Merge branch 'PERPLEXITY_CLIENT'

app.py

@@ -73,6 +73,7 @@ from google.oauth2.service_account import Credentials
 import vertexai
 from vertexai.generative_models import GenerativeModel, Part
 
+
 # import boto3
 
 from chatbot import Chatbot
@@ -96,6 +97,7 @@ if is_env_local:
         OPEN_AI_ASSISTANT_ID_GPT4_BOT1 = config["OPEN_AI_ASSISTANT_ID_GPT4_BOT1"]
         OPEN_AI_ASSISTANT_ID_GPT3_BOT1 = config["OPEN_AI_ASSISTANT_ID_GPT3_BOT1"]
         GROQ_API_KEY = config["GROQ_API_KEY"]
+        PERPLEXITY_API_KEY = config["PERPLEXITY_API_KEY"]
         JUTOR_CHAT_KEY = config["JUTOR_CHAT_KEY"]
         AWS_ACCESS_KEY = config["AWS_ACCESS_KEY"]
         AWS_SECRET_KEY = config["AWS_SECRET_KEY"]
@@ -112,6 +114,7 @@ else:
     OPEN_AI_ASSISTANT_ID_GPT4_BOT1 = os.getenv("OPEN_AI_ASSISTANT_ID_GPT4_BOT1")
     OPEN_AI_ASSISTANT_ID_GPT3_BOT1 = os.getenv("OPEN_AI_ASSISTANT_ID_GPT3_BOT1")
     GROQ_API_KEY = os.getenv("GROQ_API_KEY")
+    PERPLEXITY_API_KEY = os.getenv("PERPLEXITY_API_KEY")
     JUTOR_CHAT_KEY = os.getenv("JUTOR_CHAT_KEY")
     AWS_ACCESS_KEY = os.getenv("AWS_ACCESS_KEY")
     AWS_SECRET_KEY = os.getenv("AWS_SECRET_KEY")
@@ -139,12 +142,7 @@ GBQ_CLIENT = bigquery.Client.from_service_account_info(json.loads(GBQ_KEY))
 GROQ_CLIENT = Groq(api_key=GROQ_API_KEY)
 GCS_SERVICE = GoogleCloudStorage(GCS_KEY)
 GCS_CLIENT = GCS_SERVICE.client
-# BEDROCK_CLIENT = boto3.client(
-#     service_name="bedrock-runtime",
-#     aws_access_key_id=AWS_ACCESS_KEY,
-#     aws_secret_access_key=AWS_SECRET_KEY,
-#     region_name=AWS_REGION_NAME,
-# )
+PERPLEXITY_CLIENT = OpenAI(api_key=PERPLEXITY_API_KEY, base_url="https://api.perplexity.ai")
 
 # check open ai access
 def check_open_ai_access(open_ai_api_key):
@@ -2369,7 +2367,7 @@ def get_instructions(content_subject, content_grade, transcript_text, key_moment
         grade: {content_grade}
         context: {key_moments}
         transcript_text: {transcript_text}
-        Assistant Role: you are a {content_subject} assistant. you can call yourself as  {content_subject} 學伴
+        Assistant Role: you are a {content_subject} assistant. you can call yourself as  {content_subject} 學伴 and your name if you know
         User Role: {content_grade} th-grade student.
         Method: {method}
         Language: Traditional Chinese ZH-TW (it's very important), suitable for {content_grade} th-grade level.
@@ -2582,23 +2580,18 @@ def get_chatbot_config(ai_name, transcript_state, key_moments, content_subject,
     ai_name_clients_model = {
         "foxcat": {
             "ai_name": "foxcat",
-            "ai_client": GROQ_CLIENT,
-            "ai_model_name": "groq_llama3",
+            "ai_client": PERPLEXITY_CLIENT,
+            "ai_model_name": "perplexity_sonar",
         },
-        # "lili": {
-        #     "ai_name": "lili",
-        #     "ai_client": BEDROCK_CLIENT,
-        #     "ai_model_name": "claude3",
-        # },
         "lili": {
             "ai_name": "lili",
-            "ai_client": GROQ_CLIENT,
-            "ai_model_name": "groq_llama3",
+            "ai_client": PERPLEXITY_CLIENT,
+            "ai_model_name": "perplexity_r1_1776",
         },
         "maimai": {
             "ai_name": "maimai",
-            "ai_client": GROQ_CLIENT,
-            "ai_model_name": "groq_mixtral",
+            "ai_client": PERPLEXITY_CLIENT,
+            "ai_model_name": "perplexity_r1_1776",
         }
     }
     ai_client = ai_name_clients_model.get(ai_name, "foxcat")["ai_client"]
@@ -3343,7 +3336,7 @@ def create_app():
                         with gr.Accordion("🦄 飛特精靈 敘述", open=False):
                             vaitor_chatbot_description_value = gr.Markdown(value=vaitor_chatbot_description, visible=True)
                     # 狐狸貓
-                    with gr.Column(scale=1, variant="panel", visible=False):
+                    with gr.Column(scale=1, variant="panel", visible=True):
                         foxcat_chatbot_avatar_url = "https://storage.googleapis.com/wpassets.junyiacademy.org/1/2020/06/%E7%A7%91%E5%AD%B8%E5%BE%BD%E7%AB%A0-2-150x150.png"
                         foxcat_avatar_images = gr.State([user_avatar, foxcat_chatbot_avatar_url])
                         foxcat_chatbot_description = """Hi，我是【狐狸貓】，可以陪你一起學習本次的內容，有什麼問題都可以問我喔！\n
@@ -3358,7 +3351,7 @@ def create_app():
                         with gr.Accordion("💜 狐狸貓 敘述", open=False):
                             foxcat_chatbot_description_value = gr.Markdown(value=foxcat_chatbot_description, visible=True)
                     # 梨梨
-                    with gr.Column(scale=1, variant="panel", visible=False): 
+                    with gr.Column(scale=1, variant="panel", visible=True): 
                         lili_chatbot_avatar_url = "https://junyitopicimg.s3.amazonaws.com/live/v1283-new-topic-44-icon.png"
                         lili_avatar_images = gr.State([user_avatar, lili_chatbot_avatar_url])
                         lili_chatbot_description = """你好，我是溫柔的【梨梨】，很高興可以在這裡陪伴你學習。���果你有任何疑問，請隨時向我提出哦！ \n
@@ -3377,7 +3370,7 @@ def create_app():
                         with gr.Accordion("🧡 梨梨 敘述", open=False):
                             lili_chatbot_description_value = gr.Markdown(value=lili_chatbot_description, visible=True)
                     # 麥麥
-                    with gr.Column(scale=1, variant="panel", visible=False): 
+                    with gr.Column(scale=1, variant="panel", visible=True): 
                         maimai_chatbot_avatar_url = "https://storage.googleapis.com/wpassets.junyiacademy.org/1/2020/07/%E6%80%9D%E8%80%83%E5%8A%9B%E8%B6%85%E4%BA%BA%E5%BE%BD%E7%AB%A0_%E5%B7%A5%E4%BD%9C%E5%8D%80%E5%9F%9F-1-%E8%A4%87%E6%9C%AC-150x150.png"
                         maimai_avatar_images = gr.State([user_avatar, maimai_chatbot_avatar_url])
                         maimai_chatbot_description = """Hi，我是迷人的【麥麥】，我在這裡等著和你一起探索新知，任何疑問都可以向我提出！\n

chatbot.py

@@ -1,6 +1,7 @@
 import gradio as gr
 import json
 import requests
+import re
 
 class Chatbot:
     def __init__(self, config):
@@ -43,8 +44,10 @@ class Chatbot:
         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！"
@@ -77,6 +80,8 @@ class Chatbot:
             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")
 
@@ -90,7 +95,6 @@ class Chatbot:
         model = "gpt-4o"
         print("======model======")
         print(model)
-        # model = "gpt-3.5-turbo-0125"
         data = {
             "data": {
                 "messages": messages,
@@ -152,3 +156,158 @@ class Chatbot:
         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
+