app.py

import gradio as gr
import json
from pathlib import Path
import yaml
import re
import logging
import io
import sys
import os
import re
from datetime import datetime, timezone, timedelta
import requests

from tools import FileUploader, ResultExtractor, audio_to_str, image_to_str, azure_speech_to_text #gege的多模态
import numpy as np
from scipy.io.wavfile import write as write_wav
from PIL import Image

# 指定保存文件的相对路径
SAVE_DIR = 'download'  # 相对路径
os.makedirs(SAVE_DIR, exist_ok=True)  # 确保目录存在

def save_audio(audio, filename):
    """保存音频为.wav文件"""
    sample_rate, audio_data = audio
    write_wav(filename, sample_rate, audio_data)

def save_image(image, filename):
    """保存图片为.jpg文件"""
    img = Image.fromarray(image.astype('uint8'))
    img.save(filename)

# --- IP获取功能 (从 se_app.py 迁移) ---
def get_client_ip(request: gr.Request, debug_mode=False):
    """获取客户端真实IP地址"""
    if request:
        # 从请求头中获取真实IP（考虑代理情况）
        x_forwarded_for = request.headers.get("x-forwarded-for", "")
        if x_forwarded_for:
            client_ip = x_forwarded_for.split(",")[0]
        else:
            client_ip = request.client.host
        if debug_mode:
            print(f"Debug: Client IP detected as {client_ip}")
        return client_ip
    return "unknown"

# --- 配置加载 (从 config_loader.py 迁移并简化) ---
CONFIG = None
HF_CONFIG_PATH = Path(__file__).parent / "todogen_LLM_config.yaml"

def load_hf_config():
    global CONFIG
    if CONFIG is None:
        try:
            with open(HF_CONFIG_PATH, 'r', encoding='utf-8') as f:
                CONFIG = yaml.safe_load(f)
            print(f"✅ 配置已加载: {HF_CONFIG_PATH}")
        except FileNotFoundError:
            print(f"❌ 错误: 配置文件 {HF_CONFIG_PATH} 未找到。请确保它在 hf 目录下。")
            CONFIG = {} # 提供一个空配置以避免后续错误
        except Exception as e:
            print(f"❌ 加载配置文件 {HF_CONFIG_PATH} 时出错: {e}")
            CONFIG = {}
    return CONFIG

def get_hf_openai_config():
    config = load_hf_config()
    return config.get('openai', {})

def get_hf_openai_filter_config():
    config = load_hf_config()
    return config.get('openai_filter', {})

def get_hf_xunfei_config():
    config = load_hf_config()
    return config.get('xunfei', {})

def get_hf_azure_speech_config():
    config = load_hf_config()
    return config.get('azure_speech', {})

def get_hf_paths_config():
    config = load_hf_config()
    # 在hf环境下，路径相对于hf目录
    base = Path(__file__).resolve().parent
    paths_cfg = config.get('paths', {})
    return {
        'base_dir': base,
        'prompt_template': base / paths_cfg.get('prompt_template', 'prompt_template.txt'),
        'true_positive_examples': base / paths_cfg.get('true_positive_examples', 'TruePositive_few_shot.txt'),
        'false_positive_examples': base / paths_cfg.get('false_positive_examples', 'FalsePositive_few_shot.txt'),
        # data_dir 和 logging_dir 在 app.py 中可能用途不大，除非需要保存 LLM 输出
    }

# --- LLM Client 初始化 (使用 NVIDIA API) ---
# 从配置加载 NVIDIA API 的 base_url, api_key 和 model
llm_config = get_hf_openai_config()
NVIDIA_API_BASE_URL = llm_config.get('base_url')
NVIDIA_API_KEY = llm_config.get('api_key')
NVIDIA_MODEL_NAME = llm_config.get('model')

# 从配置加载 Filter API 的 base_url, api_key 和 model
filter_config = get_hf_openai_filter_config()
Filter_API_BASE_URL = filter_config.get('base_url_filter')
Filter_API_KEY = filter_config.get('api_key_filter')
Filter_MODEL_NAME = filter_config.get('model_filter')


if not NVIDIA_API_BASE_URL or not NVIDIA_API_KEY or not NVIDIA_MODEL_NAME:
    print("❌ 错误: NVIDIA API 配置不完整。请检查 todogen_LLM_config.yaml 中的 openai 部分。")
    # 提供默认值或退出，以便程序可以继续运行，但LLM调用会失败
    NVIDIA_API_BASE_URL = ""
    NVIDIA_API_KEY = ""
    NVIDIA_MODEL_NAME = ""

if not Filter_API_BASE_URL or not Filter_API_KEY or not Filter_MODEL_NAME:
    print("❌ 错误: Filter API 配置不完整。请检查 todogen_LLM_config.yaml 中的 openai_filter 部分。")
    # 提供默认值或退出，以便程序可以继续运行，但Filter LLM调用会失败
    Filter_API_BASE_URL = ""
    Filter_API_KEY = ""
    Filter_MODEL_NAME = ""

# --- 日志配置 (简化版) ---
# 修正后的标准流编码设置 (如果需要，但 Gradio 通常处理自己的输出)
# sys.stdin = io.TextIOWrapper(sys.stdin.buffer, encoding='utf-8')
# sys.stdout = io.TextIOWrapper(sys.stdout.buffer, encoding='utf-8', write_through=True)
# sys.stderr = io.TextIOWrapper(sys.stderr.buffer, encoding='utf-8', write_through=True)
logging.basicConfig(level=logging.INFO, format='%(asctime)s - %(levelname)s - %(message)s')
logger = logging.getLogger(__name__)

# --- Prompt 和 Few-Shot 加载 (从 todogen_llm.py 迁移并适配) ---
def load_single_few_shot_file_hf(file_path: Path) -> str:
    """加载单个 few-shot 文件并转义 { 和 }"""
    try:
        with open(file_path, 'r', encoding='utf-8') as f:
            content = f.read()
            escaped_content = content.replace('{', '{{').replace('}', '}}')
            logger.info(f"✅ 成功加载并转义文件: {file_path}")
            return escaped_content
    except FileNotFoundError:
        logger.warning(f"⚠️ 警告：找不到文件 {file_path}。")
        return ""
    except Exception as e:
        logger.error(f"❌ 加载文件 {file_path} 时出错: {e}", exc_info=True)
        return ""

PROMPT_TEMPLATE_CONTENT = ""
TRUE_POSITIVE_EXAMPLES_CONTENT = ""
FALSE_POSITIVE_EXAMPLES_CONTENT = ""

def load_prompt_data_hf():
    global PROMPT_TEMPLATE_CONTENT, TRUE_POSITIVE_EXAMPLES_CONTENT, FALSE_POSITIVE_EXAMPLES_CONTENT
    paths = get_hf_paths_config()
    try:
        with open(paths['prompt_template'], 'r', encoding='utf-8') as f:
            PROMPT_TEMPLATE_CONTENT = f.read()
            logger.info(f"✅ 成功加载 Prompt 模板文件: {paths['prompt_template']}")
    except FileNotFoundError:
        logger.error(f"❌ 错误：找不到 Prompt 模板文件：{paths['prompt_template']}")
        PROMPT_TEMPLATE_CONTENT = "Error: Prompt template not found."
    
    TRUE_POSITIVE_EXAMPLES_CONTENT = load_single_few_shot_file_hf(paths['true_positive_examples'])
    FALSE_POSITIVE_EXAMPLES_CONTENT = load_single_few_shot_file_hf(paths['false_positive_examples'])

# 应用启动时加载 prompts
load_prompt_data_hf()

# --- JSON 解析器 (从 todogen_llm.py 迁移) ---
def json_parser(text: str) -> dict:
    # 改进的JSON解析器，更健壮地处理各种格式
    logger.info(f"Attempting to parse: {text[:200]}...")
    try:
        # 1. 尝试直接将整个文本作为JSON解析
        try:
            parsed_data = json.loads(text)
            # 使用_process_parsed_json处理解析结果
            return _process_parsed_json(parsed_data)
        except json.JSONDecodeError:
            pass # 如果直接解析失败，继续尝试提取代码块

        # 2. 尝试从 ```json ... ``` 代码块中提取和解析
        match = re.search(r'```(?:json)?\n(.*?)```', text, re.DOTALL)
        if match:
            json_str = match.group(1).strip()
            # 修复常见的JSON格式问题
            json_str = re.sub(r',\s*]', ']', json_str)
            json_str = re.sub(r',\s*}', '}', json_str)
            try:
                parsed_data = json.loads(json_str)
                # 使用_process_parsed_json处理解析结果
                return _process_parsed_json(parsed_data)
            except json.JSONDecodeError as e_block:
                logger.warning(f"JSONDecodeError from code block: {e_block} while parsing: {json_str[:200]}")
                # 如果从代码块解析也失败，则继续

        # 3. 尝试查找最外层的 '{...}' 或 '[...]' 作为JSON
        # 先尝试查找数组格式 [...]
        array_match = re.search(r'\[\s*\{.*?\}\s*(?:,\s*\{.*?\}\s*)*\]', text, re.DOTALL)
        if array_match:
            potential_json = array_match.group(0).strip()
            try:
                parsed_data = json.loads(potential_json)
                # 使用_process_parsed_json处理解析结果
                return _process_parsed_json(parsed_data)
            except json.JSONDecodeError:
                logger.warning(f"Could not parse potential JSON array: {potential_json[:200]}")
                pass
        
        # 再尝试查找单个对象格式 {...}
        object_match = re.search(r'\{.*?\}', text, re.DOTALL)
        if object_match:
            potential_json = object_match.group(0).strip()
            try:
                parsed_data = json.loads(potential_json)
                # 使用_process_parsed_json处理解析结果
                return _process_parsed_json(parsed_data)
            except json.JSONDecodeError:
                logger.warning(f"Could not parse potential JSON object: {potential_json[:200]}")
                pass

        # 4. 如果所有尝试都失败，返回错误信息
        logger.error(f"Failed to find or parse JSON block in text: {text[:500]}") # 增加日志长度
        return {"error": "No valid JSON block found or failed to parse", "raw_text": text}

    except Exception as e: # 捕获所有其他意外错误
        logger.error(f"Unexpected error in json_parser: {e} for text: {text[:200]}", exc_info=True)
        return {"error": f"Unexpected error in json_parser: {e}", "raw_text": text}

def _process_parsed_json(parsed_data):
    """处理解析后的JSON数据，确保返回有效的数据结构"""
    try:
        # 如果解析结果是空列表，返回包含空字典的列表
        if isinstance(parsed_data, list):
            if not parsed_data:
                logger.warning("JSON解析结果为空列表，返回包含空字典的列表")
                return [{}]
            
            # 确保列表中的每个元素都是字典
            processed_list = []
            for item in parsed_data:
                if isinstance(item, dict):
                    processed_list.append(item)
                else:
                    # 如果不是字典，将其转换为字典
                    try:
                        processed_list.append({"content": str(item)})
                    except:
                        processed_list.append({"content": "无法转换的项目"})
            
            # 如果处理后的列表为空，返回包含空字典的列表
            if not processed_list:
                logger.warning("处理后的JSON列表为空，返回包含空字典的列表")
                return [{}]
            
            return processed_list
        
        # 如果是字典，直接返回
        elif isinstance(parsed_data, dict):
            return parsed_data
        
        # 如果是其他类型，转换为字典
        else:
            logger.warning(f"JSON解析结果不是列表或字典，而是{type(parsed_data)}，转换为字典")
            return {"content": str(parsed_data)}
    
    except Exception as e:
        logger.error(f"处理解析后的JSON数据时出错: {e}")
        return {"error": f"Error processing parsed JSON: {e}"}

# --- Filter 模块的 System Prompt (从 filter_message/libs.py 迁移) ---
FILTER_SYSTEM_PROMPT = """

# 角色

你是一个专业的短信内容分析助手，根据输入判断内容的类型及可信度，为用户使用信息提供依据和便利。



# 任务

对于输入的多条数据，分析每一条数据内容（主键：`message_id`）属于【物流取件、缴费充值、待付(还)款、会议邀约、其他】的可能性百分比。

主要对于聊天、问候、回执、结果通知、上月账单等信息不需要收件人进行下一步处理的信息，直接归到其他类进行忽略



# 要求

1. 以json格式输出

2. content简洁提炼关键词，字符数<20以内

3. 输入条数和输出条数完全一样



# 输出示例

```

[

    {"message_id":"1111111","content":"账单805.57元待还","物流取件":0,"欠费缴纳":99,"待付(还)款":1,"会议邀约":0,"其他":0, "分类":"欠费缴纳"},

    {"message_id":"222222","content":"邀请你加入飞书视频会议","物流取件":0,"欠费缴纳":0,"待付(还)款":1,"会议邀约":100,"其他":0, "分类":"会议"}

]



```

"""

# --- Filter 核心逻辑 (从ToDoAgent集成) ---
def filter_message_with_llm(text_input: str, message_id: str = "user_input_001"):
    logger.info(f"调用 filter_message_with_llm 处理输入: {text_input} (msg_id: {message_id})")

    # 构造发送给 LLM 的消息
    # filter 模块的 send_llm_with_prompt 接收的是 tuple[tuple] 格式的数据
    # 这里我们只有一个文本输入，需要模拟成那种格式
    mock_data = [(text_input, message_id)]

    # 使用与ToDoAgent相同的system prompt
    system_prompt = """

    # 角色

    你是一个专业的短信内容分析助手，根据输入判断内容的类型及可信度，为用户使用信息提供依据和便利。



    # 任务

    对于输入的多条数据，分析每一条数据内容（主键：`message_id`）属于【物流取件、缴费充值、待付(还)款、会议邀约、其他】的可能性百分比。

    主要对于聊天、问候、回执、结果通知、上月账单等信息不需要收件人进行下一步处理的信息，直接归到其他类进行忽略



    # 要求

    1. 以json格式输出

    2. content简洁提炼关键词，字符数<20以内

    3. 输入条数和输出条数完全一样

    

    # 输出示例

    ```

    [

        {"message_id":"1111111","content":"账单805.57元待还","物流取件":0,"欠费缴纳":99,"待付(还)款":1,"会议邀约":0,"其他":0, "分类":"欠费缴纳"},

        {"message_id":"222222","content":"邀请你加入飞书视频会议","物流取件":0,"欠费缴纳":0,"待付(还)款":1,"会议邀约":100,"其他":0, "分类":"会议邀约"}

    ]

    ```

    """

    llm_messages = [
        {"role": "system", "content": system_prompt},
        {"role": "user", "content": str(mock_data)}
    ]

    try:
        if not Filter_API_BASE_URL or not Filter_API_KEY or not Filter_MODEL_NAME:
            logger.error("Filter API 配置不完整，无法调用 Filter LLM。")
            return [{"error": "Filter API configuration incomplete", "-": "-"}]

        headers = {
            "Authorization": f"Bearer {Filter_API_KEY}",
            "Accept": "application/json"
        }
        payload = {
            "model": Filter_MODEL_NAME,
            "messages": llm_messages,
            "temperature": 0.0,  # 为提高准确率，温度为0（与ToDoAgent一致）
            "top_p": 0.95,
            "max_tokens": 1024,
            "stream": False
        }
        
        api_url = f"{Filter_API_BASE_URL}/chat/completions"

        try:
            response = requests.post(api_url, headers=headers, json=payload)
            response.raise_for_status() # 检查 HTTP 错误
            raw_llm_response = response.json()["choices"][0]["message"]["content"]
            logger.info(f"LLM 原始回复 (部分): {raw_llm_response[:200]}...")
        except requests.exceptions.RequestException as e:
            logger.error(f"调用 Filter API 失败: {e}")
            return [{"error": f"Filter API call failed: {e}", "-": "-"}]
        logger.info(f"Filter LLM 原始回复 (部分): {raw_llm_response[:200]}...")

        # 解析 LLM 响应
        # 移除可能的代码块标记
        raw_llm_response = raw_llm_response.replace("```json", "").replace("```", "")
        parsed_filter_data = json_parser(raw_llm_response)

        if "error" in parsed_filter_data:
            logger.error(f"解析 Filter LLM 响应失败: {parsed_filter_data['error']}")
            return [{"error": f"Filter LLM response parsing error: {parsed_filter_data['error']}"}]
        
        # 返回解析后的数据
        if isinstance(parsed_filter_data, list) and parsed_filter_data:
            # 应用规则：如果分类是欠费缴纳且内容包含"缴费支出"，归类为"其他"
            for item in parsed_filter_data:
                if isinstance(item, dict) and item.get("分类") == "欠费缴纳" and "缴费支出" in item.get("content", ""):
                    item["分类"] = "其他"
            
            # 检查是否有遗漏的消息ID（ToDoAgent的补充逻辑）
            request_id_list = {message_id}
            response_id_list = {item.get('message_id') for item in parsed_filter_data if isinstance(item, dict)}
            diff = request_id_list - response_id_list
            
            if diff:
                logger.warning(f"Filter LLM 响应中有遗漏的消息ID: {diff}")
                # 对于遗漏的消息，添加一个默认分类为"其他"的项
                for missed_id in diff:
                    parsed_filter_data.append({
                        "message_id": missed_id,
                        "content": text_input[:20],  # 截取前20个字符作为content
                        "物流取件": 0,
                        "欠费缴纳": 0,
                        "待付(还)款": 0,
                        "会议邀约": 0,
                        "其他": 100,
                        "分类": "其他"
                    })
            
            return parsed_filter_data
        else:
            logger.warning(f"Filter LLM 返回空列表或非预期格式: {parsed_filter_data}")
            # 返回默认分类为"其他"的项
            return [{
                "message_id": message_id,
                "content": text_input[:20],  # 截取前20个字符作为content
                "物流取件": 0,
                "欠费缴纳": 0,
                "待付(还)款": 0,
                "会议邀约": 0,
                "其他": 100,
                "分类": "其他",
                "error": "Filter LLM returned empty or unexpected format"
            }]

    except Exception as e:
        logger.exception(f"调用 Filter LLM 或解析时发生错误 (filter_message_with_llm)")
        return [{
            "message_id": message_id,
            "content": text_input[:20],  # 截取前20个字符作为content
            "物流取件": 0,
            "欠费缴纳": 0,
            "待付(还)款": 0,
            "会议邀约": 0,
            "其他": 100,
            "分类": "其他",
            "error": f"Filter LLM call/parse error: {str(e)}"
        }]

# --- ToDo List 生成核心逻辑 (使用迁移的代码) ---
def generate_todolist_from_text(text_input: str, message_id: str = "user_input_001"):
    """根据输入文本生成 ToDoList (使用迁移的逻辑)"""
    logger.info(f"调用 generate_todolist_from_text 处理输入: {text_input} (msg_id: {message_id})")
    
    if not PROMPT_TEMPLATE_CONTENT or "Error:" in PROMPT_TEMPLATE_CONTENT:
        logger.error("Prompt 模板未正确加载，无法生成 ToDoList。")
        return [["error", "Prompt template not loaded", "-"]]

    current_time_iso = datetime.now(timezone.utc).isoformat()
    # 转义输入内容中的 { 和 }
    content_escaped = text_input.replace('{', '{{').replace('}', '}}')

    # 构造 prompt
    formatted_prompt = PROMPT_TEMPLATE_CONTENT.format(
        true_positive_examples=TRUE_POSITIVE_EXAMPLES_CONTENT,
        false_positive_examples=FALSE_POSITIVE_EXAMPLES_CONTENT,
        current_time=current_time_iso,
        message_id=message_id,
        content_escaped=content_escaped
    )
    
    # 添加明确的JSON输出指令
    enhanced_prompt = formatted_prompt + """



# 重要提示

请确保你的回复是有效的JSON格式，并且只包含JSON内容。不要添加任何额外的解释或文本。

你的回复应该严格按照上面的输出示例格式，只包含JSON对象，不要有任何其他文本。

"""

    # 构造发送给 LLM 的消息
    llm_messages = [
        {"role": "user", "content": enhanced_prompt}
    ]
    
    logger.info(f"发送给 LLM 的消息 (部分): {str(llm_messages)[:300]}...")

    try:
        # 根据输入文本智能生成 ToDo List
        # 如果是移动话费充值提醒类消息
        if ("充值" in text_input or "缴费" in text_input) and ("移动" in text_input or "话费" in text_input or "余额" in text_input):
            # 直接生成待办事项，不调用API
            todo_item = {
                message_id: {
                    "is_todo": True,
                    "end_time": (datetime.now(timezone.utc) + timedelta(days=3)).isoformat(),
                    "location": "线上:中国移动APP",
                    "todo_content": "缴纳话费",
                    "urgency": "important"
                }
            }
            
            # 转换为表格显示格式 - 合并为一行
            todo_content = "缴纳话费"
            end_time = todo_item[message_id]["end_time"].split("T")[0]
            location = todo_item[message_id]["location"]
            
            # 合并所有信息到任务内容中
            combined_content = f"{todo_content} (截止时间: {end_time}, 地点: {location})"
            
            output_for_df = []
            output_for_df.append([1, combined_content, "重要"])
            
            return output_for_df
        
        # 如果是会议邀约类消息
        elif "会议" in text_input and ("邀请" in text_input or "参加" in text_input):
            # 提取可能的会议时间
            meeting_time = None
            meeting_pattern = r'(\d{1,2}[月/-]\d{1,2}[日号]?\s*\d{1,2}[点:]\d{0,2}|\d{4}[年/-]\d{1,2}[月/-]\d{1,2}[日号]?\s*\d{1,2}[点:]\d{0,2})'
            meeting_match = re.search(meeting_pattern, text_input)
            
            if meeting_match:
                # 简单处理，实际应用中应该更精确地解析日期时间
                meeting_time = (datetime.now(timezone.utc) + timedelta(days=1, hours=2)).isoformat()
            else:
                meeting_time = (datetime.now(timezone.utc) + timedelta(days=1)).isoformat()
            
            todo_item = {
                message_id: {
                    "is_todo": True,
                    "end_time": meeting_time,
                    "location": "线上:会议软件",
                    "todo_content": "参加会议",
                    "urgency": "important"
                }
            }
            
            # 转换为表格显示格式 - 合并为一行
            todo_content = "参加会议"
            end_time = todo_item[message_id]["end_time"].split("T")[0]
            location = todo_item[message_id]["location"]
            
            # 合并所有信息到任务内容中
            combined_content = f"{todo_content} (截止时间: {end_time}, 地点: {location})"
            
            output_for_df = []
            output_for_df.append([1, combined_content, "重要"])
            
            return output_for_df
            
        # 如果是物流取件类消息
        elif ("快递" in text_input or "物流" in text_input or "取件" in text_input) and ("到达" in text_input or "取件码" in text_input or "柜" in text_input):
            # 提取可能的取件码
            pickup_code = None
            code_pattern = r'取件码[是为:]?\s*(\d{4,6})'
            code_match = re.search(code_pattern, text_input)
            
            todo_content = "取快递"
            if code_match:
                pickup_code = code_match.group(1)
                todo_content = f"取快递(取件码:{pickup_code})"
            
            todo_item = {
                message_id: {
                    "is_todo": True,
                    "end_time": (datetime.now(timezone.utc) + timedelta(days=2)).isoformat(),
                    "location": "线下:快递柜",
                    "todo_content": todo_content,
                    "urgency": "important"
                }
            }
            
            # 转换为表格显示格式 - 合并为一行
            end_time = todo_item[message_id]["end_time"].split("T")[0]
            location = todo_item[message_id]["location"]
            
            # 合并所有信息到任务内容中
            combined_content = f"{todo_content} (截止时间: {end_time}, 地点: {location})"
            
            output_for_df = []
            output_for_df.append([1, combined_content, "重要"])
            
            return output_for_df
        
        # 对于其他类型的消息，调用LLM API进行处理
        if not Filter_API_BASE_URL or not Filter_API_KEY or not Filter_MODEL_NAME:
            logger.error("Filter API 配置不完整，无法调用 Filter LLM。")
            return [["error", "Filter API configuration incomplete", "-"]]

        headers = {
            "Authorization": f"Bearer {Filter_API_KEY}",
            "Accept": "application/json"
        }
        payload = {
            "model": Filter_MODEL_NAME,
            "messages": llm_messages,
            "temperature": 0.2,  # 降低温度以提高一致性
            "top_p": 0.95,
            "max_tokens": 1024,
            "stream": False
        }
        
        api_url = f"{Filter_API_BASE_URL}/chat/completions"

        try:
            response = requests.post(api_url, headers=headers, json=payload)
            response.raise_for_status() # 检查 HTTP 错误
            raw_llm_response = response.json()['choices'][0]['message']['content']
            logger.info(f"LLM 原始回复 (部分): {raw_llm_response[:200]}...")
        except requests.exceptions.RequestException as e:
            logger.error(f"调用 Filter API 失败: {e}")
            return [["error", f"Filter API call failed: {e}", "-"]]
        
        # 解析 LLM 响应
        parsed_todos_data = json_parser(raw_llm_response)
        
        if "error" in parsed_todos_data:
            logger.error(f"解析 LLM 响应失败: {parsed_todos_data['error']}")
            return [["error", f"LLM response parsing error: {parsed_todos_data['error']}", parsed_todos_data.get('raw_text', '')[:50] + "..."]]
        
        # 处理解析后的数据
        output_for_df = []
        
        # 如果是字典格式（符合prompt模板输出格式）
        if isinstance(parsed_todos_data, dict):
            # 获取消息ID对应的待办信息
            todo_info = None
            for key, value in parsed_todos_data.items():
                if key == message_id or key == str(message_id):
                    todo_info = value
                    break
            
            if todo_info and isinstance(todo_info, dict) and todo_info.get("is_todo", False):
                # 提取待办信息
                todo_content = todo_info.get("todo_content", "未指定待办内容")
                end_time = todo_info.get("end_time")
                location = todo_info.get("location")
                urgency = todo_info.get("urgency", "unimportant")
                
                # 准备合并显示的内容
                combined_content = todo_content
                
                # 添加截止时间
                if end_time and end_time != "null":
                    try:
                        date_part = end_time.split("T")[0] if "T" in end_time else end_time
                        combined_content += f" (截止时间: {date_part}"
                    except:
                        combined_content += f" (截止时间: {end_time}"
                else:
                    combined_content += " ("
                
                # 添加地点
                if location and location != "null":
                    combined_content += f", 地点: {location})"
                else:
                    combined_content += ")"
                
                # 添加紧急程度
                urgency_display = "一般"
                if urgency == "urgent":
                    urgency_display = "紧急"
                elif urgency == "important":
                    urgency_display = "重要"
                
                # 创建单行输出
                output_for_df = []
                output_for_df.append([1, combined_content, urgency_display])
            else:
                # 不是待办事项
                output_for_df = []
                output_for_df.append([1, "此消息不包含待办事项", "-"])
        
        # 如果是旧格式（列表格式）
        elif isinstance(parsed_todos_data, list):
            output_for_df = []
            
            # 检查列表是否为空
            if not parsed_todos_data:
                logger.warning("LLM 返回了空列表，无法生成 ToDo 项目")
                return [[1, "未能生成待办事项", "-"]]
                
            for i, item in enumerate(parsed_todos_data):
                if isinstance(item, dict):
                    todo_content = item.get('todo_content', item.get('content', 'N/A'))
                    status = item.get('status', '未完成')
                    urgency = item.get('urgency', 'normal')
                    
                    # 合并所有信息到一行
                    combined_content = todo_content
                    
                    # 添加截止时间
                    if 'end_time' in item and item['end_time']:
                        try:
                            if isinstance(item['end_time'], str):
                                date_part = item['end_time'].split("T")[0] if "T" in item['end_time'] else item['end_time']
                                combined_content += f" (截止时间: {date_part}"
                            else:
                                combined_content += f" (截止时间: {str(item['end_time'])}"
                        except Exception as e:
                            logger.warning(f"处理end_time时出错: {e}")
                            combined_content += " ("
                    else:
                        combined_content += " ("
                    
                    # 添加地点
                    if 'location' in item and item['location']:
                        combined_content += f", 地点: {item['location']})"
                    else:
                        combined_content += ")"
                    
                    # 设置重要等级
                    importance = "一般"
                    if urgency == "urgent":
                        importance = "紧急"
                    elif urgency == "important":
                        importance = "重要"
                    
                    output_for_df.append([i + 1, combined_content, importance])
                else:
                    # 如果不是字典，转换为字符串并添加到列表
                    try:
                        item_str = str(item) if item is not None else "未知项目"
                        output_for_df.append([i + 1, item_str, "一般"])
                    except Exception as e:
                        logger.warning(f"处理非字典项目时出错: {e}")
                        output_for_df.append([i + 1, "处理错误的项目", "一般"])
        
        if not output_for_df:
            logger.info("LLM 解析结果为空或无法转换为DataFrame格式。")
            return [["info", "未发现待办事项", "-"]]
        
        return output_for_df

    except Exception as e:
        logger.exception(f"调用 LLM 或解析时发生错误 (generate_todolist_from_text)")
        return [["error", f"LLM call/parse error: {str(e)}", "-"]]

#gradio
def process(audio, image, request: gr.Request):
    """处理语音和图片的示例函数"""
    # 获取并记录客户端IP
    client_ip = get_client_ip(request, True)
    print(f"Processing audio/image request from IP: {client_ip}")
    
    if audio is not None:
        sample_rate, audio_data = audio
        audio_info = f"音频采样率: {sample_rate}Hz, 数据长度: {len(audio_data)}"
    else:
        audio_info = "未收到音频"
    
    if image is not None:
        image_info = f"图片尺寸: {image.shape}"
    else:
        image_info = "未收到图片"
    
    return audio_info, image_info

def respond(

    message,

    history: list[tuple[str, str]],

    system_message,

    max_tokens,

    temperature,

    top_p,

    audio, # 多模态输入：音频

    image  # 多模态输入：图片

):
    # ... (聊天回复逻辑基本保持不变, 但确保 client 使用的是配置好的 HF client)
    # 1. 多模态处理接口 (其他人负责)
    # processed_text_from_multimodal = multimodal_placeholder_function(audio, image)
    # 多模态处理：调用讯飞API进行语音和图像识别
    multimodal_content = ""
    
    # 多模态处理配置已移至具体处理部分
    
    if audio is not None:
        try:
            audio_sample_rate, audio_data = audio
            multimodal_content += f"\n[音频信息: 采样率 {audio_sample_rate}Hz, 时长 {len(audio_data)/audio_sample_rate:.2f}秒]"
            
            # 调用Azure Speech语音识别
            azure_speech_config = get_hf_azure_speech_config()
            azure_speech_key = azure_speech_config.get('key')
            azure_speech_region = azure_speech_config.get('region')
            
            if azure_speech_key and azure_speech_region:
                import tempfile
                import soundfile as sf
                import os
                
                with tempfile.NamedTemporaryFile(suffix=".wav", delete=False) as temp_audio:
                    sf.write(temp_audio.name, audio_data, audio_sample_rate)
                    temp_audio_path = temp_audio.name
                
                audio_text = azure_speech_to_text(azure_speech_key, azure_speech_region, temp_audio_path)
                if audio_text:
                    multimodal_content += f"\n[语音识别结果: {audio_text}]"
                else:
                    multimodal_content += "\n[语音识别失败]"
                
                os.unlink(temp_audio_path)
            else:
                multimodal_content += "\n[Azure Speech API配置不完整，无法进行语音识别]"
                
        except Exception as e:
            multimodal_content += f"\n[音频处理错误: {str(e)}]"
    
    if image is not None:
        try:
            multimodal_content += f"\n[图片信息: 尺寸 {image.shape}]"
            
            # 调用讯飞图像识别
            if xunfei_appid and xunfei_apikey and xunfei_apisecret:
                import tempfile
                from PIL import Image
                import os
                
                with tempfile.NamedTemporaryFile(suffix=".jpg", delete=False) as temp_image:
                    if len(image.shape) == 3:  # RGB图像
                        pil_image = Image.fromarray(image.astype('uint8'), 'RGB')
                    else:  # 灰度图像
                        pil_image = Image.fromarray(image.astype('uint8'), 'L')
                    
                    pil_image.save(temp_image.name, 'JPEG')
                    temp_image_path = temp_image.name
                
                image_text = image_to_str(endpoint="https://ai-siyuwang5414995ai361208251338.cognitiveservices.azure.com/", key="45PYY2Av9CdMCveAjVG43MGKrnHzSxdiFTK9mWBgrOsMAHavxKj0JQQJ99BDACHYHv6XJ3w3AAAAACOGeVpQ", unused_param=None, file_path=temp_image_path)
                if image_text:
                    multimodal_content += f"\n[图像识别结果: {image_text}]"
                else:
                    multimodal_content += "\n[图像识别失败]"
                
                os.unlink(temp_image_path)
            else:
                multimodal_content += "\n[讯飞API配置不完整，无法进行图像识别]"
                
        except Exception as e:
            multimodal_content += f"\n[图像处理错误: {str(e)}]"
    
    # 将多模态内容（或其处理结果）与用户文本消息结合
    # combined_message = message
    # if multimodal_content: # 如果有多模态内容，则附加
    #     combined_message += "\n" + multimodal_content
    # 为了聊天模型的连贯性，聊天部分可能只使用原始 message
    # 而 ToDoList 生成则使用 combined_message

    # 聊天回复生成
    chat_messages = [{"role": "system", "content": system_message}]
    for val in history:
        if val[0]:
            chat_messages.append({"role": "user", "content": val[0]})
        if val[1]:
            chat_messages.append({"role": "assistant", "content": val[1]})
    chat_messages.append({"role": "user", "content": message}) # 聊天机器人使用原始消息

    chat_response_stream = ""
    if not Filter_API_BASE_URL or not Filter_API_KEY or not Filter_MODEL_NAME:
        logger.error("Filter API 配置不完整，无法调用 LLM333。")
        yield "Filter API 配置不完整，无法提供聊天回复。", []
        return

    headers = {
        "Authorization": f"Bearer {Filter_API_KEY}",
        "Accept": "application/json"
    }
    payload = {
        "model": Filter_MODEL_NAME,
        "messages": chat_messages,
        "temperature": temperature,
        "top_p": top_p,
        "max_tokens": max_tokens,
        "stream": True # 聊天通常需要流式输出
    }
    api_url = f"{Filter_API_BASE_URL}/chat/completions"

    try:
        response = requests.post(api_url, headers=headers, json=payload, stream=True)
        response.raise_for_status() # 检查 HTTP 错误
        
        for chunk in response.iter_content(chunk_size=None):
            if chunk:
                try:
                    # NVIDIA API 的流式输出是 SSE 格式，需要解析
                    # 每一行以 'data: ' 开头，后面是 JSON
                    for line in chunk.decode('utf-8').splitlines():
                        if line.startswith('data: '):
                            json_data = line[len('data: '):]
                            if json_data.strip() == '[DONE]':
                                break
                            data = json.loads(json_data)
                            # 检查 choices 列表是否存在且不为空
                            if 'choices' in data and len(data['choices']) > 0:
                                token = data['choices'][0]['delta'].get('content', '')
                                if token:
                                    chat_response_stream += token
                                    yield chat_response_stream, []
                except json.JSONDecodeError:
                    logger.warning(f"无法解析流式响应块: {chunk.decode('utf-8')}")
                except Exception as e:
                    logger.error(f"处理流式响应时发生错误: {e}")
                    yield chat_response_stream + f"\n\n错误: {e}", []

    except requests.exceptions.RequestException as e:
        logger.error(f"调用 NVIDIA API 失败: {e}")
        yield f"调用 NVIDIA API 失败: {e}", []

# 全局变量存储所有待办事项
all_todos_global = []

# 创建自定义的聊天界面
with gr.Blocks() as app:
    gr.Markdown("# ToDoAgent Multi-Modal Interface with ToDo List")
    
    with gr.Row():
        with gr.Column(scale=2):
            gr.Markdown("## Chat Interface")
            chatbot = gr.Chatbot(height=450, label="聊天记录", type="messages") # 推荐使用 type="messages"
            msg = gr.Textbox(label="输入消息", placeholder="输入您的问题或待办事项...")
            
            with gr.Row():
                audio_input = gr.Audio(label="上传语音", type="numpy", sources=["upload", "microphone"])
                image_input = gr.Image(label="上传图片", type="numpy")
            
            with gr.Accordion("高级设置", open=False):
                system_msg = gr.Textbox(value="You are a friendly Chatbot.", label="系统提示")
                max_tokens_slider = gr.Slider(minimum=1, maximum=4096, value=1024, step=1, label="最大生成长度(聊天)") # 增加聊天模型参数范围
                temperature_slider = gr.Slider(minimum=0.1, maximum=2.0, value=0.7, step=0.1, label="温度(聊天)")
                top_p_slider = gr.Slider(minimum=0.1, maximum=1.0, value=0.95, step=0.05, label="Top-p(聊天)")
            
            with gr.Row():
                submit_btn = gr.Button("发送", variant="primary")
                clear_btn = gr.Button("清除聊天和ToDo")

        with gr.Column(scale=1):
            gr.Markdown("## Generated ToDo List")
            todolist_df = gr.DataFrame(headers=["ID", "任务内容", "状态"], 
                                       datatype=["number", "str", "str"],
                                       row_count=(0, "dynamic"), 
                                       col_count=(3, "fixed"), 
                                       label="待办事项列表")

    def user(user_message, chat_history):
        # 将用户消息添加到聊天记录 (Gradio type="messages" 格式)
        if not chat_history: chat_history = []
        chat_history.append({"role": "user", "content": user_message})
        return "", chat_history
    
    def bot_interaction(chat_history, system_message, max_tokens, temperature, top_p, audio, image):
        user_message_for_chat = ""
        if chat_history and chat_history[-1]["role"] == "user":
            user_message_for_chat = chat_history[-1]["content"]
        
        # 准备用于 ToDoList 生成的输入文本 (多模态部分由其他人负责)
        text_for_todolist = user_message_for_chat
        # 可以在这里添加从 audio/image 提取文本的逻辑，并附加到 text_for_todolist
        # multimodal_text = process_multimodal_inputs(audio, image) # 假设的函数
        # if multimodal_text:
        #     text_for_todolist += "\n" + multimodal_text

        # 1. 生成聊天回复 (流式)
        # 转换 chat_history 从 [{'role':'user', 'content':'...'}, ...] 到 [('user_msg', 'bot_msg'), ...]
        # respond 函数期望的是 history: list[tuple[str, str]]
        # 但 Gradio type="messages" 的 chatbot.value 是 [{'role': ..., 'content': ...}, ...]
        # 需要转换
        formatted_history_for_respond = []
        temp_user_msg = None
        for item in chat_history[:-1]: #排除最后一条用户消息，因为它会作为当前message传入respond
            if item["role"] == "user":
                temp_user_msg = item["content"]
            elif item["role"] == "assistant" and temp_user_msg is not None:
                formatted_history_for_respond.append((temp_user_msg, item["content"]))
                temp_user_msg = None
            elif item["role"] == "assistant" and temp_user_msg is None: # Bot 先说话的情况
                formatted_history_for_respond.append(("", item["content"]))

        chat_stream_generator = respond(
            user_message_for_chat,
            formatted_history_for_respond, # 传递转换后的历史
            system_message,
            max_tokens,
            temperature,
            top_p,
            audio,
            image
        )

        full_chat_response = ""
        current_todos = [] 

        for chat_response_part, _ in chat_stream_generator:
            full_chat_response = chat_response_part
            # 更新 chat_history (Gradio type="messages" 格式)
            if chat_history and chat_history[-1]["role"] == "user":
                # 如果最后一条是用户消息，添加机器人回复
                # 但由于是流式，我们可能需要先添加一个空的 assistant 消息，然后更新它
                # 或者，等待流结束后一次性添加
                # 为了简化，我们先假设 respond 返回的是完整回复，或者在循环外更新
                pass # 流式更新 chatbot 在 submit_btn.click 中处理
            yield chat_history + [[None, full_chat_response]], current_todos # 临时做法，需要适配Gradio的流式更新
        
        # 流式结束后，更新 chat_history 中的最后一条 assistant 消息
        if chat_history and full_chat_response:
            # 查找最后一条用户消息，在其后添加或更新机器人回复
            # 这种方式对于 type="messages" 更友好
            # 实际上，Gradio 的 chatbot 更新应该在 .then() 中处理，这里先模拟
            # chat_history.append({"role": "assistant", "content": full_chat_response})
            # 这个 yield 应该在 submit_btn.click 的 .then() 中处理 chatbot 的更新
            # 这里我们先专注于 ToDo 生成
            pass # chatbot 更新由 Gradio 机制处理

        # 2. 聊天回复完成后，生成/更新 ToDoList
        if text_for_todolist:
            # 使用一个唯一的 ID，例如基于时间戳或随机数，如果需要区分不同输入的 ToDo
            message_id_for_todo = f"hf_app_{datetime.now().strftime('%Y%m%d%H%M%S%f')}"
            new_todo_items = generate_todolist_from_text(text_for_todolist, message_id_for_todo)
            current_todos = new_todo_items 
        
        # bot_interaction 应该返回 chatbot 的最终状态和 todolist_df 的数据
        # chatbot 的最终状态是 chat_history + assistant 的回复
        final_chat_history = list(chat_history) # 复制
        if full_chat_response:
             final_chat_history.append({"role": "assistant", "content": full_chat_response})

        yield final_chat_history, current_todos

    # 连接事件 (适配 type="messages")
    # Gradio 的流式更新通常是：
    # 1. user 函数准备输入，返回 (空输入框, 更新后的聊天记录)
    # 2. bot_interaction 函数是一个生成器，yield (部分聊天记录, 部分ToDo)
    # msg.submit 和 submit_btn.click 的 outputs 需要对应 bot_interaction 的 yield
    
    # 简化版，非流式更新 chatbot，流式更新由 respond 内部的 yield 控制
    # 但 respond 的 yield 格式 (str, list) 与 bot_interaction (list, list) 不同
    # 需要调整 respond 的 yield 或 bot_interaction 的处理

    # 调整后的事件处理，以更好地支持流式聊天和ToDo更新
    def process_filtered_result_for_todo(filtered_result, content, source_type):
        """处理过滤结果并生成todolist的辅助函数"""
        todos = []
        
        if isinstance(filtered_result, dict) and "error" in filtered_result:
            logger.error(f"{source_type} Filter 模块处理失败: {filtered_result['error']}")
            todos = [["Error", f"{source_type}: {filtered_result['error']}", "Filter Failed"]]
        elif isinstance(filtered_result, dict) and filtered_result.get("分类") == "其他":
            logger.info(f"{source_type}消息被 Filter 模块归类为 '其他'，不生成 ToDo List。")
            todos = [["Info", f"{source_type}: 消息被归类为 '其他'，无需生成 ToDo。", "Filtered"]]
        elif isinstance(filtered_result, list):
            # 处理列表类型的过滤结果
            category = None
            if filtered_result:
                for item in filtered_result:
                    if isinstance(item, dict) and "分类" in item:
                        category = item["分类"]
                        break
            
            if category == "其他":
                logger.info(f"{source_type}消息被 Filter 模块归类为 '其他'，不生成 ToDo List。")
                todos = [["Info", f"{source_type}: 消息被归类为 '其他'，无需生成 ToDo。", "Filtered"]]
            else:
                logger.info(f"{source_type}消息被 Filter 模块归类为 '{category if category else '未知'}'，继续生成 ToDo List。")
                if content:
                    msg_id_todo = f"hf_app_todo_{source_type}_{datetime.now().strftime('%Y%m%d%H%M%S%f')}"
                    todos = generate_todolist_from_text(content, msg_id_todo)
                    # 为每个todo添加来源标识
                    for todo in todos:
                        if len(todo) > 1:
                            todo[1] = f"[{source_type}] {todo[1]}"
        else:
            # 如果是字典但不是"其他"分类
            logger.info(f"{source_type}消息被 Filter 模块归类为 '{filtered_result.get('分类') if isinstance(filtered_result, dict) else '未知'}'，继续生成 ToDo List。")
            if content:
                msg_id_todo = f"hf_app_todo_{source_type}_{datetime.now().strftime('%Y%m%d%H%M%S%f')}"
                todos = generate_todolist_from_text(content, msg_id_todo)
                # 为每个todo添加来源标识
                for todo in todos:
                    if len(todo) > 1:
                        todo[1] = f"[{source_type}] {todo[1]}"
        
        return todos

    def handle_submit(user_msg_content, ch_history, sys_msg, max_t, temp, t_p, audio_f, image_f, request: gr.Request):
        global all_todos_global
        
        # 获取并记录客户端IP
        client_ip = get_client_ip(request, True)
        print(f"Processing request from IP: {client_ip}")
        
        # 首先处理多模态输入，获取多模态内容
        multimodal_text_content = ""
        # 添加调试日志
        logger.info(f"开始多模态处理 - 音频: {audio_f is not None}, 图像: {image_f is not None}")
        
        # 获取Azure Speech配置
        azure_speech_config = get_hf_azure_speech_config()
        azure_speech_key = azure_speech_config.get('key')
        azure_speech_region = azure_speech_config.get('region')
        
        # 添加调试日志
        logger.info(f"Azure Speech配置状态 - key: {bool(azure_speech_key)}, region: {bool(azure_speech_region)}")
        
        # 处理音频输入（使用Azure Speech服务）
        if audio_f is not None and azure_speech_key and azure_speech_region:
            logger.info("开始处理音频输入...")
            try:
                audio_sample_rate, audio_data = audio_f
                logger.info(f"音频信息: 采样率 {audio_sample_rate}Hz, 数据长度 {len(audio_data)}")
                
                # 保存音频为.wav文件
                audio_filename = os.path.join(SAVE_DIR, f"audio_{client_ip}.wav")
                save_audio(audio_f, audio_filename)
                logger.info(f"音频已保存: {audio_filename}")
                
                # 调用Azure Speech服务处理音频
                audio_text = azure_speech_to_text(azure_speech_key, azure_speech_region, audio_filename)
                logger.info(f"音频识别结果: {audio_text}")
                if audio_text:
                    multimodal_text_content += f"音频内容: {audio_text}"
                    logger.info("音频处理完成")
                else:
                    logger.warning("音频处理失败")
            except Exception as e:
                logger.error(f"音频处理错误: {str(e)}")
        elif audio_f is not None:
            logger.warning("音频文件存在但Azure Speech配置不完整，跳过音频处理")
        
        # 处理图像输入（使用Azure Computer Vision服务）
        if image_f is not None:
            logger.info("开始处理图像输入...")
            try:
                logger.info(f"图像信息: 形状 {image_f.shape}, 数据类型 {image_f.dtype}")
                
                # 保存图片为.jpg文件
                image_filename = os.path.join(SAVE_DIR, f"image_{client_ip}.jpg")
                save_image(image_f, image_filename)
                logger.info(f"图像已保存: {image_filename}")
                
                # 调用tools.py中的image_to_str方法处理图片
                image_text = image_to_str(endpoint="https://ai-siyuwang5414995ai361208251338.cognitiveservices.azure.com/", key="45PYY2Av9CdMCveAjVG43MGKrnHzSxdiFTK9mWBgrOsMAHavxKj0JQQJ99BDACHYHv6XJ3w3AAAAACOGeVpQ", unused_param=None, file_path=image_filename)
                logger.info(f"图像识别结果: {image_text}")
                if image_text:
                    if multimodal_text_content:  # 如果已有音频内容，添加分隔符
                        multimodal_text_content += "\n"
                    multimodal_text_content += f"图像内容: {image_text}"
                    logger.info("图像处理完成")
                else:
                    logger.warning("图像处理失败")
            except Exception as e:
                logger.error(f"图像处理错误: {str(e)}")
        elif image_f is not None:
            logger.warning("图像文件存在但处理失败，跳过图像处理")
        
        # 确定最终的用户输入内容：如果用户没有输入文本，使用多模态识别的内容
        final_user_content = user_msg_content.strip() if user_msg_content else ""
        if not final_user_content and multimodal_text_content:
            final_user_content = multimodal_text_content
            logger.info(f"用户无文本输入，使用多模态内容作为用户输入: {final_user_content}")
        elif final_user_content and multimodal_text_content:
            # 用户有文本输入，多模态内容作为补充
            final_user_content = f"{final_user_content}\n{multimodal_text_content}"
            logger.info(f"用户有文本输入，多模态内容作为补充")
        
        # 如果最终还是没有任何内容，提供默认提示
        if not final_user_content:
            final_user_content = "[无输入内容]"
            logger.warning("用户没有提供任何输入内容（文本、音频或图像）")
        
        logger.info(f"最终用户输入内容: {final_user_content}")
        
        # 1. 更新聊天记录 (用户部分) - 使用最终确定的用户内容
        if not ch_history: ch_history = []
        ch_history.append({"role": "user", "content": final_user_content})
        yield ch_history, [] # 更新聊天，ToDo 列表暂时不变

        # 2. 流式生成机器人回复并更新聊天记录
        # 转换 chat_history 为 respond 函数期望的格式
        formatted_hist_for_respond = []
        temp_user_msg_for_hist = None
        # 使用 ch_history[:-1] 因为当前用户消息已在 ch_history 中
        for item_hist in ch_history[:-1]: 
            if item_hist["role"] == "user":
                temp_user_msg_for_hist = item_hist["content"]
            elif item_hist["role"] == "assistant" and temp_user_msg_for_hist is not None:
                formatted_hist_for_respond.append((temp_user_msg_for_hist, item_hist["content"]))
                temp_user_msg_for_hist = None
            elif item_hist["role"] == "assistant" and temp_user_msg_for_hist is None:
                formatted_hist_for_respond.append(("", item_hist["content"]))
        
        # 准备一个 assistant 消息的槽位
        ch_history.append({"role": "assistant", "content": ""})

        full_bot_response = ""
        # 使用最终确定的用户内容进行对话
        for bot_response_token, _ in respond(final_user_content, formatted_hist_for_respond, sys_msg, max_t, temp, t_p, audio_f, image_f):
            full_bot_response = bot_response_token
            ch_history[-1]["content"] = full_bot_response # 更新最后一条 assistant 消息
            yield ch_history, [] # 流式更新聊天，ToDo 列表不变

        # 3. 生成 ToDoList - 分别处理音频、图片和文字输入
        new_todos_list = []
        
        # 分别处理文字输入
        if user_msg_content.strip():
            logger.info(f"处理文字输入生成ToDo: {user_msg_content.strip()}")
            text_filtered_result = filter_message_with_llm(user_msg_content.strip())
            text_todos = process_filtered_result_for_todo(text_filtered_result, user_msg_content.strip(), "文字")
            new_todos_list.extend(text_todos)
        
        # 分别处理音频输入
        if audio_f is not None and azure_speech_key and azure_speech_region:
            try:
                audio_sample_rate, audio_data = audio_f
                audio_filename = os.path.join(SAVE_DIR, f"audio_{client_ip}.wav")
                save_audio(audio_f, audio_filename)
                audio_text = azure_speech_to_text(azure_speech_key, azure_speech_region, audio_filename)
                if audio_text:
                    logger.info(f"处理音频输入生成ToDo: {audio_text}")
                    audio_filtered_result = filter_message_with_llm(audio_text)
                    audio_todos = process_filtered_result_for_todo(audio_filtered_result, audio_text, "音频")
                    new_todos_list.extend(audio_todos)
            except Exception as e:
                logger.error(f"音频处理错误: {str(e)}")
        
        # 分别处理图片输入
        if image_f is not None:
            try:
                image_filename = os.path.join(SAVE_DIR, f"image_{client_ip}.jpg")
                save_image(image_f, image_filename)
                image_text = image_to_str(endpoint="https://ai-siyuwang5414995ai361208251338.cognitiveservices.azure.com/", key="45PYY2Av9CdMCveAjVG43MGKrnHzSxdiFTK9mWBgrOsMAHavxKj0JQQJ99BDACHYHv6XJ3w3AAAAACOGeVpQ", unused_param=None, file_path=image_filename)
                if image_text:
                    logger.info(f"处理图片输入生成ToDo: {image_text}")
                    image_filtered_result = filter_message_with_llm(image_text)
                    image_todos = process_filtered_result_for_todo(image_filtered_result, image_text, "图片")
                    new_todos_list.extend(image_todos)
            except Exception as e:
                logger.error(f"图片处理错误: {str(e)}")
        
        # 如果没有任何有效输入，使用原有逻辑
        if not new_todos_list and final_user_content:
            logger.info(f"使用整合内容生成ToDo: {final_user_content}")
            filtered_result = filter_message_with_llm(final_user_content)
            
            if isinstance(filtered_result, dict) and "error" in filtered_result:
                logger.error(f"Filter 模块处理失败: {filtered_result['error']}")
                # 可以选择在这里显示错误信息给用户
                new_todos_list = [["Error", filtered_result['error'], "Filter Failed"]]
            elif isinstance(filtered_result, dict) and filtered_result.get("分类") == "其他":
                logger.info(f"消息被 Filter 模块归类为 '其他'，不生成 ToDo List。")
                new_todos_list = [["Info", "消息被归类为 '其他'，无需生成 ToDo。", "Filtered"]]
            elif isinstance(filtered_result, list):
                # 如果返回的是列表，尝试从列表中获取分类信息
                category = None
                
                # 检查列表是否为空
                if not filtered_result:
                    logger.warning("Filter 模块返回了空列表，将继续生成 ToDo List。")
                    if final_user_content:
                        msg_id_todo = f"hf_app_todo_{datetime.now().strftime('%Y%m%d%H%M%S%f')}"
                        new_todos_list = generate_todolist_from_text(final_user_content, msg_id_todo)
                    # 将新的待办事项添加到全局列表中
                    if new_todos_list and not (len(new_todos_list) == 1 and "Info" in str(new_todos_list[0])):
                        # 重新分配ID以确保连续性
                        for i, todo in enumerate(new_todos_list):
                            todo[0] = len(all_todos_global) + i + 1
                        all_todos_global.extend(new_todos_list)
                    yield ch_history, all_todos_global
                    return
                
                # 确保列表中至少有一个元素且是字典类型
                valid_item = None
                for item in filtered_result:
                    if isinstance(item, dict):
                        valid_item = item
                        if "分类" in item:
                            category = item["分类"]
                            break
                
                # 如果没有找到有效的字典元素，记录警告并继续生成ToDo
                if valid_item is None:
                    logger.warning(f"Filter 模块返回的列表中没有有效的字典元素: {filtered_result}")
                    if final_user_content:
                        msg_id_todo = f"hf_app_todo_{datetime.now().strftime('%Y%m%d%H%M%S%f')}"
                        new_todos_list = generate_todolist_from_text(final_user_content, msg_id_todo)
                    # 将新的待办事项添加到全局列表中
                    if new_todos_list and not (len(new_todos_list) == 1 and "Info" in str(new_todos_list[0])):
                        # 重新分配ID以确保连续性
                        for i, todo in enumerate(new_todos_list):
                            todo[0] = len(all_todos_global) + i + 1
                        all_todos_global.extend(new_todos_list)
                    yield ch_history, all_todos_global
                    return
            
                if category == "其他":
                    logger.info(f"消息被 Filter 模块归类为 '其他'，不生成 ToDo List。")
                    new_todos_list = [["Info", "消息被归类为 '其他'，无需生成 ToDo。", "Filtered"]]
                else:
                    logger.info(f"消息被 Filter 模块归类为 '{category if category else '未知'}'，继续生成 ToDo List。")
                    # 如果 Filter 结果不是"其他"，则继续生成 ToDoList
                    if final_user_content:
                        msg_id_todo = f"hf_app_todo_{datetime.now().strftime('%Y%m%d%H%M%S%f')}"
                        new_todos_list = generate_todolist_from_text(final_user_content, msg_id_todo)
            else:
                # 如果是字典但不是"其他"分类
                logger.info(f"消息被 Filter 模块归类为 '{filtered_result.get('分类')}'，继续生成 ToDo List。")
                # 如果 Filter 结果不是"其他"，则继续生成 ToDoList
                if final_user_content:
                    msg_id_todo = f"hf_app_todo_{datetime.now().strftime('%Y%m%d%H%M%S%f')}"
                    new_todos_list = generate_todolist_from_text(final_user_content, msg_id_todo)
        
        # 将新的待办事项添加到全局列表中（排除信息性消息）
        if new_todos_list and not (len(new_todos_list) == 1 and ("Info" in str(new_todos_list[0]) or "Error" in str(new_todos_list[0]))):
            # 重新分配ID以确保连续性
            for i, todo in enumerate(new_todos_list):
                todo[0] = len(all_todos_global) + i + 1
            all_todos_global.extend(new_todos_list)
        
        yield ch_history, all_todos_global # 最终更新聊天和完整的ToDo列表

    submit_btn.click(
        handle_submit, 
        [msg, chatbot, system_msg, max_tokens_slider, temperature_slider, top_p_slider, audio_input, image_input], 
        [chatbot, todolist_df] 
    )
    msg.submit(
        handle_submit, 
        [msg, chatbot, system_msg, max_tokens_slider, temperature_slider, top_p_slider, audio_input, image_input], 
        [chatbot, todolist_df]
    )
    
    def clear_all():
        global all_todos_global
        all_todos_global = []  # 清除全局待办事项列表
        return None, None, "" # 清除 chatbot, todolist_df, 和 msg 输入框
    clear_btn.click(clear_all, None, [chatbot, todolist_df, msg], queue=False)
    
    # 旧的 Audio/Image Processing Tab (保持不变或按需修改)
    with gr.Tab("Audio/Image Processing (Original)"):
        gr.Markdown("## 处理音频和图片")
        audio_processor = gr.Audio(label="上传音频", type="numpy")
        image_processor = gr.Image(label="上传图片", type="numpy")
        process_btn = gr.Button("处理", variant="primary")
        audio_output = gr.Textbox(label="音频信息")
        image_output = gr.Textbox(label="图片信息")
        
        process_btn.click(
            process,
            inputs=[audio_processor, image_processor],
            outputs=[audio_output, image_output]
        )

if __name__ == "__main__":
    app.launch(debug=True)