chatbot.py

import gradio as gr
import json
import requests
import re

class Chatbot:
    def __init__(self, config):
        self.video_id = config.get('video_id')
        self.content_subject = config.get('content_subject')
        self.content_grade = config.get('content_grade')
        self.jutor_chat_key = config.get('jutor_chat_key')
        self.transcript_text = self.get_transcript_text(config.get('transcript'))
        self.key_moments_text = self.get_key_moments_text(config.get('key_moments'))
        self.ai_model_name = config.get('ai_model_name')
        self.ai_client = config.get('ai_client')
        self.instructions = config.get('instructions')

    def get_transcript_text(self, transcript_data):
        if isinstance(transcript_data, str):
            transcript_json = json.loads(transcript_data)
        else:
            transcript_json = transcript_data
        for entry in transcript_json:
            entry.pop('end_time', None)
        transcript_text = json.dumps(transcript_json, ensure_ascii=False)
        return transcript_text
    
    def get_key_moments_text(self, key_moments_data):
        if isinstance(key_moments_data, str):
            key_moments_json = json.loads(key_moments_data)
        else:
            key_moments_json = key_moments_data
        # key_moments_json remove images
        for moment in key_moments_json:
            moment.pop('images', None)
            moment.pop('end', None)
            moment.pop('transcript', None)
            
        key_moments_text = json.dumps(key_moments_json, ensure_ascii=False)
        return key_moments_text


    def chat(self, user_message, chat_history):
        try:
            messages = self.prepare_messages(chat_history, user_message)
            system_prompt = self.instructions
            system_prompt += "\n\n告知用戶你現在是誰，第一次加上科目學伴及名字，後面就只說名字就好，但不用每次都說，自然就好，不用每一句都特別說明，口氣請符合給予的人設，請用繁體中文回答"
            service_type = self.ai_model_name
            response_text = self.chat_with_service(service_type, system_prompt, messages)

        except Exception as e:
            print(f"Error: {e}")
            response_text = "學習精靈有點累，請稍後再試！"

        return response_text

    def prepare_messages(self, chat_history, user_message):
        messages = []
        if chat_history is not None:
            if len(chat_history) > 10:
                chat_history = chat_history[-10:]

            for user_msg, assistant_msg in chat_history:
                if user_msg:
                    messages.append({"role": "user", "content": user_msg})
                if assistant_msg:
                    messages.append({"role": "assistant", "content": assistant_msg})
                
        if user_message:
            user_message += "/n (請一定要用繁體中文回答 zh-TW，並用台灣人的禮貌口語表達，回答時不要特別說明這是台灣人的語氣，不要提到「台灣腔」，不用提到「逐字稿」這個詞，用「內容」代替)，回答時如果有用到數學式，請用數學符號代替純文字（Latex 用 $ 字號 render)"
            messages.append({"role": "user", "content": user_message})
        return messages

    def chat_with_service(self, service_type, system_prompt, messages):
        if service_type == 'openai':
            return self.chat_with_jutor(system_prompt, messages)
        elif service_type == 'groq_llama3':
            return self.chat_with_groq(service_type, system_prompt, messages)
        elif service_type == 'groq_mixtral':
            return self.chat_with_groq(service_type, system_prompt, messages)
        elif service_type == 'claude3':
            return self.chat_with_claude3(system_prompt, messages)
        elif service_type in ['perplexity_sonar', 'perplexity_sonar_pro', 'perplexity_r1_1776']:
            return self.chat_with_perplexity(service_type, system_prompt, messages)
        else:
            raise gr.Error("不支持的服务类型")

    def chat_with_jutor(self, system_prompt, messages):
        messages.insert(0, {"role": "system", "content": system_prompt})
        api_endpoint = "https://ci-live-feat-video-ai-dot-junyiacademy.appspot.com/api/v2/jutor/hf-chat"
        headers = {
            "Content-Type": "application/json",
            "x-api-key": self.jutor_chat_key,
        }
        model = "gpt-4o"
        print("======model======")
        print(model)
        data = {
            "data": {
                "messages": messages,
                "max_tokens": 512,
                "temperature": 0.9,
                "model": model,
                "stream": False,
            }
        }

        response = requests.post(api_endpoint, headers=headers, data=json.dumps(data))
        response_data = response.json()
        response_completion = response_data['data']['choices'][0]['message']['content'].strip()
        return response_completion

    def chat_with_groq(self, model_name, system_prompt, messages):
        # system_prompt insert to messages 的最前面 {"role": "system", "content": system_prompt}
        messages.insert(0, {"role": "system", "content": system_prompt})
        model_name_dict = {
            "groq_llama3": "llama-3.1-70b-versatile",
            "groq_mixtral": "mixtral-8x7b-32768"
        }
        model = model_name_dict.get(model_name)
        print("======model======")
        print(model)

        request_payload = {
            "model": model,
            "messages": messages,
            "max_tokens": 500  # 設定一個較大的值，可根據需要調整
        }
        groq_client = self.ai_client
        response = groq_client.chat.completions.create(**request_payload)
        response_completion = response.choices[0].message.content.strip()
        return response_completion

    def chat_with_claude3(self, system_prompt, messages):
        if not system_prompt.strip():
            raise ValueError("System prompt cannot be empty")
        
        model_id = "anthropic.claude-3-sonnet-20240229-v1:0"
        # model_id = "anthropic.claude-3-haiku-20240307-v1:0"
        print("======model_id======")
        print(model_id)
        kwargs = {
            "modelId": model_id,
            "contentType": "application/json",
            "accept": "application/json",
            "body": json.dumps({
                "anthropic_version": "bedrock-2023-05-31",
                "max_tokens": 500,
                "system": system_prompt,
                "messages": messages
            })
        }
        # 建立 message API，讀取回應
        bedrock_client = self.ai_client
        response = bedrock_client.invoke_model(**kwargs)
        response_body = json.loads(response.get('body').read())
        response_completion = response_body.get('content')[0].get('text').strip()
        return response_completion

    def chat_with_perplexity(self, service_type, system_prompt, messages):
        """使用 Perplexity API 進行對話"""
        if not system_prompt.strip():
            raise ValueError("System prompt cannot be empty")
        
        # 清理用戶訊息中的特殊指令
        for msg in messages:
            if msg["role"] == "user":
                # 移除可能導致問題的特殊指令
                msg["content"] = msg["content"].replace("/n", "\n")
                # 移除括號內的特殊指令
                msg["content"] = re.sub(r'\(請一定要用繁體中文回答.*?\)', '', msg["content"])
        
        # 系統提示放在最前面
        clean_messages = [{"role": "system", "content": system_prompt}]
        # 添加其他訊息
        for msg in messages:
            if msg["role"] != "system":  # 避免重複添加系統提示
                clean_messages.append(msg)
        
        # 在系統提示中添加 Markdown 和 LaTeX 格式指導
        system_prompt += "\n\n重要：使用 LaTeX 數學符號時，請確保格式正確。數學表達式應該使用 $ 符號包圍，例如：$7 \\times 10^4$。不要使用 ** 符號來強調數字，而是使用 $ 符號，例如：$7$個萬 ($7 \\times 10000$)。不要使用 \\text 或 \\quad 等命令。"

        # 根據服務類型選擇模型
        model_name_dict = {
            "perplexity_sonar": "sonar",
            "perplexity_sonar_pro": "sonar-pro",
            "perplexity_r1_1776": "r1-1776"
        }
        model = model_name_dict.get(service_type, "sonar")
        
        print("======model======")
        print(model)
        print("======clean_messages======")
        print(json.dumps(clean_messages[:1], ensure_ascii=False))  # 只打印系統提示的前部分
        
        try:
            perplexity_client = self.ai_client
            
            # 針對 r1-1776 模型調整參數
            if service_type == "perplexity_r1_1776":
                # 增加 max_tokens 並添加特殊指令
                response = perplexity_client.chat.completions.create(
                    model=model,
                    messages=clean_messages,
                    max_tokens=1000,  # 增加 token 限制
                    temperature=0.7,
                    top_p=0.9
                )
            else:
                response = perplexity_client.chat.completions.create(
                    model=model,
                    messages=clean_messages,
                    max_tokens=500,
                    temperature=0.7,
                    top_p=0.9
                )
            
            # 檢查回應是否為空
            if not hasattr(response, 'choices') or len(response.choices) == 0:
                print("警告：API 回傳無效回應結構")
                return "學習精靈暫時無法回答，請稍後再試！"
            
            response_completion = response.choices[0].message.content
            if not response_completion or response_completion.strip() == "":
                print("警告：API 回傳空回應")
                return "學習精靈暫時無法回答，請稍後再試！"
            
            # 處理回應中的思考過程標籤和修正 LaTeX 格式
            response_completion = self._process_response(response_completion)
            
            # 打印處理後的回應以便調試
            print("======processed_response======")
            print(response_completion)
            
            return response_completion.strip()
            
        except Exception as e:
            print(f"Perplexity API Error: {e}")
            print(f"Error details: {str(e)}")
            # 嘗試使用備用模型
            try:
                if service_type == "perplexity_r1_1776":
                    print("嘗試使用備用模型 sonar")
                    backup_response = perplexity_client.chat.completions.create(
                        model="sonar",
                        messages=clean_messages,
                        max_tokens=500,
                        temperature=0.7
                    )
                    backup_completion = backup_response.choices[0].message.content
                    backup_completion = self._process_response(backup_completion)
                    return backup_completion.strip()
            except Exception as backup_error:
                print(f"備用模型也失敗: {backup_error}")
            
            return "學習精靈暫時無法回答，請稍後再試！"

    def _process_response(self, response_text):
        """處理回應中的思考過程標籤和修正 LaTeX 格式"""
        # 移除 <think>...</think> 區塊
        import re
        response_text = re.sub(r'<think>.*?</think>', '', response_text, flags=re.DOTALL)
        
        # 移除其他可能的標籤或指令
        response_text = re.sub(r'（偷偷說.*?）', '', response_text, flags=re.DOTALL)
        
        # 修正 Markdown 格式
        # 1. 確保項目符號前後有正確的空格和換行
        response_text = re.sub(r'(\n|^)(\s*)([-•○●◦])\s*', r'\1\2\3 ', response_text)
        
        # 2. 確保數字列表前後有正確的空格和換行
        response_text = re.sub(r'(\n|^)(\s*)(\d+\.)\s*', r'\1\2\3 ', response_text)
        
        # 3. 修正 LaTeX 格式
        # 移除不正確的 LaTeX 命令
        response_text = re.sub(r'\\text\{([^}]+)\}', r'\1', response_text)
        response_text = re.sub(r'\\quad', ' ', response_text)
        
        # 4. 修正數學表達式
        # 確保數學表達式中的乘法符號格式正確
        response_text = re.sub(r'(\d+)個「([^」]+)」→\s*(\d+)\\times(\d+)', r'\1個「\2」→ $\3\\times\4$', response_text)
        
        # 5. 修正單獨數字的 LaTeX 格式
        # 將單獨的數字包裹在 $ 符號中
        response_text = re.sub(r'([^$\d])(\d+)([^$\d\w])', r'\1$\2$\3', response_text)
        
        # 6. 修正連續的 LaTeX 表達式
        # 確保連續的 LaTeX 表達式之間有空格
        response_text = re.sub(r'\$([^$]+)\$\$([^$]+)\$', r'$\1$ $\2$', response_text)
        
        # 7. 移除單獨的 $ 符號
        response_text = re.sub(r'(?<!\$)\$(?!\$)\s*$', '', response_text)
        response_text = re.sub(r'^\s*\$(?!\$)', '', response_text)
        response_text = re.sub(r'(?<!\$)\$(?!\$)\s*\n', '\n', response_text)
        
        # 8. 確保成對的 $ 符號
        dollar_count = response_text.count('$')
        if dollar_count % 2 != 0:
            # 如果 $ 符號數量為奇數，移除最後一個 $
            last_dollar_pos = response_text.rfind('$')
            if last_dollar_pos != -1:
                response_text = response_text[:last_dollar_pos] + response_text[last_dollar_pos+1:]
        
        # 9. 修正錯誤的粗體標記
        # 將 **數字** 格式修正為正確的數字格式
        response_text = re.sub(r'\*\*(\d+)\*\*', r'$\1$', response_text)
        
        # 如果處理後的回應為空，返回原始回應
        if not response_text.strip():
            return "學習精靈暫時無法回答，請稍後再試！"
        
        return response_text