开源模型应用落地-FastAPI-助力模型交互-WebSocket篇（六）

一、前言

使用 FastAPI 可以帮助我们更简单高效地部署 AI 交互业务。FastAPI 提供了快速构建 API 的能力,开发者可以轻松地定义模型需要的输入和输出格式,并编写好相应的业务逻辑。

FastAPI 的异步高性能架构,可以有效支持大量并发的预测请求,为用户提供流畅的交互体验。此外,FastAPI 还提供了容器化部署能力,开发者可以轻松打包 AI 模型为 Docker 镜像,实现跨环境的部署和扩展。

总之,使用 FastAPI 可以大大提高 AI 应用程序的开发效率和用户体验,为 AI 模型的部署和交互提供全方位的支持。

本篇在开源模型应用落地-FastAPI-助力模型交互-WebSocket篇（五）基础上，学习如何*++集成Tool获取实时数据，并以流式方式返回++*

---

二、术语

2.1.Tool

Tool（工具）是为了增强其语言模型的功能和实用性而设计的一系列辅助手段，用于扩展模型的能力。例如代码解释器（Code Interpreter）和知识检索（Knowledge Retrieval）等都属于其工具。

2.2.langchain预置的tools

https://github.com/langchain-ai/langchain/tree/v0.1.16/docs/docs/integrations/tools

基本这些工具能满足大部分需求，具体使用参见：

2.3.LangChain支持流式输出的方法

• stream：基本的流式传输方式，能逐步给出代理的动作和观察结果。
• astream：异步的流式传输，用于异步处理需求的情况。
• astream_events：更细致的流式传输，能流式传输代理的每个具体事件，如工具调用和结束、模型启动和结束等，便于深入了解和监控代理执行的详细过程。

2.4.langchainhub

是 LangChain 相关工具的集合中心，其作用在于方便开发者发现和共享常用的提示（Prompt）、链、代理等。

它受 Hugging Face Hub 启发，促进社区交流与协作，推动 LangChain 生态发展。当前，它在新架构中被置于 LangSmith 里，主要聚焦于 Prompt。

2.5.asyncio

是一个用于编写并发代码的标准库,它提供了构建异步应用程序的基础框架。

---

三、前置条件

3.1. 创建虚拟环境&安装依赖

增加Google Search以及langchainhub的依赖包

conda create -n fastapi_test python=3.10
conda activate fastapi_test
pip install fastapi websockets uvicorn
pip install --quiet  langchain-core langchain-community langchain-openai
pip install google-search-results langchainhub

3.2. 注册Google Search API账号

参见：开源模型应用落地-FastAPI-助力模型交互-WebSocket篇（五）

3.3. 生成Google Search API的KEY

---

四、技术实现

4.1. 使用Tool&流式输出

# -*- coding: utf-8 -*-
import asyncio
import os
from langchain.agents import  create_structured_chat_agent, AgentExecutor
from langchain_community.utilities.serpapi import SerpAPIWrapper
from langchain_core.prompts import SystemMessagePromptTemplate, HumanMessagePromptTemplate, ChatPromptTemplate
from langchain_core.tools import tool
from langchain_openai import ChatOpenAI

os.environ["OPENAI_API_KEY"] = 'sk-xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx'  # 你的Open AI Key
os.environ["SERPAPI_API_KEY"] = "xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx"


llm = ChatOpenAI(model="gpt-3.5-turbo",temperature=0,max_tokens=512)


@tool
def search(query:str):
    """只有需要了解实时信息或不知道的事情的时候才会使用这个工具，需要传入要搜索的内容。"""
    serp = SerpAPIWrapper()
    result = serp.run(query)
    print("实时搜索结果:", result)
    return result


tools = [search]

template='''
Respond to the human as helpfully and accurately as possible. You have access to the following tools:

{tools}

Use a json blob to specify a tool by providing an action key (tool name) and an action_input key (tool input).

Valid "action" values: "Final Answer" or {tool_names}

Provide only ONE action per $JSON_BLOB, as shown:

```

{{

  "action": $TOOL_NAME,

  "action_input": $INPUT

}}

```

Follow this format:

Question: input question to answer

Thought: consider previous and subsequent steps

Action:

```

$JSON_BLOB

```

Observation: action result

... (repeat Thought/Action/Observation N times)

Thought: I know what to respond

Action:

```

{{

  "action": "Final Answer",

  "action_input": "Final response to human"

}}

Begin! Reminder to ALWAYS respond with a valid json blob of a single action. Use tools if necessary. Respond directly if appropriate. Format is Action:```$JSON_BLOB```then Observation
'''
system_message_prompt = SystemMessagePromptTemplate.from_template(template)
human_template='''
{input}

{agent_scratchpad}

 (reminder to respond in a JSON blob no matter what)
'''
human_message_prompt = HumanMessagePromptTemplate.from_template(human_template)
prompt = ChatPromptTemplate.from_messages([system_message_prompt, human_message_prompt])


print(prompt)

agent = create_structured_chat_agent(
    llm, tools, prompt
)

agent_executor = AgentExecutor(agent=agent, tools=tools, verbose=True, handle_parsing_errors=True)


async def chat(params):
    events = agent_executor.astream_events(params,version="v2")
    async for event in events:
        type = event['event']
        if 'on_chat_model_stream' == type:
            data = event['data']
            chunk =  data['chunk']
            content =  chunk.content
            if content and len(content) > 0:
                print(content)



asyncio.run(chat({"input": "广州现在天气如何?"}))

调用结果：

说明：

流式输出的数据结构为：

{'event': 'on_chat_model_stream', 'data': {'chunk': AIMessageChunk(content='天', id='run-92515b63-4b86-4af8-8515-2f84def9dfab')}, 'run_id': '92515b63-4b86-4af8-8515-2f84def9dfab', 'name': 'ChatOpenAI', 'tags': ['seq:step:3'], 'metadata': {'ls_provider': 'openai', 'ls_model_name': 'gpt-3.5-turbo', 'ls_model_type': 'chat', 'ls_temperature': 0.0, 'ls_max_tokens': 512, 'ls_stop': ['\nObservation']}}
type: on_chat_model_stream
{'event': 'on_chat_model_stream', 'data': {'chunk': AIMessageChunk(content='气', id='run-92515b63-4b86-4af8-8515-2f84def9dfab')}, 'run_id': '92515b63-4b86-4af8-8515-2f84def9dfab', 'name': 'ChatOpenAI', 'tags': ['seq:step:3'], 'metadata': {'ls_provider': 'openai', 'ls_model_name': 'gpt-3.5-turbo', 'ls_model_type': 'chat', 'ls_temperature': 0.0, 'ls_max_tokens': 512, 'ls_stop': ['\nObservation']}}

4.2. 通过langchainhub使用公共prompt

在4.1使用Tool&流式输出的代码基础上进行调整

# -*- coding: utf-8 -*-
import asyncio
import os
from langchain.agents import  create_structured_chat_agent, AgentExecutor
from langchain_community.utilities.serpapi import SerpAPIWrapper
from langchain_core.tools import tool
from langchain_openai import ChatOpenAI

os.environ["OPENAI_API_KEY"] = 'sk-xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx'  # 你的Open AI Key
os.environ["SERPAPI_API_KEY"] = "xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx"

from langchain import hub

llm = ChatOpenAI(model="gpt-3.5-turbo",temperature=0,max_tokens=512)


@tool
def search(query:str):
    """只有需要了解实时信息或不知道的事情的时候才会使用这个工具，需要传入要搜索的内容。"""
    serp = SerpAPIWrapper()
    result = serp.run(query)
    print("实时搜索结果:", result)
    return result


tools = [search]

prompt = hub.pull("hwchase17/structured-chat-agent")

print(prompt)

agent = create_structured_chat_agent(
    llm, tools, prompt
)

agent_executor = AgentExecutor(agent=agent, tools=tools, verbose=True, handle_parsing_errors=True)



async def chat(params):
    events = agent_executor.astream_events(params,version="v2")
    async for event in events:
        type = event['event']
        if 'on_chat_model_stream' == type:
            data = event['data']
            chunk =  data['chunk']
            content =  chunk.content
            if content and len(content) > 0:
                print(content)


asyncio.run(chat({"input": "广州现在天气如何?"}))

调用结果：

4.3. 整合代码

在开源模型应用落地-FastAPI-助力模型交互-WebSocket篇（五）的代码基础上进行调整

import uvicorn
import os

from typing import Annotated
from fastapi import (
    Depends,
    FastAPI,
    WebSocket,
    WebSocketException,
    WebSocketDisconnect,
    status,
)
from langchain import hub
from langchain.agents import create_structured_chat_agent, AgentExecutor
from langchain_community.utilities import SerpAPIWrapper

from langchain_core.tools import tool
from langchain_openai import ChatOpenAI

os.environ["OPENAI_API_KEY"] = 'sk-xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx'  # 你的Open AI Key
os.environ["SERPAPI_API_KEY"] = "xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx"


class ConnectionManager:
    def __init__(self):
        self.active_connections: list[WebSocket] = []

    async def connect(self, websocket: WebSocket):
        await websocket.accept()
        self.active_connections.append(websocket)

    def disconnect(self, websocket: WebSocket):
        self.active_connections.remove(websocket)

    async def send_personal_message(self, message: str, websocket: WebSocket):
        await websocket.send_text(message)

    async def broadcast(self, message: str):
        for connection in self.active_connections:
            await connection.send_text(message)

manager = ConnectionManager()

app = FastAPI()

async def authenticate(
    websocket: WebSocket,
    userid: str,
    secret: str,
):
    if userid is None or secret is None:
        raise WebSocketException(code=status.WS_1008_POLICY_VIOLATION)

    print(f'userid: {userid},secret: {secret}')
    if '12345' == userid and 'xxxxxxxxxxxxxxxxxxxxxxxxxx' == secret:
        return 'pass'
    else:
        return 'fail'

@tool
def search(query:str):
    """只有需要了解实时信息或不知道的事情的时候才会使用这个工具，需要传入要搜索的内容。"""
    serp = SerpAPIWrapper()
    result = serp.run(query)
    print("实时搜索结果:", result)
    return result


def get_prompt():
    prompt = hub.pull("hwchase17/structured-chat-agent")

    return prompt

async def chat(query):
    global llm,tools
    agent = create_structured_chat_agent(
        llm, tools, get_prompt()
    )

    agent_executor = AgentExecutor(agent=agent, tools=tools, verbose=True, handle_parsing_errors=True)

    events = agent_executor.astream_events({"input": query}, version="v1")
    async for event in events:
        type = event['event']
        if 'on_chat_model_stream' == type:
            data = event['data']
            chunk = data['chunk']
            content = chunk.content
            if content and len(content) > 0:
                print(content)
                yield content


@app.websocket("/ws")
async def websocket_endpoint(*,websocket: WebSocket,userid: str,permission: Annotated[str, Depends(authenticate)],):
    await manager.connect(websocket)
    try:
        while True:
            text = await websocket.receive_text()

            if 'fail' == permission:
                await manager.send_personal_message(
                    f"authentication failed", websocket
                )
            else:
                if text is not None and len(text) > 0:
                    async for msg in chat(text):
                        await manager.send_personal_message(msg, websocket)

    except WebSocketDisconnect:
        manager.disconnect(websocket)
        print(f"Client #{userid} left the chat")
        await manager.broadcast(f"Client #{userid} left the chat")

if __name__ == '__main__':
    tools = [search]
    llm = ChatOpenAI(model="gpt-3.5-turbo", temperature=0, max_tokens=512)
    uvicorn.run(app, host='0.0.0.0',port=7777)

客户端：

<!DOCTYPE html>
<html>
    <head>
        <title>Chat</title>
    </head>
    <body>
        <h1>WebSocket Chat</h1>
        <form action="" onsubmit="sendMessage(event)">
            <label>USERID: <input type="text" id="userid" autocomplete="off" value="12345"/></label>
            <label>SECRET: <input type="text" id="secret" autocomplete="off" value="xxxxxxxxxxxxxxxxxxxxxxxxxx"/></label>
            <br/>
            <button onclick="connect(event)">Connect</button>
            <hr>
            <label>Message: <input type="text" id="messageText" autocomplete="off"/></label>
            <button>Send</button>
        </form>
        <ul id='messages'>
        </ul>
        <script>
            var ws = null;
            function connect(event) {
                var userid = document.getElementById("userid")
                var secret = document.getElementById("secret")
                ws = new WebSocket("ws://localhost:7777/ws?userid="+userid.value+"&secret=" + secret.value);
                ws.onmessage = function(event) {
                    var messages = document.getElementById('messages')
                    var message = document.createElement('li')
                    var content = document.createTextNode(event.data)
                    message.appendChild(content)
                    messages.appendChild(message)
                };
                event.preventDefault()
            }
            function sendMessage(event) {
                var input = document.getElementById("messageText")
                ws.send(input.value)
                input.value = ''
                event.preventDefault()
            }
        </script>
    </body>
</html>

调用结果：

用户输入：你好

​

不需要触发工具调用

模型输出：

用户输入：广州现在天气如何？

​

需要调用工具

模型输出：

```
Action:
```
{
  "action": "Final Answer",
  "action_input": "广州现在的天气是多云，温度为87华氏度，降水概率为7%，湿度为76%，风力为7英里/小时。"
}
```

PS:

1. 上面仅用于演示流式输出的效果，里面包含一些冗余的信息，例如："action": "Final Answer"，要根据实际情况过滤。

2. 页面输出的样式可以根据实际需要进行调整，此处仅用于演示效果。