(file_info):
|
| 794 |
+
"""處理單個檔案的函數"""
|
| 795 |
+
idx, file = file_info
|
| 796 |
try:
|
| 797 |
file_name = file.name if hasattr(file, 'name') else str(file)
|
| 798 |
+
print(f"\n處理檔案 {idx+1}/{total_files}: {file_name}")
|
| 799 |
|
| 800 |
+
# 從檔名檢測受訪者
|
| 801 |
detected_speakers = extract_speaker_from_filename(file_name, all_speakers)
|
| 802 |
|
| 803 |
if detected_speakers:
|
| 804 |
speakers_to_use = detected_speakers
|
| 805 |
+
status_msg = f"檔案 {idx+1}: 檢測到受訪者 {detected_speakers[0]}"
|
| 806 |
else:
|
| 807 |
speakers_to_use = default_speakers
|
| 808 |
+
status_msg = f"檔案 {idx+1}: 使用預設受訪者"
|
| 809 |
|
| 810 |
print(status_msg)
|
|
|
|
| 811 |
|
| 812 |
+
# 處理單個檔案(含並行處理問題)
|
| 813 |
output_file, process_status = single_interviewee_guide_filling(
|
| 814 |
file.name if hasattr(file, 'name') else file,
|
| 815 |
speakers_to_use,
|
| 816 |
+
file_name,
|
| 817 |
+
None # 批量處理時不使用個別進度回調
|
| 818 |
)
|
| 819 |
|
| 820 |
if output_file:
|
| 821 |
+
return {
|
| 822 |
+
'success': True,
|
| 823 |
+
'file_name': file_name,
|
| 824 |
+
'output_file': output_file,
|
| 825 |
+
'status': process_status
|
| 826 |
+
}
|
| 827 |
else:
|
| 828 |
+
return {
|
| 829 |
+
'success': False,
|
| 830 |
+
'file_name': file_name,
|
| 831 |
+
'error': process_status
|
| 832 |
+
}
|
| 833 |
+
|
| 834 |
except Exception as e:
|
| 835 |
+
return {
|
| 836 |
+
'success': False,
|
| 837 |
+
'file_name': file_name if 'file_name' in locals() else f"檔案 {idx+1}",
|
| 838 |
+
'error': str(e)
|
| 839 |
+
}
|
| 840 |
+
|
| 841 |
+
# 使用執行緒池並行處理多個檔案
|
| 842 |
+
with ThreadPoolExecutor(max_workers=min(MAX_WORKERS, total_files)) as executor:
|
| 843 |
+
# 提交所有檔案處理任務
|
| 844 |
+
future_to_file = {
|
| 845 |
+
executor.submit(process_file, (i, file)): i
|
| 846 |
+
for i, file in enumerate(files)
|
| 847 |
+
}
|
| 848 |
+
|
| 849 |
+
# 處理完成的任務
|
| 850 |
+
for future in as_completed(future_to_file):
|
| 851 |
+
file_idx = future_to_file[future]
|
| 852 |
+
try:
|
| 853 |
+
result = future.result(timeout=300) # 5分鐘超時
|
| 854 |
+
|
| 855 |
+
if result['success']:
|
| 856 |
+
processed_files.append(result['output_file'])
|
| 857 |
+
results.append(f"✅ {result['file_name']} -> {result['output_file']}")
|
| 858 |
+
|
| 859 |
+
with status_lock:
|
| 860 |
+
processing_status.completed_items += 1
|
| 861 |
+
else:
|
| 862 |
+
results.append(f"❌ {result['file_name']}: {result['error']}")
|
| 863 |
+
|
| 864 |
+
with status_lock:
|
| 865 |
+
processing_status.failed_items += 1
|
| 866 |
+
processing_status.errors.append(result['error'])
|
| 867 |
+
|
| 868 |
+
# 更新進度
|
| 869 |
+
if progress_callback:
|
| 870 |
+
progress = (processing_status.completed_items + processing_status.failed_items) / total_files
|
| 871 |
+
progress_callback(
|
| 872 |
+
progress,
|
| 873 |
+
f"已處理 {processing_status.completed_items + processing_status.failed_items}/{total_files} 個檔案"
|
| 874 |
+
)
|
| 875 |
+
|
| 876 |
+
except Exception as e:
|
| 877 |
+
error_msg = f"檔案 {file_idx+1} 處理超時或失敗: {str(e)}"
|
| 878 |
+
print(error_msg)
|
| 879 |
+
results.append(error_msg)
|
| 880 |
+
|
| 881 |
+
with status_lock:
|
| 882 |
+
processing_status.failed_items += 1
|
| 883 |
+
processing_status.errors.append(str(e))
|
| 884 |
|
| 885 |
# 彙總結果
|
| 886 |
+
summary = f"\n處理完成!\n成功: {processing_status.completed_items}/{total_files} 個檔案\n"
|
| 887 |
+
if processing_status.failed_items > 0:
|
| 888 |
+
summary += f"失敗: {processing_status.failed_items} 個檔案\n"
|
| 889 |
summary += "\n詳細結果:\n" + "\n".join(results)
|
| 890 |
|
| 891 |
return processed_files, summary
|
|
|
|
| 906 |
.gradio-container {
|
| 907 |
font-family: 'Microsoft JhengHei', sans-serif;
|
| 908 |
}
|
| 909 |
+
.progress-bar {
|
| 910 |
+
background-color: #4CAF50;
|
| 911 |
}
|
| 912 |
"""
|
| 913 |
) as app:
|
| 914 |
# 標題
|
| 915 |
gr.Markdown("""
|
| 916 |
+
# 🎙️ 訪談轉錄稿智慧分析系統 v2.0
|
| 917 |
|
| 918 |
+
**技術架構:** Multilingual-E5-Large + GPT-4o-mini + 冠軍級 RAG + 並行處理
|
| 919 |
|
| 920 |
**核心功能:**
|
| 921 |
+
- 🔍 智慧語義搜尋(顯示所有 ≥0.4 分結果)
|
| 922 |
+
- 💬 AI 對話(含匯出功能)
|
| 923 |
+
- 📝 訪綱自動填答(5組並行處理)
|
| 924 |
+
- 📊 批量處理(自動上傳 HF Dataset)
|
| 925 |
+
- ⚡ 錯誤重試機制 + 進度追蹤
|
| 926 |
""")
|
| 927 |
|
| 928 |
# 系統狀態
|
|
|
|
| 938 |
with gr.Tabs():
|
| 939 |
# Tab 1: AI 對話
|
| 940 |
with gr.Tab("💬 AI 對話"):
|
| 941 |
+
gr.Markdown("### 智慧問答系統(顯示所有相關 RAG)")
|
| 942 |
|
| 943 |
with gr.Row():
|
| 944 |
with gr.Column(scale=1):
|
|
|
|
| 966 |
|
| 967 |
with gr.Row():
|
| 968 |
clear_btn = gr.Button("清除對話")
|
| 969 |
+
export_btn = gr.Button("📥 匯出對話為 Word", variant="secondary")
|
| 970 |
+
|
| 971 |
+
export_status = gr.Textbox(
|
| 972 |
+
label="匯出狀態",
|
| 973 |
+
visible=False
|
| 974 |
+
)
|
| 975 |
+
|
| 976 |
+
export_file = gr.File(
|
| 977 |
+
label="下載匯出檔案",
|
| 978 |
+
visible=False
|
| 979 |
+
)
|
| 980 |
|
| 981 |
# Tab 2: 訪綱填答
|
| 982 |
with gr.Tab("📝 訪綱填答"):
|
|
|
|
| 984 |
### 智慧訪綱填答系統
|
| 985 |
|
| 986 |
**特色功能:**
|
| 987 |
+
- 🚀 5組並行處理(速度提升5倍)
|
| 988 |
+
- 📊 顯示所有 ≥0.4 分的 RAG 結果
|
| 989 |
+
- 🔄 錯誤自動重試(最多3次)
|
| 990 |
+
- 📤 自動上傳到 HF Dataset
|
| 991 |
+
- 🏷️ 檔名自動識別受訪者
|
|
|
|
|
|
|
|
|
|
|
|
|
| 992 |
""")
|
| 993 |
|
| 994 |
with gr.Row():
|
|
|
|
| 999 |
info="檔名中有受訪者名稱時會自動覆蓋此選擇"
|
| 1000 |
)
|
| 1001 |
|
| 1002 |
+
# 單檔上傳
|
| 1003 |
with gr.Accordion("單檔處理", open=False):
|
| 1004 |
single_file_input = gr.File(
|
| 1005 |
label="上傳單個訪綱 (Word 格式)",
|
|
|
|
| 1007 |
)
|
| 1008 |
single_process_btn = gr.Button("處理單檔", variant="secondary")
|
| 1009 |
|
| 1010 |
+
# 批量上傳
|
| 1011 |
batch_file_input = gr.File(
|
| 1012 |
label="批量上傳訪綱(最多15個 Word 檔案)",
|
| 1013 |
file_types=[".docx"],
|
| 1014 |
file_count="multiple"
|
| 1015 |
)
|
| 1016 |
|
| 1017 |
+
batch_process_btn = gr.Button("🚀 批量並行處理", variant="primary", size="lg")
|
| 1018 |
|
| 1019 |
with gr.Column():
|
| 1020 |
+
# 進度條
|
| 1021 |
+
progress_bar = gr.Progress()
|
| 1022 |
+
|
| 1023 |
process_status = gr.Textbox(
|
| 1024 |
label="處理狀態",
|
| 1025 |
interactive=False,
|
| 1026 |
lines=10
|
| 1027 |
)
|
| 1028 |
|
| 1029 |
+
# 處理統計
|
| 1030 |
+
with gr.Row():
|
| 1031 |
+
stat_total = gr.Number(label="總計", value=0)
|
| 1032 |
+
stat_success = gr.Number(label="成功", value=0)
|
| 1033 |
+
stat_failed = gr.Number(label="失敗", value=0)
|
| 1034 |
+
stat_time = gr.Number(label="預估時間(秒)", value=0)
|
| 1035 |
+
|
| 1036 |
+
# 下載區
|
| 1037 |
single_download_file = gr.File(
|
| 1038 |
label="下載單檔結果",
|
| 1039 |
visible=False
|
| 1040 |
)
|
| 1041 |
|
|
|
|
| 1042 |
batch_download_files = gr.File(
|
| 1043 |
label="下載所有結果",
|
| 1044 |
visible=False,
|
|
|
|
| 1048 |
# 技術細節
|
| 1049 |
with gr.Accordion("🔧 技術細節", open=False):
|
| 1050 |
gr.Markdown("""
|
| 1051 |
+
### 系統特色
|
| 1052 |
+
|
| 1053 |
+
**1. 並行處理架構**
|
| 1054 |
+
- 5組同時處理不同問題
|
| 1055 |
+
- ThreadPoolExecutor 管理
|
| 1056 |
+
- 自動負載平衡
|
| 1057 |
+
|
| 1058 |
+
**2. RAG 結果顯示**
|
| 1059 |
+
- 所有 weighted_score ≥ 0.4 的結果都納入分析
|
| 1060 |
+
- 完整顯示原始內容(不截斷)
|
| 1061 |
+
- 顯示相關性分數
|
| 1062 |
+
|
| 1063 |
+
**3. 錯誤處理**
|
| 1064 |
+
- API 調用失敗自動重試(最多3次)
|
| 1065 |
+
- 指數退避策略(2^n 秒)
|
| 1066 |
+
- 失敗隔離(不影響其他任務)
|
| 1067 |
+
|
| 1068 |
+
**4. HF Dataset 自動上傳**
|
| 1069 |
+
- 位置:s880453/interview-outputs
|
| 1070 |
+
- 自動分類:guide_outputs / chat_exports
|
| 1071 |
+
- 包含處理 metadata
|
| 1072 |
+
|
| 1073 |
+
**5. 進度追蹤**
|
| 1074 |
+
- 即時顯示處理進度
|
| 1075 |
+
- 預估剩餘時間
|
| 1076 |
+
- 詳細錯誤日誌
|
| 1077 |
""")
|
| 1078 |
|
| 1079 |
# 事件處理
|
|
|
|
| 1086 |
def clear_chat():
|
| 1087 |
return []
|
| 1088 |
|
| 1089 |
+
def export_chat(history):
|
| 1090 |
+
if not history:
|
| 1091 |
+
return gr.Textbox(value="沒有對話可匯出", visible=True), gr.File(visible=False)
|
| 1092 |
+
|
| 1093 |
+
file_path, status = export_chat_to_word(history)
|
| 1094 |
+
if file_path:
|
| 1095 |
+
return (
|
| 1096 |
+
gr.Textbox(value=status, visible=True),
|
| 1097 |
+
gr.File(value=file_path, visible=True)
|
| 1098 |
+
)
|
| 1099 |
+
else:
|
| 1100 |
+
return gr.Textbox(value=status, visible=True), gr.File(visible=False)
|
| 1101 |
+
|
| 1102 |
+
def process_single_guide_with_progress(file, speakers):
|
| 1103 |
+
"""處理單個檔案(含進度顯示)"""
|
| 1104 |
if not file:
|
| 1105 |
+
return (
|
| 1106 |
+
"請上傳文件",
|
| 1107 |
+
gr.File(visible=False),
|
| 1108 |
+
0, 0, 0, 0
|
| 1109 |
+
)
|
| 1110 |
|
| 1111 |
+
def progress_update(progress, message):
|
| 1112 |
+
return message
|
| 1113 |
+
|
| 1114 |
+
result_file, status = single_interviewee_guide_filling(
|
| 1115 |
+
file.name,
|
| 1116 |
+
speakers,
|
| 1117 |
+
file.name,
|
| 1118 |
+
progress_update
|
| 1119 |
+
)
|
| 1120 |
|
| 1121 |
if result_file:
|
| 1122 |
+
return (
|
| 1123 |
+
status,
|
| 1124 |
+
gr.File(value=result_file, visible=True),
|
| 1125 |
+
processing_status.total_items,
|
| 1126 |
+
processing_status.completed_items,
|
| 1127 |
+
processing_status.failed_items,
|
| 1128 |
+
processing_status.estimated_time
|
| 1129 |
+
)
|
| 1130 |
else:
|
| 1131 |
+
return (
|
| 1132 |
+
status,
|
| 1133 |
+
gr.File(visible=False),
|
| 1134 |
+
processing_status.total_items,
|
| 1135 |
+
processing_status.completed_items,
|
| 1136 |
+
processing_status.failed_items,
|
| 1137 |
+
0
|
| 1138 |
+
)
|
| 1139 |
|
| 1140 |
+
def process_batch_guides_with_progress(files, speakers):
|
| 1141 |
+
"""批量處理(含進度顯示)"""
|
| 1142 |
if not files:
|
| 1143 |
+
return (
|
| 1144 |
+
"請上傳至少一個檔案",
|
| 1145 |
+
gr.File(visible=False),
|
| 1146 |
+
0, 0, 0, 0
|
| 1147 |
+
)
|
| 1148 |
|
|
|
|
| 1149 |
if len(files) > 15:
|
| 1150 |
+
return (
|
| 1151 |
+
f"檔案數量超過限制(最多15個),您上傳了 {len(files)} 個",
|
| 1152 |
+
gr.File(visible=False),
|
| 1153 |
+
0, 0, 0, 0
|
| 1154 |
+
)
|
| 1155 |
+
|
| 1156 |
+
def progress_update(progress, message):
|
| 1157 |
+
return message
|
| 1158 |
|
| 1159 |
# 批量處理
|
| 1160 |
+
processed_files, status = batch_process_guides(files, speakers, progress_update)
|
| 1161 |
|
| 1162 |
if processed_files:
|
| 1163 |
+
return (
|
| 1164 |
+
status,
|
| 1165 |
+
gr.File(value=processed_files, visible=True, file_count="multiple"),
|
| 1166 |
+
processing_status.total_items,
|
| 1167 |
+
processing_status.completed_items,
|
| 1168 |
+
processing_status.failed_items,
|
| 1169 |
+
processing_status.estimated_time
|
| 1170 |
+
)
|
| 1171 |
else:
|
| 1172 |
+
return (
|
| 1173 |
+
status,
|
| 1174 |
+
gr.File(visible=False),
|
| 1175 |
+
processing_status.total_items,
|
| 1176 |
+
processing_status.completed_items,
|
| 1177 |
+
processing_status.failed_items,
|
| 1178 |
+
0
|
| 1179 |
+
)
|
| 1180 |
|
| 1181 |
def update_status():
|
| 1182 |
success, message = initialize_system()
|
|
|
|
| 1203 |
|
| 1204 |
clear_btn.click(clear_chat, outputs=[chatbot])
|
| 1205 |
|
| 1206 |
+
export_btn.click(
|
| 1207 |
+
export_chat,
|
| 1208 |
+
inputs=[chatbot],
|
| 1209 |
+
outputs=[export_status, export_file]
|
| 1210 |
+
)
|
| 1211 |
+
|
| 1212 |
# 單檔處理
|
| 1213 |
single_process_btn.click(
|
| 1214 |
+
process_single_guide_with_progress,
|
| 1215 |
inputs=[single_file_input, guide_speakers],
|
| 1216 |
+
outputs=[
|
| 1217 |
+
process_status,
|
| 1218 |
+
single_download_file,
|
| 1219 |
+
stat_total,
|
| 1220 |
+
stat_success,
|
| 1221 |
+
stat_failed,
|
| 1222 |
+
stat_time
|
| 1223 |
+
]
|
| 1224 |
)
|
| 1225 |
|
| 1226 |
# 批量處理
|
| 1227 |
batch_process_btn.click(
|
| 1228 |
+
process_batch_guides_with_progress,
|
| 1229 |
inputs=[batch_file_input, guide_speakers],
|
| 1230 |
+
outputs=[
|
| 1231 |
+
process_status,
|
| 1232 |
+
batch_download_files,
|
| 1233 |
+
stat_total,
|
| 1234 |
+
stat_success,
|
| 1235 |
+
stat_failed,
|
| 1236 |
+
stat_time
|
| 1237 |
+
]
|
| 1238 |
)
|
| 1239 |
|
| 1240 |
init_btn.click(
|
|
|
|
| 1247 |
update_status,
|
| 1248 |
outputs=[status_text, speaker_selector, guide_speakers]
|
| 1249 |
)
|
| 1250 |
+
|
| 1251 |
+
# 啟用 Queue(支援並行處理)
|
| 1252 |
+
app.queue(
|
| 1253 |
+
concurrency_count=MAX_WORKERS,
|
| 1254 |
+
max_size=50
|
| 1255 |
+
)
|
| 1256 |
|
| 1257 |
return app
|
| 1258 |
|
|
|
|
| 1260 |
# 主程式入口
|
| 1261 |
# ==========================================
|
| 1262 |
if __name__ == "__main__":
|
| 1263 |
+
# 創建並啟動應用(加入密碼保護)
|
| 1264 |
app = create_interface()
|
| 1265 |
+
|
| 1266 |
+
# 設定身份驗證
|
| 1267 |
+
# 支援多種驗證方式:
|
| 1268 |
+
# 1. 任意使用者名稱 + 正確密碼
|
| 1269 |
+
# 2. 可以設定多組帳號密碼
|
| 1270 |
+
def authenticate(username, password):
|
| 1271 |
+
"""驗證函數 - 確保安全性"""
|
| 1272 |
+
# 主要密碼驗證
|
| 1273 |
+
if password == ACCESS_PASSWORD:
|
| 1274 |
+
return True
|
| 1275 |
+
# 可選:添加特定使用者帳號
|
| 1276 |
+
valid_users = {
|
| 1277 |
+
"admin": ACCESS_PASSWORD,
|
| 1278 |
+
"user": ACCESS_PASSWORD
|
| 1279 |
+
}
|
| 1280 |
+
return username in valid_users and valid_users[username] == password
|
| 1281 |
+
|
| 1282 |
+
# 啟動應用(需要密碼才能訪問)
|
| 1283 |
+
app.launch(
|
| 1284 |
+
share=False,
|
| 1285 |
+
server_name="0.0.0.0",
|
| 1286 |
+
server_port=7860,
|
| 1287 |
+
auth=authenticate,
|
| 1288 |
+
auth_message="🔒 訪談轉錄稿 RAG 系統\n請輸入密碼以訪問系統",
|
| 1289 |
+
show_api=False, # 隱藏 API 文檔
|
| 1290 |
+
show_error=False # 不顯示詳細錯誤(增加安全性)
|
| 1291 |
+
)
|