perf: completion dispatch

client/src/service/moduleDispatch/agent/classifyQuestion.ts

@@ -0,0 +1,103 @@
+import { adaptChatItem_openAI } from '@/utils/plugin/openai';
+import { ChatContextFilter } from '@/service/utils/chat/index';
+import type { ChatHistoryItemResType, ChatItemType } from '@/types/chat';
+import { ChatRoleEnum, TaskResponseKeyEnum } from '@/constants/chat';
+import { getOpenAIApi, axiosConfig } from '@/service/ai/openai';
+import type { ClassifyQuestionAgentItemType } from '@/types/app';
+import { countModelPrice } from '@/service/events/pushBill';
+
+export type CQProps = {
+  systemPrompt?: string;
+  history?: ChatItemType[];
+  userChatInput: string;
+  agents: ClassifyQuestionAgentItemType[];
+};
+export type CQResponse = {
+  [TaskResponseKeyEnum.responseData]: ChatHistoryItemResType;
+  [key: string]: any;
+};
+
+const moduleName = 'Classify Question';
+const agentModel = 'gpt-3.5-turbo';
+const agentFunName = 'agent_user_question';
+const maxTokens = 2000;
+
+/* request openai chat */
+export const dispatchClassifyQuestion = async (props: Record<string, any>): Promise<CQResponse> => {
+  const { agents, systemPrompt, history = [], userChatInput } = props as CQProps;
+
+  const messages: ChatItemType[] = [
+    ...(systemPrompt
+      ? [
+          {
+            obj: ChatRoleEnum.System,
+            value: systemPrompt
+          }
+        ]
+      : []),
+    ...history,
+    {
+      obj: ChatRoleEnum.Human,
+      value: userChatInput
+    }
+  ];
+  const filterMessages = ChatContextFilter({
+    model: agentModel,
+    prompts: messages,
+    maxTokens
+  });
+  const adaptMessages = adaptChatItem_openAI({ messages: filterMessages, reserveId: false });
+
+  //   function body
+  const agentFunction = {
+    name: agentFunName,
+    description: '判断用户问题的类型，并返回指定值',
+    parameters: {
+      type: 'object',
+      properties: {
+        type: {
+          type: 'string',
+          description: agents.map((item) => `${item.value}，返回：'${item.key}'`).join('；'),
+          enum: agents.map((item) => item.key)
+        }
+      },
+      required: ['type']
+    }
+  };
+  const chatAPI = getOpenAIApi();
+
+  const response = await chatAPI.createChatCompletion(
+    {
+      model: agentModel,
+      temperature: 0,
+      messages: [...adaptMessages],
+      function_call: { name: agentFunName },
+      functions: [agentFunction]
+    },
+    {
+      ...axiosConfig()
+    }
+  );
+
+  const arg = JSON.parse(response.data.choices?.[0]?.message?.function_call?.arguments || '');
+
+  if (!arg.type) {
+    throw new Error('');
+  }
+
+  const tokens = response.data.usage?.total_tokens || 0;
+
+  const result = agents.find((item) => item.key === arg.type) || agents[0];
+
+  return {
+    [result.key]: 1,
+    [TaskResponseKeyEnum.responseData]: {
+      moduleName,
+      price: countModelPrice({ model: agentModel, tokens }),
+      model: agentModel,
+      tokens,
+      cqList: agents,
+      cqResult: result.value
+    }
+  };
+};

client/src/service/moduleDispatch/agent/extract.ts

@@ -0,0 +1,100 @@
+// Next.js API route support: https://nextjs.org/docs/api-routes/introduction
+import type { NextApiRequest, NextApiResponse } from 'next';
+import { jsonRes } from '@/service/response';
+import { adaptChatItem_openAI } from '@/utils/plugin/openai';
+import { ChatContextFilter } from '@/service/utils/chat/index';
+import type { ChatItemType } from '@/types/chat';
+import { ChatRoleEnum } from '@/constants/chat';
+import { getOpenAIApi, axiosConfig } from '@/service/ai/openai';
+import type { ClassifyQuestionAgentItemType } from '@/types/app';
+import { authUser } from '@/service/utils/auth';
+
+export type Props = {
+  history?: ChatItemType[];
+  userChatInput: string;
+  agents: ClassifyQuestionAgentItemType[];
+  description: string;
+};
+export type Response = { history: ChatItemType[] };
+
+const agentModel = 'gpt-3.5-turbo-16k';
+const agentFunName = 'agent_extract_data';
+
+export default async function handler(req: NextApiRequest, res: NextApiResponse) {
+  try {
+    await authUser({ req, authRoot: true });
+
+    const response = await extract(req.body);
+
+    jsonRes(res, {
+      data: response
+    });
+  } catch (err) {
+    jsonRes(res, {
+      code: 500,
+      error: err
+    });
+  }
+}
+
+/* request openai chat */
+export async function extract({ agents, history = [], userChatInput, description }: Props) {
+  const messages: ChatItemType[] = [
+    ...history.slice(-4),
+    {
+      obj: ChatRoleEnum.Human,
+      value: userChatInput
+    }
+  ];
+  const filterMessages = ChatContextFilter({
+    // @ts-ignore
+    model: agentModel,
+    prompts: messages,
+    maxTokens: 3000
+  });
+  const adaptMessages = adaptChatItem_openAI({ messages: filterMessages, reserveId: false });
+
+  const properties: Record<
+    string,
+    {
+      type: string;
+      description: string;
+    }
+  > = {};
+  agents.forEach((item) => {
+    properties[item.key] = {
+      type: 'string',
+      description: item.value
+    };
+  });
+
+  //   function body
+  const agentFunction = {
+    name: agentFunName,
+    description,
+    parameters: {
+      type: 'object',
+      properties,
+      required: agents.map((item) => item.key)
+    }
+  };
+
+  const chatAPI = getOpenAIApi();
+
+  const response = await chatAPI.createChatCompletion(
+    {
+      model: agentModel,
+      temperature: 0,
+      messages: [...adaptMessages],
+      function_call: { name: agentFunName },
+      functions: [agentFunction]
+    },
+    {
+      ...axiosConfig()
+    }
+  );
+
+  const arg = JSON.parse(response.data.choices?.[0]?.message?.function_call?.arguments || '');
+
+  return arg;
+}

client/src/service/moduleDispatch/chat/oneapi.ts

@@ -0,0 +1,255 @@
+import type { NextApiResponse } from 'next';
+import { sseResponse } from '@/service/utils/tools';
+import { OpenAiChatEnum } from '@/constants/model';
+import { adaptChatItem_openAI, countOpenAIToken } from '@/utils/plugin/openai';
+import { modelToolMap } from '@/utils/plugin';
+import { ChatContextFilter } from '@/service/utils/chat/index';
+import type { ChatItemType, QuoteItemType } from '@/types/chat';
+import type { ChatHistoryItemResType } from '@/types/chat';
+import { ChatRoleEnum, sseResponseEventEnum } from '@/constants/chat';
+import { parseStreamChunk, textAdaptGptResponse } from '@/utils/adapt';
+import { getOpenAIApi, axiosConfig } from '@/service/ai/openai';
+import { TaskResponseKeyEnum } from '@/constants/chat';
+import { getChatModel } from '@/service/utils/data';
+import { countModelPrice } from '@/service/events/pushBill';
+
+export type ChatProps = {
+  res: NextApiResponse;
+  model: `${OpenAiChatEnum}`;
+  temperature?: number;
+  maxToken?: number;
+  history?: ChatItemType[];
+  userChatInput: string;
+  stream?: boolean;
+  quoteQA?: QuoteItemType[];
+  systemPrompt?: string;
+  limitPrompt?: string;
+};
+export type ChatResponse = {
+  [TaskResponseKeyEnum.answerText]: string;
+  [TaskResponseKeyEnum.responseData]: ChatHistoryItemResType;
+};
+
+const moduleName = 'AI Chat';
+
+/* request openai chat */
+export const dispatchChatCompletion = async (props: Record<string, any>): Promise<ChatResponse> => {
+  let {
+    res,
+    model,
+    temperature = 0,
+    maxToken = 4000,
+    stream = false,
+    history = [],
+    quoteQA = [],
+    userChatInput,
+    systemPrompt = '',
+    limitPrompt = ''
+  } = props as ChatProps;
+
+  // temperature adapt
+  const modelConstantsData = getChatModel(model);
+
+  if (!modelConstantsData) {
+    return Promise.reject('The chat model is undefined, you need to select a chat model.');
+  }
+
+  // FastGpt temperature range: 1~10
+  temperature = +(modelConstantsData.maxTemperature * (temperature / 10)).toFixed(2);
+
+  const limitText = (() => {
+    if (limitPrompt) return limitPrompt;
+    if (quoteQA.length > 0 && !limitPrompt) {
+      return '根据知识库内容回答问题，仅回复知识库提供的内容，不要对知识库内容做补充说明。';
+    }
+    return '';
+  })();
+
+  const quotePrompt =
+    quoteQA.length > 0
+      ? `下面是知识库内容:
+${quoteQA.map((item, i) => `${i + 1}. [${item.q}\n${item.a}]`).join('\n')}
+`
+      : '';
+
+  const messages: ChatItemType[] = [
+    ...(quotePrompt
+      ? [
+          {
+            obj: ChatRoleEnum.System,
+            value: quotePrompt
+          }
+        ]
+      : []),
+    ...(systemPrompt
+      ? [
+          {
+            obj: ChatRoleEnum.System,
+            value: systemPrompt
+          }
+        ]
+      : []),
+    ...history,
+    ...(limitText
+      ? [
+          {
+            obj: ChatRoleEnum.System,
+            value: limitText
+          }
+        ]
+      : []),
+    {
+      obj: ChatRoleEnum.Human,
+      value: userChatInput
+    }
+  ];
+  const modelTokenLimit = getChatModel(model)?.contextMaxToken || 4000;
+
+  const filterMessages = ChatContextFilter({
+    model,
+    prompts: messages,
+    maxTokens: Math.ceil(modelTokenLimit - 300) // filter token. not response maxToken
+  });
+
+  const adaptMessages = adaptChatItem_openAI({ messages: filterMessages, reserveId: false });
+  const chatAPI = getOpenAIApi();
+  // console.log(adaptMessages);
+
+  /* count response max token */
+  const promptsToken = modelToolMap.countTokens({
+    model,
+    messages: filterMessages
+  });
+  maxToken = maxToken + promptsToken > modelTokenLimit ? modelTokenLimit - promptsToken : maxToken;
+
+  const response = await chatAPI.createChatCompletion(
+    {
+      model,
+      temperature: Number(temperature || 0),
+      max_tokens: maxToken,
+      messages: adaptMessages,
+      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
+    },
+    {
+      timeout: stream ? 60000 : 480000,
+      responseType: stream ? 'stream' : 'json',
+      ...axiosConfig()
+    }
+  );
+
+  const { answerText, totalTokens, finishMessages } = await (async () => {
+    if (stream) {
+      // sse response
+      const { answer } = await streamResponse({ res, response });
+      // count tokens
+      const finishMessages = filterMessages.concat({
+        obj: ChatRoleEnum.AI,
+        value: answer
+      });
+
+      const totalTokens = countOpenAIToken({
+        messages: finishMessages
+      });
+
+      return {
+        answerText: answer,
+        totalTokens,
+        finishMessages
+      };
+    } else {
+      const answer = stream ? '' : response.data.choices?.[0].message?.content || '';
+      const totalTokens = stream ? 0 : response.data.usage?.total_tokens || 0;
+
+      const finishMessages = filterMessages.concat({
+        obj: ChatRoleEnum.AI,
+        value: answer
+      });
+
+      return {
+        answerText: answer,
+        totalTokens,
+        finishMessages
+      };
+    }
+  })();
+
+  return {
+    [TaskResponseKeyEnum.answerText]: answerText,
+    [TaskResponseKeyEnum.responseData]: {
+      moduleName,
+      price: countModelPrice({ model, tokens: totalTokens }),
+      model: modelConstantsData.name,
+      tokens: totalTokens,
+      question: userChatInput,
+      answer: answerText,
+      maxToken,
+      finishMessages
+    }
+  };
+};
+
+async function streamResponse({ res, response }: { res: NextApiResponse; response: any }) {
+  let answer = '';
+  let error: any = null;
+
+  const clientRes = async (data: string) => {
+    const { content = '' } = (() => {
+      try {
+        const json = JSON.parse(data);
+        const content: string = json?.choices?.[0].delta.content || '';
+        error = json.error;
+        answer += content;
+        return { content };
+      } catch (error) {
+        return {};
+      }
+    })();
+
+    if (res.closed || error) return;
+
+    if (data === '[DONE]') {
+      sseResponse({
+        res,
+        event: sseResponseEventEnum.answer,
+        data: textAdaptGptResponse({
+          text: null,
+          finish_reason: 'stop'
+        })
+      });
+      sseResponse({
+        res,
+        event: sseResponseEventEnum.answer,
+        data: '[DONE]'
+      });
+    } else {
+      sseResponse({
+        res,
+        event: sseResponseEventEnum.answer,
+        data: textAdaptGptResponse({
+          text: content
+        })
+      });
+    }
+  };
+
+  try {
+    for await (const chunk of response.data as any) {
+      if (res.closed) break;
+      const parse = parseStreamChunk(chunk);
+      parse.forEach((item) => clientRes(item.data));
+    }
+  } catch (error) {
+    console.log('pipe error', error);
+  }
+
+  if (error) {
+    console.log(error);
+    return Promise.reject(error);
+  }
+
+  return {
+    answer
+  };
+}

client/src/service/moduleDispatch/index.ts

@@ -0,0 +1,6 @@
+export * from './init/history';
+export * from './init/userChatInput';
+export * from './chat/oneapi';
+export * from './kb/search';
+export * from './tools/answer';
+export * from './agent/classifyQuestion';

client/src/service/moduleDispatch/init/history.tsx

@@ -0,0 +1,15 @@
+import { SystemInputEnum } from '@/constants/app';
+import { ChatItemType } from '@/types/chat';
+
+export type HistoryProps = {
+  maxContext: number;
+  [SystemInputEnum.history]: ChatItemType[];
+};
+
+export const dispatchHistory = (props: Record<string, any>) => {
+  const { maxContext = 5, history = [] } = props as HistoryProps;
+
+  return {
+    history: history.slice(-maxContext)
+  };
+};

client/src/service/moduleDispatch/init/userChatInput.tsx

@@ -0,0 +1,12 @@
+import { SystemInputEnum } from '@/constants/app';
+
+export type UserChatInputProps = {
+  [SystemInputEnum.userChatInput]: string;
+};
+
+export const dispatchChatInput = (props: Record<string, any>) => {
+  const { userChatInput } = props as UserChatInputProps;
+  return {
+    userChatInput
+  };
+};

client/src/service/moduleDispatch/kb/search.ts

@@ -0,0 +1,76 @@
+import { PgClient } from '@/service/pg';
+import type { ChatHistoryItemResType, ChatItemType } from '@/types/chat';
+import { TaskResponseKeyEnum } from '@/constants/chat';
+import { getVector } from '@/pages/api/openapi/plugin/vector';
+import { countModelPrice } from '@/service/events/pushBill';
+import type { SelectedKbType } from '@/types/plugin';
+import type { QuoteItemType } from '@/types/chat';
+
+type KBSearchProps = {
+  kbList: SelectedKbType;
+  history: ChatItemType[];
+  similarity: number;
+  limit: number;
+  userChatInput: string;
+};
+export type KBSearchResponse = {
+  [TaskResponseKeyEnum.responseData]: ChatHistoryItemResType;
+  isEmpty?: boolean;
+  unEmpty?: boolean;
+  quoteQA: QuoteItemType[];
+};
+
+const moduleName = 'KB Search';
+
+export async function dispatchKBSearch(props: Record<string, any>): Promise<KBSearchResponse> {
+  const {
+    kbList = [],
+    history = [],
+    similarity = 0.8,
+    limit = 5,
+    userChatInput
+  } = props as KBSearchProps;
+
+  if (kbList.length === 0) {
+    return Promise.reject("You didn't choose the knowledge base");
+  }
+
+  if (!userChatInput) {
+    return Promise.reject('Your input is empty');
+  }
+
+  // get vector
+  const vectorModel = global.vectorModels[0];
+  const { vectors, tokenLen } = await getVector({
+    model: vectorModel.model,
+    input: [userChatInput]
+  });
+
+  // search kb
+  const res: any = await PgClient.query(
+    `BEGIN;
+    SET LOCAL ivfflat.probes = ${global.systemEnv.pgIvfflatProbe || 10};
+    select kb_id,id,q,a,source from modelData where kb_id IN (${kbList
+      .map((item) => `'${item.kbId}'`)
+      .join(',')}) AND vector <#> '[${vectors[0]}]' < -${similarity} order by vector <#> '[${
+      vectors[0]
+    }]' limit ${limit};
+    COMMIT;`
+  );
+
+  const searchRes: QuoteItemType[] = res?.[2]?.rows || [];
+
+  return {
+    isEmpty: searchRes.length === 0 ? true : undefined,
+    unEmpty: searchRes.length > 0 ? true : undefined,
+    quoteQA: searchRes,
+    responseData: {
+      moduleName,
+      price: countModelPrice({ model: vectorModel.model, tokens: tokenLen }),
+      model: vectorModel.name,
+      tokens: tokenLen,
+      similarity,
+      limit
+    }
+  };
+}

client/src/service/moduleDispatch/tools/answer.ts

@@ -0,0 +1,31 @@
+import { sseResponseEventEnum, TaskResponseKeyEnum } from '@/constants/chat';
+import { sseResponse } from '@/service/utils/tools';
+import { textAdaptGptResponse } from '@/utils/adapt';
+import type { NextApiResponse } from 'next';
+
+export type AnswerProps = {
+  res: NextApiResponse;
+  text: string;
+  stream: boolean;
+};
+export type AnswerResponse = {
+  [TaskResponseKeyEnum.answerText]: string;
+};
+
+export const dispatchAnswer = (props: Record<string, any>): AnswerResponse => {
+  const { res, text = '', stream } = props as AnswerProps;
+
+  if (stream) {
+    sseResponse({
+      res,
+      event: sseResponseEventEnum.answer,
+      data: textAdaptGptResponse({
+        text: text.replace(/\\n/g, '\n')
+      })
+    });
+  }
+
+  return {
+    [TaskResponseKeyEnum.answerText]: text
+  };
+};