253 lines
8.0 KiB
TypeScript
253 lines
8.0 KiB
TypeScript
import type { NextApiRequest, NextApiResponse } from 'next';
|
||
import { connectToDatabase, Model } from '@/service/mongo';
|
||
import { getOpenAIApi } from '@/service/utils/auth';
|
||
import { authOpenApiKey } from '@/service/utils/tools';
|
||
import { httpsAgent, openaiChatFilter, systemPromptFilter } from '@/service/utils/tools';
|
||
import { ChatCompletionRequestMessage, ChatCompletionRequestMessageRoleEnum } from 'openai';
|
||
import { ChatItemType } from '@/types/chat';
|
||
import { jsonRes } from '@/service/response';
|
||
import { PassThrough } from 'stream';
|
||
import {
|
||
ChatModelNameEnum,
|
||
modelList,
|
||
ChatModelNameMap,
|
||
ModelVectorSearchModeMap
|
||
} from '@/constants/model';
|
||
import { pushChatBill } from '@/service/events/pushBill';
|
||
import { openaiCreateEmbedding, gpt35StreamResponse } from '@/service/utils/openai';
|
||
import { PgClient } from '@/service/pg';
|
||
|
||
/* 发送提示词 */
|
||
export default async function handler(req: NextApiRequest, res: NextApiResponse) {
|
||
let step = 0; // step=1时，表示开始了流响应
|
||
const stream = new PassThrough();
|
||
stream.on('error', () => {
|
||
console.log('error: ', 'stream error');
|
||
stream.destroy();
|
||
});
|
||
res.on('close', () => {
|
||
stream.destroy();
|
||
});
|
||
res.on('error', () => {
|
||
console.log('error: ', 'request error');
|
||
stream.destroy();
|
||
});
|
||
|
||
try {
|
||
const {
|
||
prompt,
|
||
modelId,
|
||
isStream = true
|
||
} = req.body as {
|
||
prompt: ChatItemType;
|
||
modelId: string;
|
||
isStream: boolean;
|
||
};
|
||
|
||
if (!prompt || !modelId) {
|
||
throw new Error('缺少参数');
|
||
}
|
||
|
||
await connectToDatabase();
|
||
let startTime = Date.now();
|
||
|
||
/* 凭证校验 */
|
||
const { apiKey, userId } = await authOpenApiKey(req);
|
||
|
||
/* 查找数据库里的模型信息 */
|
||
const model = await Model.findById(modelId);
|
||
if (!model) {
|
||
throw new Error('找不到模型');
|
||
}
|
||
|
||
const modelConstantsData = modelList.find(
|
||
(item) => item.model === ChatModelNameEnum.VECTOR_GPT
|
||
);
|
||
if (!modelConstantsData) {
|
||
throw new Error('模型已下架');
|
||
}
|
||
console.log('laf gpt start');
|
||
|
||
// 获取 chatAPI
|
||
const chatAPI = getOpenAIApi(apiKey);
|
||
|
||
// 请求一次 chatgpt 拆解需求
|
||
const promptResponse = await chatAPI.createChatCompletion(
|
||
{
|
||
model: ChatModelNameMap[ChatModelNameEnum.GPT35],
|
||
temperature: 0,
|
||
frequency_penalty: 0.5, // 越大，重复内容越少
|
||
presence_penalty: -0.5, // 越大，越容易出现新内容
|
||
messages: [
|
||
{
|
||
role: 'system',
|
||
content: `服务端逻辑生成器.根据用户输入的需求,拆解成 laf 云函数实现的步骤,只返回步骤,按格式返回步骤: 1.\n2.\n3.\n ......
|
||
下面是一些例子:
|
||
一个 hello world 例子
|
||
1. 返回字符串: "hello world"
|
||
|
||
计算圆的面积
|
||
1. 从 body 中获取半径 radius.
|
||
2. 校验 radius 是否为有效的数字.
|
||
3. 计算圆的面积.
|
||
4. 返回圆的面积: {area}
|
||
|
||
实现一个手机号发生注册验证码方法.
|
||
1. 从 query 中获取 phone.
|
||
2. 校验手机号格式是否正确,不正确则返回错误原因:手机号格式错误.
|
||
3. 给 phone 发送一个短信验证码,验证码长度为6位字符串,内容为:你正在注册laf,验证码为:code.
|
||
4. 数据库添加数据,表为"codes",内容为 {phone, code}.
|
||
|
||
实现一个云函数，使用手机号注册账号,需要验证手机验证码.
|
||
1. 从 body 中获取 phone 和 code.
|
||
2. 校验手机号格式是否正确,不正确则返回错误原因:手机号格式错误.
|
||
2. 获取数据库数据,表为"codes",查找是否有符合 phone, code 等于body参数的记录,没有的话返回错误原因:验证码不正确.
|
||
4. 添加数据库数据,表为"users" ,内容为{phone, code, createTime}.
|
||
5. 删除数据库数据,删除 code 记录.
|
||
6. 返回新建用户的Id: return {userId}
|
||
|
||
更新博客记录。传入blogId,blogText,tags,还需要记录更新的时间.
|
||
1. 从 body 中获取 blogId,blogText 和 tags.
|
||
2. 校验 blogId 是否为空,为空则返回错误原因:博客ID不能为空.
|
||
3. 校验 blogText 是否为空,为空则返回错误原因:博客内容不能为空.
|
||
4. 校验 tags 是否为数组,不是则返回错误原因:标签必须为数组.
|
||
5. 获取当前时间,记录为 updateTime.
|
||
6. 更新数据库数据,表为"blogs",更新符合 blogId 的记录的内容为{blogText, tags, updateTime}.
|
||
7. 返回结果 "更新博客记录成功"`
|
||
},
|
||
{
|
||
role: 'user',
|
||
content: prompt.value
|
||
}
|
||
]
|
||
},
|
||
{
|
||
timeout: 120000,
|
||
httpsAgent: httpsAgent(true)
|
||
}
|
||
);
|
||
|
||
const promptResolve = promptResponse.data.choices?.[0]?.message?.content || '';
|
||
if (!promptResolve) {
|
||
throw new Error('gpt 异常');
|
||
}
|
||
|
||
prompt.value += ` ${promptResolve}`;
|
||
console.log('prompt resolve success, time:', `${(Date.now() - startTime) / 1000}s`);
|
||
|
||
// 获取提示词的向量
|
||
const { vector: promptVector } = await openaiCreateEmbedding({
|
||
isPay: true,
|
||
apiKey,
|
||
userId,
|
||
text: prompt.value
|
||
});
|
||
|
||
// 读取对话内容
|
||
const prompts = [prompt];
|
||
|
||
// 相似度搜索
|
||
const similarity = ModelVectorSearchModeMap[model.search.mode]?.similarity || 0.22;
|
||
const vectorSearch = await PgClient.select<{ id: string; q: string; a: string }>('modelData', {
|
||
fields: ['id', 'q', 'a'],
|
||
order: [{ field: 'vector', mode: `<=> '[${promptVector}]'` }],
|
||
where: [
|
||
['model_id', model._id],
|
||
'AND',
|
||
['user_id', userId],
|
||
'AND',
|
||
`vector <=> '[${promptVector}]' < ${similarity}`
|
||
],
|
||
limit: 30
|
||
});
|
||
|
||
const formatRedisPrompt: string[] = vectorSearch.rows.map((item) => `${item.q}\n${item.a}`);
|
||
|
||
// textArr 筛选，最多 2500 tokens
|
||
const systemPrompt = systemPromptFilter(formatRedisPrompt, 2500);
|
||
|
||
prompts.unshift({
|
||
obj: 'SYSTEM',
|
||
value: `${model.systemPrompt} 知识库是最新的,下面是知识库内容:${systemPrompt}`
|
||
});
|
||
|
||
// 控制在 tokens 数量，防止超出
|
||
const filterPrompts = openaiChatFilter(prompts, modelConstantsData.contextMaxToken);
|
||
|
||
// 格式化文本内容成 chatgpt 格式
|
||
const map = {
|
||
Human: ChatCompletionRequestMessageRoleEnum.User,
|
||
AI: ChatCompletionRequestMessageRoleEnum.Assistant,
|
||
SYSTEM: ChatCompletionRequestMessageRoleEnum.System
|
||
};
|
||
const formatPrompts: ChatCompletionRequestMessage[] = filterPrompts.map(
|
||
(item: ChatItemType) => ({
|
||
role: map[item.obj],
|
||
content: item.value
|
||
})
|
||
);
|
||
// console.log(formatPrompts);
|
||
// 计算温度
|
||
const temperature = modelConstantsData.maxTemperature * (model.temperature / 10);
|
||
|
||
// 发出请求
|
||
const chatResponse = await chatAPI.createChatCompletion(
|
||
{
|
||
model: model.service.chatModel,
|
||
temperature,
|
||
messages: formatPrompts,
|
||
frequency_penalty: 0.5, // 越大，重复内容越少
|
||
presence_penalty: -0.5, // 越大，越容易出现新内容
|
||
stream: isStream
|
||
},
|
||
{
|
||
timeout: 120000,
|
||
responseType: isStream ? 'stream' : 'json',
|
||
httpsAgent: httpsAgent(true)
|
||
}
|
||
);
|
||
|
||
console.log('code response. time:', `${(Date.now() - startTime) / 1000}s`);
|
||
|
||
step = 1;
|
||
let responseContent = '';
|
||
|
||
if (isStream) {
|
||
const streamResponse = await gpt35StreamResponse({
|
||
res,
|
||
stream,
|
||
chatResponse
|
||
});
|
||
responseContent = streamResponse.responseContent;
|
||
} else {
|
||
responseContent = chatResponse.data.choices?.[0]?.message?.content || '';
|
||
jsonRes(res, {
|
||
data: responseContent
|
||
});
|
||
}
|
||
|
||
console.log('laf gpt done. time:', `${(Date.now() - startTime) / 1000}s`);
|
||
|
||
const promptsContent = formatPrompts.map((item) => item.content).join('');
|
||
|
||
pushChatBill({
|
||
isPay: true,
|
||
modelName: model.service.modelName,
|
||
userId,
|
||
text: promptsContent + responseContent
|
||
});
|
||
} catch (err: any) {
|
||
if (step === 1) {
|
||
// 直接结束流
|
||
console.log('error，结束');
|
||
stream.destroy();
|
||
} else {
|
||
res.status(500);
|
||
jsonRes(res, {
|
||
code: 500,
|
||
error: err
|
||
});
|
||
}
|
||
}
|
||
}
|