init commit（包含LLM交互部分）

LLM/spark_desk.py

@@ -0,0 +1,105 @@
+# # -*- coding: utf-8 -*-
+# # @Time : 2023/6/28 16:08
+# # @Author : Fishead_East
+# # @Email : ytzd2696@foxmail.com
+# # @File : spark_desk.py
+# # @Project : PromptArt
+"""
+    使用LangChain自定义LLM
+"""
+import os
+from typing import Optional, List, Mapping, Any
+import ssl
+
+from langchain.llms.base import LLM
+from langchain.callbacks.manager import CallbackManagerForLLMRun
+from langchain.prompts import PromptTemplate
+from langchain.chains import LLMChain, SimpleSequentialChain
+
+import websocket
+from LLM.webInteract.web_param import WsParamGPT
+from LLM.webInteract.web_interact_gpt import Singleton, WS
+from LLM.webInteract.web_interact_gpt import (on_close,
+                                              on_open,
+                                              on_error,
+                                              on_message)
+
+
+@Singleton
+class SparkDesk(LLM):
+    """
+    讯飞星火的语言模型
+    """
+
+    url = "wss://spark-api.xf-yun.com/v1.1/chat"
+    APPID = os.getenv("APPID")      # 环境变量
+    APIKey = os.getenv("APIKEY")
+    APISecret = os.getenv("APISECRET")
+
+    @property
+    def _llm_type(self) -> str:
+        return "SparkDesk"
+
+    @property
+    def _identifying_params(self) -> Mapping[str, Any]:
+        _param_dict = {
+            "url": self.url,
+            "APPID": self.APPID,
+            "APIKey": self.APIKey,
+            "APISecret": self.APISecret
+        }
+        return _param_dict
+
+    def _call(
+            self,
+            prompt: str,
+            stop: Optional[List[str]] = None,
+            run_manager: Optional[CallbackManagerForLLMRun] = None,
+    ) -> str:
+        ws_param = WsParamGPT(self.url, self.APPID, self.APIKey, self.APISecret)
+        websocket.enableTrace(False)
+        wsUrl = ws_param.create_url()
+        ws = WS(appid=ws_param.APPID,
+                url=wsUrl,
+                on_message=on_message,
+                on_error=on_error,
+                on_close=on_close,
+                on_open=on_open)
+        ws.question = prompt
+        ws.run_forever(sslopt={"cert_reqs": ssl.CERT_NONE})  # 建立长连接
+        return ws.received_message
+
+
+if __name__ == "__main__":
+    llm = SparkDesk()
+
+    # test
+    prompt_1 = PromptTemplate(
+        input_variables=["lastname"],
+        template="我的邻居姓{lastname}，他生了个儿子，给他儿子起个名字",
+    )
+    chain_1 = LLMChain(llm=llm,
+                       prompt=prompt_1)
+    # 创建第二条链
+    prompt_2 = PromptTemplate(
+        input_variables=["child_name"],
+        template="邻居的儿子名字叫{child_name}，给他起一个小名",
+    )
+    chain_2 = LLMChain(llm=llm, prompt=prompt_2)
+
+    # 链接两条链
+    overall_chain = SimpleSequentialChain(chains=[chain_1, chain_2], verbose=True)
+
+    # 执行链，只需要传入第一个参数
+    catchphrase = overall_chain.run("王")
+    print(catchphrase)
+
+# # 使用代理来确定如何使用LLM来采取行动
+# # 这里省略代理的定义
+#
+# # 使用内存来在链或调用之间存储状态
+# # 这里省略内存的定义
+#
+# # 测试应用程序
+# response = chain.run("the meaning of life")
+# print(response)

LLM/spark_desk_embedding.py

@@ -0,0 +1,92 @@
+# -*- coding: utf-8 -*-
+# @Time : 2023/7/3 15:41
+# @Author : Fishead_East
+# @Email : ytzd2696@foxmail.com
+# @File : spark_desk_embedding.py
+# @Project : PromptArt
+import os
+import json
+import requests
+from typing import Optional, List, Mapping, Any
+
+from langchain.llms.base import LLM
+from langchain.callbacks.manager import CallbackManagerForLLMRun
+from langchain.embeddings.base import Embeddings
+from langchain.embeddings.huggingface import HuggingFaceEmbeddings
+
+from LLM.webInteract.web_param import WsParamEmb
+
+
+class SparkDeskEmbedding(object):
+    """
+    讯飞星火的Embedding模型
+    """
+    url = r'https://knowledge-retrieval.cn-huabei-1.xf-yun.com/v1/aiui/embedding/query'
+    APPID: str = os.getenv("APPID")
+    APIKey: str = os.getenv("APIKEY")
+    APISecret: str = os.getenv("APISECRET")
+
+    def _get_param(self, text) -> Mapping[str, Any]:
+        """
+        组织请求消息
+        :param text: 待向量化的文本
+        :return:
+        """
+        param_dict = {
+            'header': {
+                'app_id': self.APPID
+            },
+            'payload': {
+                'text': text
+            }
+        }
+        return param_dict
+
+    def embed_query(self, text: str,) -> List[float]:
+        """Compute query embeddings using the Spark Desk Model.
+
+        Args:
+            text: The text to embed.
+
+        Returns:
+            Embeddings for the text.
+        """
+        ws_param = WsParamEmb(self.url, self.APPID, self.APIKey, self.APISecret)
+        wsUrl = ws_param.create_url()
+        param_dict = self._get_param(text)
+        response = requests.post(url=wsUrl, json=param_dict)    # 得到响应串
+        result_dict = json.loads(response.content.decode('utf-8'))
+        embed = json.loads(result_dict['payload']['text']['vector'])
+        return embed
+
+
+if __name__ == '__main__':
+    llm_embed = SparkDeskEmbedding()
+    embed1 = llm_embed.embed_query("你好吗？")
+    print(embed1)
+    print(len(embed1))
+    print(embed1[:5])
+
+
+# class SparkDeskEmbeddings(HuggingFaceEmbeddings):
+#     """重写HuggingFaceEmbeddings加载类"""
+#
+#     client: Any  #: :meta private:
+#     model_name: str = "SparkDeskEmbeddings"
+#
+#     def __init__(self, **kwargs: Any):
+#         super().__init__(**kwargs)
+#         # self.client即向量化工具，为sentence_transformers包中的类
+#
+#     def embed_query(self, text: str) -> List[float]:
+#         """Compute query embeddings using a HuggingFace transformer model.
+#
+#         Args:
+#             text: The text to embed.
+#
+#         Returns:
+#             Embeddings for the text.
+#         """
+#         text = text.replace("\n", " ")
+#         embedding = self.client.encode(text, normalize_embeddings=True)
+#         return embedding.tolist()

LLM/webInteract/web_interact_gpt.py

@@ -0,0 +1,96 @@
+"""
+    GPT交互的功能模块
+"""
+import _thread as thread
+import json
+import websocket
+
+from LLM.webInteract.web_param import gen_params_gpt
+
+
+class Singleton(object):
+    """
+    类装饰器,用于实现单例模式
+    """
+
+    def __init__(self, cls):
+        self._cls = cls
+        self._instance = {}
+
+    def __call__(self):
+        if self._cls not in self._instance:
+            self._instance[self._cls] = self._cls()
+        return self._instance[self._cls]
+
+
+class WS(websocket.WebSocketApp):
+    """
+    WebSocketApp子类，添加消息变量
+    """
+
+    def __init__(self, appid, url, on_message, on_error, on_close, on_open):
+        self.appid = appid
+        self.received_message = ""
+        super(WS, self).__init__(url=url,
+                                 on_message=on_message,
+                                 on_error=on_error,
+                                 on_close=on_close,
+                                 on_open=on_open)
+
+
+def on_error(ws, error):
+    """
+    收到websocket错误的处理
+    """
+    print("### error:", error)
+
+
+def on_close(ws):
+    """
+    收到websocket关闭的处理
+    """
+    print("### closed ###")
+
+
+def on_open(ws):
+    """
+    收到websocket连接建立的处理
+    """
+    # print("### open ###")
+    thread.start_new_thread(run, (ws,))
+
+
+def run(ws, *args):
+    # print("### run ###")
+    data = json.dumps(gen_params_gpt(appid=ws.appid, question=ws.question))
+    ws.send(data)
+
+
+def on_message(ws, message):
+    """
+    到websocket消息的处理
+    :param ws:
+    :param message:
+    :return:
+    """
+    # print("### message ###")
+    data = json.loads(message)  # 将JSON字符串转化为Python对象
+    code = data['header']['code']
+    if code != 0:
+        print(f'请求错误: {code}, {data}')
+        ws.close()
+        # print("# closed 1 #")
+    else:
+        choices = data["payload"]["choices"]  # 有效载荷数据
+        status = choices["status"]  # 消息的状态
+        content = choices["text"][0]["content"]  # 消息的内容
+        # 存储返回的消息
+        ws.received_message += content
+
+        # print(ws.received_message)
+        if status == 2:  # 判断末尾消息
+            ws.close()
+            # print("# closed 2 #")
+            # print(ws.received_message)
+
+

LLM/webInteract/web_param.py

@@ -0,0 +1,125 @@
+"""
+    交互的参数模块
+"""
+# -*- coding: utf-8 -*-
+# @Time : 2023/7/4 09:35
+# @Author : Fishead_East
+# @Email : ytzd2696@foxmail.com
+# @File : web_param.py
+# @Project : PromptArt
+import _thread as thread
+import base64
+import datetime
+import hashlib
+import hmac
+from urllib.parse import urlparse
+from datetime import datetime
+from time import mktime
+from urllib.parse import urlencode
+from wsgiref.handlers import format_date_time
+
+
+class WsParamGPT(object):
+    """
+    交互参数类(GPT)
+    """
+    def __init__(self, url, app_id, api_key, api_secret):
+        self.url = url
+        self.host = urlparse(self.url).netloc
+        self.path = urlparse(self.url).path
+        self.APPID = app_id
+        self.APIKey = api_key
+        self.APISecret = api_secret
+
+    def create_signature_origin(self, date) -> str:
+        """拼接字符串,得到初始签名signature_origin"""
+        signature_origin = "host: " + self.host + "\n"
+        signature_origin += "date: " + date + "\n"
+        signature_origin += "GET " + self.path + " HTTP/1.1"    # GPT交互使用Get
+        return signature_origin
+
+    def create_url(self):
+        """
+        生成url [host主机 + date时间戳(RFC1123格式) + authorization认证信息]
+        """
+        # print("create_url启动")
+        # (可对参数进行逐步打印确认)
+        # 生成参数:时间戳date(RFC1123格式)
+        now = datetime.now()
+        date = format_date_time(mktime(now.timetuple()))
+
+        # 生成参数:认证信息authorization(base64编码)
+        signature_origin = self.create_signature_origin(date)
+        # 进行hmac-sha256算法进行加密; 得到签名的摘要signature_sha
+        signature_sha = hmac.new(self.APISecret.encode('utf-8'), signature_origin.encode('utf-8'),
+                                 digestmod=hashlib.sha256).digest()
+
+        # 进行base64编码生成签名signature_sha_base64
+        signature_sha_base64 = base64.b64encode(signature_sha).decode(encoding='utf-8')
+
+        # 将字符串拼接成原始认证
+        authorization_origin = f'api_key="{self.APIKey}", algorithm="hmac-sha256", headers="host date request-line", ' \
+                               f'signature="{signature_sha_base64}"'
+
+        # 进行base64编码生成最终认证信息authorization
+        authorization = base64.b64encode(authorization_origin.encode('utf-8')).decode(encoding='utf-8')
+        # 将请求的鉴权参数组合为字典
+        v = {
+            "authorization": authorization,
+            "date": date,
+            "host": self.host
+        }
+        # 拼接鉴权参数，生成url
+        url_final = self.url + '?' + urlencode(v)
+        # print("create_url完成")
+        return url_final
+
+
+class WsParamEmb(WsParamGPT):
+    """
+    交互参数类（Embedding）
+    """
+    def __init__(self, url, app_id, api_key, api_secret):
+        super(WsParamEmb, self).__init__(url, app_id, api_key, api_secret)
+
+    def create_signature_origin(self, date) -> str:
+        """拼接字符串,得到初始签名signature_origin"""
+        signature_origin = "host: " + self.host + "\n"
+        signature_origin += "date: " + date + "\n"
+        signature_origin += "POST " + self.path + " HTTP/1.1"   # Embedding交互使用Post
+        return signature_origin
+
+
+def gen_params_gpt(appid, question):
+    """
+        通过appid和用户的提问, 生成请求参数
+        """
+    # print("gen_params启动")
+    data = {
+        "header": {
+            "app_id": appid,  # AppID
+            "uid": "1234"  # 用于区分不同的用户
+        },
+        "parameter": {
+            "chat": {
+                "domain": "general",  # (必)指定访问的领域
+                "random_threshold": 0,  # 温度系数temperature
+                "max_tokens": 2048,  # 模型回答的tokens的最大长度,范围为[1,4096]
+                "auditing": "default"
+                # "top_k": 4                # 从k个候选中随机选择⼀个（⾮等概率）,范围为[1,6]
+                # "chat_id": "1234"         # 用于关联用户会话,需要保障用户对话的唯一性
+            }
+        },
+        "payload": {
+            "message": {
+                "text": [
+                    {
+                        "role": "user",  # 对话角色,范围为[user,assistant]
+                        "content": question  # 用户和AI的对话内容,text下所有content累计tokens需要控制在8192内
+                    }
+                ]
+            }
+        }
+    }
+    # print("gen_params完成")
+    return data