元旦快乐

2024.01.01:
openai token计算✓
优化ReAct Agent 与 LLM Chain自动切换逻辑✓
*添加双记忆机制：长时记忆流及短时聊天记忆✓
修复record.py asr bug✓
提高远程音频（android 连接器）的稳定性✓
修复执行时间计算bug✓
优化语音输出逻辑✓

README.md

@@ -10,8 +10,7 @@
 
 
 
-
+**请先想明白**
-**12月迟来的报到，Fay数字人 AI Agent版（含智慧农业箱的操作demo代码，如果你需要完整代码可以公众号留言申请获取）第5版正式上传！**
 
 如果你需要是一个线上线下的销售员，请移步[`带货完整版`](https://github.com/TheRamU/Fay/tree/fay-sales-edition)                       
 
@@ -19,13 +18,25 @@
 
 
 
-***“优秀的产品都值得用数字人从新做一遍”***
+***然后，“优秀的产品都值得用数字人从新做一遍”***
+1、基于日程维护的助理模式：执行及维护你的日程，绝不是一个简单的闹钟
+<img src="images/you1.png" alt="Fay">
 
-亮点：计划任务主动执行，无需一问一答，自动规划及调用agent tool去完成工作；使用open ai tts；使用向量数据库实现永久记忆及记忆检索；
+2、强大的规划执行（ReAct）能力：规划->执行<->反思->总结
+<img src="images/you2.png" alt="Fay">
 
-![](images/agent_demo.gif)
+3、LLM Chain与React Agent自动切换：保留规划执行能力的同时兼顾聊天能力（还需优化）
+<img src="images/you3.png" alt="Fay">
+
+4、双记忆机制：斯坦福AI小镇的记忆流（时间、重要性、相关度）实现长时记忆，邻近对话记忆实现连贯对话
+<img src="images/you4.png" alt="Fay">
+
+5、易于扩展的agent 工具
+<img src="images/you5.png" alt="Fay">
+
+6、配套24小时后台运行的android 连接器
+<img src="images/you6.png" alt="Fay">
 
-​                                                                       (上图：实测ReAct能力）
 
 ## **安装说明**
 
@@ -62,20 +73,23 @@ python main.py
 + 仓库地址：https://github.com/xszyou/fay-android
 
 ### **更新日志**
+2024.01.01:
+openai token计算✓
+优化ReAct Agent 与 LLM Chain自动切换逻辑✓
+*添加双记忆机制：长时记忆流及短时聊天记忆✓
+修复record.py asr bug✓
+提高远程音频（android 连接器）的稳定性✓
+修复执行时间计算bug✓
+优化语音输出逻辑✓
+
 2023.12.25:
 
 *实现agent ReAct与LLM chain自动切换逻辑✓
-
 聊天窗区分任务消息✓
-
 修复删除日程bug✓
-
 优化远程音频逻辑✓
-
 等待处理引入加载中效果✓
-
 优化prompt以解决日程任务递归调用问题✓
-
 修复一次性日程清除的bug✓
 
 

README_EN.md

@@ -10,20 +10,30 @@
 
 
 
-
+**Please Understand First**
-**Belated December announcement, the 5th edition of Fay Digital Human AI Agent Version (complete code for smart agriculture box can be requested via our public channel) is officially uploaded!**
 
 If you need an online and offline salesperson, please go to [`Complete Retail Version`](https://github.com/TheRamU/Fay/tree/fay-sales-edition)
 
 If you need a digital human assistant for human-computer interaction (and yes, you can command it to switch devices on and off), please go to [`Complete Assistant Version`](https://github.com/TheRamU/Fay/tree/fay-assistant-edition)
 
-***“Exceptional products deserve to be reimagined with digital humans”***
+**"Excellent products deserve to be redone with digital humans."**
+1.Assistant mode based on schedule maintenance: Managing and maintaining your schedule, not just a simple alarm clock.
+<img src="images/you1.png" alt="Fay">
 
-Highlights: Proactive execution of planned tasks without the need for question-and-answer interactions, automatic planning and use of the agent tool to complete tasks; use of OpenAI TTS; use of a vector database for permanent memory and memory retrieval;
+2.Powerful planning and execution (ReAct) capability: Plan -> Execute <-> Reflect -> Summarize.
+<img src="images/you2.png" alt="Fay">
 
-![](images/agent_demo.gif)
+3.Automatic switching between LLM Chain and React Agent: Retains planning and execution capabilities while considering chatting abilities (still needs optimization).
+<img src="images/you3.png" alt="Fay">
 
-​                                                                       (Above image: Testing ReAct capabilities）
+4.Dual memory mechanism: Stanford AI Town's memory stream (time, importance, relevance) for long-term memory, and adjacent conversation memory for coherent conversations.
+<img src="images/you4.png" alt="Fay">
+
+5.Easily expandable agent tools.
+<img src="images/you5.png" alt="Fay">
+
+6.Accompanying 24-hour background running Android connector.
+<img src="images/you6.png" alt="Fay">
 
 ## **Installation Instructions**
 
@@ -60,19 +70,24 @@ Repository URL: https://github.com/xszyou/fay-android
 
 
 ### **Changelog**
+2024.01.01:
+
+OpenAI token calculation ✓
+Optimized ReAct Agent and LLM Chain auto-switching logic ✓
+*Added dual memory mechanism: long-term memory stream and short-term chat memory ✓
+Fixed record.py ASR bug ✓
+Improved stability of remote audio (Android connector) ✓
+Fixed execution time calculation bug ✓
+Optimized voice output logic ✓
+
 2023.12.25:
 
 Implemented the automatic switching logic between agent ReAct and LLM chain ✓
 Distinguished task messages in the chat window ✓
-
 Fixed the bug in deleting schedules ✓
-
 Optimized remote audio logic ✓
-
 Introduced loading effects for pending processes ✓
-
 Optimized prompts to resolve recursive calling issues in schedule tasks ✓
-
 Fixed the bug in clearing one-time schedules ✓
 
 

agent/agent_service.py

@@ -3,6 +3,7 @@ import threading
 import datetime
 import time
 from agent.fay_agent import FayAgentCore
+from core import fay_core
 
 scheduled_tasks = {}
 agent_running = False
@@ -52,8 +53,10 @@ def parse_repeat_rule(rule, task_time):
 
 # 执行任务
 def execute_task(task_time, id, content):
-    agent.run("执行任务->立刻" + content)
+    agent.is_chat = False
-    del scheduled_tasks[id]
+    fay_core.send_for_answer("执行任务->立刻" + content)
+    if id in scheduled_tasks:
+        del scheduled_tasks[id]
     # 如果不重复，执行后删除记录
     conn = sqlite3.connect('timer.db')
     cursor = conn.cursor()
@@ -89,24 +92,24 @@ def agent_start():
     global agent
     
     agent_running = True
-    init_db()
+    # init_db()
     check_and_execute_thread = threading.Thread(target=check_and_execute)
     check_and_execute_thread.start()
 
     #初始计划
-    agent.run("""执行任务-->
+    # fay_core.send_for_answer("""执行任务-->
-        你是一个智慧农业实验箱里的ai，你的责任是陪伴主人生活、工作，以及协助主人打理好农业种植箱里的农作物：
+    #     你是一个智慧农业实验箱里的ai，你的责任是陪伴主人生活、工作，以及协助主人打理好农业种植箱里的农作物：
-        1、在每天早上8点、中午12点、晚上10点检查农业种植箱的状态是否附合设定的预期执行，如果不符合请告知我调整。
+    #     1、在每天早上8点、中午12点、晚上10点检查农业种植箱的状态是否附合设定的预期执行，如果不符合请告知我调整。
-        2、每天12点“语音提醒主人吃饭”;  
+    #     2、每天12点“语音提醒主人吃饭”;  
-        3、在星期一到星期五13:30“语音提醒主人开始工作”; 
+    #     3、在星期一到星期五13:30“语音提醒主人开始工作”; 
-        4、在星期一到星期五15:15“语音提醒主人冲咖啡”; 
+    #     4、在星期一到星期五15:15“语音提醒主人冲咖啡”; 
-        5、在星期一、星期三的11:15“语音提醒主人开会”; 
+    #     5、在星期一、星期三的11:15“语音提醒主人开会”; 
-        6、在星期五17:30“语音提醒主人开会”;
+    #     6、在星期五17:30“语音提醒主人开会”;
-        7、在星期一到星期五18:00“语音提醒主人下班”;
+    #     7、在星期一到星期五18:00“语音提醒主人下班”;
-        8、在每天21点陪主人聊聊天;  
+    #     8、在每天21点陪主人聊聊天;  
-        9、在每天晚上10:30会跟据第二天的天气预报信息和当天的运行情况，检查iotm系统当天的控制规则；
+    #     9、在每天晚上10:30会跟据第二天的天气预报信息和当天的运行情况，检查iotm系统当天的控制规则；
 
-        """)
+    #     """)
 
 def agent_stop():
     global agent_running 

agent/fay_agent.py

@@ -1,3 +1,7 @@
+import os
+import time
+import math
+
 from langchain.embeddings.openai import OpenAIEmbeddings
 from langchain.chat_models import ChatOpenAI
 from langchain.memory import VectorStoreRetrieverMemory
@@ -21,18 +25,18 @@ from agent.tools.DeleteTimer import DeleteTimer
 from agent.tools.GetSwitchLog import GetSwitchLog
 from agent.tools.getOnRunLinkage import getOnRunLinkage
 from agent.tools.SetChatStatus import SetChatStatus
-
+from langchain.callbacks import get_openai_callback
+from langchain.retrievers import TimeWeightedVectorStoreRetriever
+from langchain.memory import ConversationBufferWindowMemory
 
 import utils.config_util as utils
-from core.content_db import Content_Db
 from core import wsa_server
-import os
+import fay_booter
-
+from utils import util
 
 
 class FayAgentCore():
     def __init__(self):
-
         utils.load_config()
         os.environ['OPENAI_API_KEY'] = utils.key_gpt_api_key
         #使用open ai embedding
@@ -41,19 +45,24 @@ class FayAgentCore():
         embedding_fn = OpenAIEmbeddings()
 
         #创建llm
-        llm = ChatOpenAI(model="gpt-4-1106-preview", verbose=True)
+        self.llm = ChatOpenAI(model="gpt-4-1106-preview", verbose=True)
 
         #创建向量数据库
-        vectorstore = FAISS(embedding_fn, index, InMemoryDocstore({}), {})
+        def relevance_score_fn(self, score: float) -> float:
+            return 1.0 - score / math.sqrt(2)
+        vectorstore = FAISS(embedding_fn, index, InMemoryDocstore({}), {}, relevance_score_fn=relevance_score_fn)
 
-        # 创建记忆
+        # 创建记忆(斯坦福小镇同款记忆检索机制:时间、相关性、重要性三个维度)
-        retriever = vectorstore.as_retriever(search_kwargs=dict(k=2))
+        retriever = TimeWeightedVectorStoreRetriever(vectorstore=vectorstore, other_score_keys=["importance"], k=3)  
-        memory = VectorStoreRetrieverMemory(memory_key="chat_history", retriever=retriever)
+        self.agent_memory = VectorStoreRetrieverMemory(memory_key="history", retriever=retriever)
 
         # 保存基本信息到记忆
         utils.load_config()
         attr_info = ", ".join(f"{key}: {value}" for key, value in utils.config["attribute"].items())
-        memory.save_context({"input": "我的基本信息是?"}, {"output": attr_info})
+        self.agent_memory.save_context({"input": "我的基本信息是?"}, {"output": attr_info})
+
+        #内存保存聊天历史
+        self.chat_history = []
 
         #创建agent chain
         my_timer = MyTimer()
@@ -70,7 +79,7 @@ class FayAgentCore():
         get_on_run_linkage = getOnRunLinkage()
         set_chat_status_tool = SetChatStatus()
 
-        tools = [
+        self.tools = [
             Tool(
                 name=my_timer.name,
                 func=my_timer.run,
@@ -140,58 +149,136 @@ class FayAgentCore():
 
         #agent用于执行任务
         self.agent = initialize_agent(agent_types=agent_types.AgentType.CHAT_CONVERSATIONAL_REACT_DESCRIPTION,
-                         tools=tools, llm=llm, verbose=True,
+                         tools=self.tools, llm=self.llm, verbose=True,
-                         max_history=5, 
+                         max_history=5, handle_parsing_errors=True)
-                         memory=memory, handle_parsing_errors=True)
         
-        #创建llm chain用于聊天
+        #llm chain  用于聊天
-        tools_prompt = "["
-        tool_names = [tool.name for tool in tools if tool.name != set_chat_status_tool.name and tool.name != say_tool.name]
-        tools_prompt += "、".join(tool_names) + "]"
-        template = """你是一个智慧农业实验箱里的ai，你的责任是陪伴主人生活、工作，以及协助主人打理好农业种植箱里的农作物.现在主人正在和你聊天，若你在聊天过程中感觉到主人想使用以下工具，请按“agent:'主人刚刚的说话'”这样的格式回复，否则请直接回复我文字内容。工具如下：
-        """ +  tools_prompt +"""
-        {chat_history}
-        Human: {human_input}
-        AI:"""
-        prompt = PromptTemplate(
-            input_variables=["chat_history", "human_input"], template=template
-        )
-        self.llm_chain = LLMChain(
-            llm=llm,
-            prompt=prompt,
-            verbose=True,
-            memory=memory
-        )
-
         self.is_chat = False#聊天状态
 
+        #记录一轮执行有无调用过say tool
+        self.is_use_say_tool = False   
+        self.say_tool_text = ""
 
-    def run(self, input_text):
+        self.total_tokens = 0
-        #消息保存
+        self.total_cost = 0
-        contentdb = Content_Db()    
+
-        contentdb.add_content('member', 'agent', input_text.replace('主人语音说了：', '').replace('主人文字说了：', ''))
+
-        wsa_server.get_web_instance().add_cmd({"panelReply": {"type":"member","content":input_text.replace('主人语音说了：', '').replace('主人文字说了：', '')}})
+    def format_history_str(self, str):
         result = ""
+        history_string = str['history']
+
+        # Split the string into lines
+        lines = history_string.split('input:')
+
+        # Initialize an empty list to store the formatted history
+        formatted_history = []
+
+        #处理记忆流格式
+        for line in lines:
+            if "output" in line:
+                input_line = line.split("output:")[0].strip()
+                output_line = line.split("output:")[1].strip()
+                formatted_history.append({"input": input_line, "output": output_line})
+
+        
+        # 记忆流转换成字符串
+        result += "-以下是与用户说话关连度最高的记忆：\n"
+        for i in range(len(formatted_history)):
+            if i >= 3:
+                break
+            line = formatted_history[i]
+            result += f"--input：{line['input']}\n--output：{line['output']}\n"
+        if len(formatted_history) == 0:
+            result += "--没有记录\n"
+
+
+        #添加内存记忆
+        formatted_history = []
+        for line in self.chat_history:
+            formatted_history.append({"input": line[0], "output": line[1]})
+        
+        #格式化内存记忆字符串
+        result += "\n-以下刚刚的对话：\n"
+        for i in range(len(formatted_history)):
+            line = formatted_history[i]
+            result += f"--input：{line['input']}\n--output：{line['output']}\n"
+        if len(formatted_history) == 0:
+            result += "--没有记录\n"
+
+        return result
+
+    
+    def get_llm_chain(self, history):
+        tools_prompt = "["
+        tool_names = [tool.name for tool in self.tools if tool.name != SetChatStatus().name and tool.name != Say().name]
+        tools_prompt += "、".join(tool_names) + "]"
+        template = """
+你是一个智能家居系统中的AI，负责协助主人处理日常事务和智能设备的操作。当主人提出要求时，如果需要使用特定的工具或执行特定的操作，请严格回复“agent: {human_input}”字符串。如果主人只是进行普通对话或询问信息，直接以文本内容回答即可。你可以使用的工具或执行的任务包括：。
+""" +  tools_prompt + "等。" +"""
+现在时间是：now_time
+请依据以下信息回复主人：
+chat_history
+
+input：
+{human_input}
+output：""".replace("chat_history", history).replace("now_time", QueryTime().run(""))
+        prompt = PromptTemplate(
+                    input_variables=["human_input"], template=template
+                )
+        
+        llm_chain = LLMChain(
+            llm=self.llm,
+            prompt=prompt,
+            verbose=True
+        )
+        return llm_chain
+    
+    def run(self, input_text):
+        self.is_use_say_tool = False
+        self.say_tool_text = ""
+        
+        result = ""
+        history = self.agent_memory.load_memory_variables({"input":input_text.replace('主人语音说了：', '').replace('主人文字说了：', '')})
+        history = self.format_history_str(history)
         try:
             #判断执行聊天模式还是agent模式，双模式在运行过程中会主动切换
             if self.is_chat:
-                result = self.llm_chain.predict(human_input=input_text.replace('主人语音说了：', '').replace('主人文字说了：', ''))
+                llm_chain = self.get_llm_chain(history)
+                with get_openai_callback() as cb:
+                    result = llm_chain.predict(human_input=input_text.replace('主人语音说了：', '').replace('主人文字说了：', ''))
+                    self.total_tokens = self.total_tokens + cb.total_tokens
+                    self.total_cost = self.total_cost + cb.total_cost
+                    util.log(1, "本次消耗token:{}， Cost (USD):{}，共消耗token:{}， Cost (USD):{}".format(cb.total_tokens, cb.total_cost, self.total_tokens, self.total_cost))
+
             if "agent:" in result.lower() or not self.is_chat:
-                print(result)
-                print(self.is_chat)
                 self.is_chat = False
-                input_text = result if  result.lower().replace("agent:", "") else input_text
+                input_text = result.lower().replace("agent:", "") if "agent:" in result.lower() else input_text.replace('主人语音说了：', '').replace('主人文字说了：', '')
-                result = self.agent.run(input_text)
+                agent_prompt = """
+现在时间是：{now_time}。请依据以下信息为主人服务 ：
+{history}
+input：{input_text}
+output：
+""".format(history=history, input_text=input_text, now_time=QueryTime().run(""))
+                print(agent_prompt)
+                with get_openai_callback() as cb:
+                    result = self.agent.run(agent_prompt)
+                    self.total_tokens = self.total_tokens + cb.total_tokens
+                    self.total_cost = self.total_cost + cb.total_cost
+                    util.log(1, "本次消耗token:{}， Cost (USD):{}，共消耗token:{}， Cost (USD):{}".format(cb.total_tokens, cb.total_cost, self.total_tokens, self.total_cost))
+                    
         except Exception as e:
             print(e)
+        
         result = "执行完毕" if result is None or result == "N/A" else result
-        
+        chat_text = self.say_tool_text if self.is_use_say_tool else result
 
-        #消息保存
+        #保存到记忆流和聊天对话
-        contentdb.add_content('fay','agent', result)
+        self.agent_memory.save_context({"input": input_text.replace('主人语音说了：', '').replace('主人文字说了：', '')},{"output": result})
-        wsa_server.get_web_instance().add_cmd({"panelReply": {"type":"fay","content":result}})
+        self.chat_history.append((input_text.replace('主人语音说了：', '').replace('主人文字说了：', ''), chat_text))
-        
+        if len(self.chat_history) > 5:
-        return result
+            self.chat_history.pop(0)
+
+        return self.is_use_say_tool, chat_text
 
 if __name__ == "__main__":
     agent = FayAgentCore()

agent/tools/Say.py

@@ -21,9 +21,11 @@ class Say(BaseTool):
 
 
     def _run(self, para: str) -> str:
+        agent_service.agent.is_chat = True
+        agent_service.agent.is_use_say_tool = True
+        agent_service.agent.say_tool_text = para
         interact = Interact("audio", 1, {'user': '', 'msg': para})
         fay_booter.feiFei.on_interact(interact)
-        agent_service.agent.is_chat = True
         return "语音输出了：" + para
 
 

core/fay_core.py

@@ -10,10 +10,8 @@ import logging
 
 # 适应模型使用
 import numpy as np
-# import tensorflow as tf
 import fay_booter
 from ai_module import xf_ltp
-# from ai_module.ms_tts_sdk import Speech
 from ai_module.openai_tts import Speech
 from core import wsa_server
 from core.interact import Interact
@@ -27,6 +25,7 @@ import platform
 from ai_module import yolov8
 from agent import agent_service
 import fay_booter
+from core.content_db import Content_Db
 if platform.system() == "Windows":
     import sys
     sys.path.append("test/ovr_lipsync")
@@ -38,6 +37,11 @@ def send_for_answer(msg):
         #记录运行时间
         fay_booter.feiFei.last_quest_time = time.time()
 
+        #消息保存
+        contentdb = Content_Db()    
+        contentdb.add_content('member', 'agent', msg.replace('主人语音说了：', '').replace('主人文字说了：', ''))
+        wsa_server.get_web_instance().add_cmd({"panelReply": {"type":"member","content":msg.replace('主人语音说了：', '').replace('主人文字说了：', '')}})
+
         # 发送给数字人端    
         if not config_util.config["interact"]["playSound"]: 
             content = {'Topic': 'Unreal', 'Data': {'Key': 'question', 'Value': msg}}
@@ -49,14 +53,19 @@ def send_for_answer(msg):
             content = {'Topic': 'Unreal', 'Data': {'Key': 'log', 'Value': "思考中..."}}
             wsa_server.get_instance().add_cmd(content)
         
-        #agent 处理
+        #agent 或llm chain处理
-        text = agent_service.agent.run(msg)
+        is_use_say_tool, text = agent_service.agent.run(msg)
 
-        #聊天模式语音输入语音输出
+        #语音输入强制语音输出
-        if text and "语音" in msg and agent_service.agent.is_chat:
+        if text and "语音说了" in msg and not is_use_say_tool:
             interact = Interact("audio", 1, {'user': '', 'msg': text})
             fay_booter.feiFei.on_interact(interact)
 
+        #消息保存
+        contentdb.add_content('fay','agent', text)
+        wsa_server.get_web_instance().add_cmd({"panelReply": {"type":"fay","content":text}})
+        util.log(1, 'ReAct Agent或LLM Chain处理总时长：{} ms'.format(math.floor((time.time() - fay_booter.feiFei.last_quest_time) * 1000)))
+
         #推送数字人
         if not cfg.config["interact"]["playSound"]: 
             content = {'Topic': 'Unreal', 'Data': {'Key': 'log', 'Value': text}}
@@ -294,7 +303,7 @@ class FeiFei:
                 except Exception as serr:
                     util.log(1,"远程音频输入输出设备已经断开：{}".format(serr))
                     self.deviceConnect = None
-            time.sleep(1)
+            time.sleep(5)
 
     def __accept_audio_device_output_connect(self):
         self.deviceSocket = socket.socket(socket.AF_INET,socket.SOCK_STREAM) 

core/recorder.py

@@ -42,16 +42,14 @@ class Recorder:
         if cfg.config['source']['wake_word_enabled']:
             self.timer = threading.Timer(60, self.reset_wakeup_status)  # 60秒后执行reset_wakeup_status方法
 
-
     def asrclient(self):
+        asrcli = None
         if self.ASRMode == "ali":
             asrcli = ALiNls()
         elif self.ASRMode == "funasr":
             asrcli = FunASR()
         return asrcli
 
-    
-
     def __get_history_average(self, number):
         total = 0
         num = 0