feat: function call prompt version (#331)

client/data/config.json

@@ -62,5 +62,21 @@
     "name": "GPT35-16k",
     "maxToken": 16000,
     "price": 0
+  },
+  "ExtractModel": {
+    "model": "gpt-3.5-turbo-16k",
+    "functionCall": false,
+    "name": "GPT35-16k",
+    "maxToken": 16000,
+    "price": 0,
+    "prompt": ""
+  },
+  "CQModel": {
+    "model": "gpt-3.5-turbo-16k",
+    "functionCall": false,
+    "name": "GPT35-16k",
+    "maxToken": 16000,
+    "price": 0,
+    "prompt": ""
   }
 }

client/public/docs/versionIntro.md

@@ -1,4 +1,4 @@
-### Fast GPT V4.4.3
+### Fast GPT V4.4.4
 
 1. 去除 - 限定词。目前旧应用仍生效，9/25 后全面去除，请及时替换。
 2. 新增 - 引用模板/引用提示词设置，可以 DIY 引用内容的格式，从而更好的适配场景。

client/src/components/ChatBox/ResponseTags.tsx

@@ -56,7 +56,7 @@ const ResponseTags = ({
 
   return responseData.length === 0 ? null : (
     <Flex alignItems={'center'} mt={2} flexWrap={'wrap'}>
-      {chatAccount === 1 ? (
+      {chatAccount === 1 && (
         <>
           {quoteList.length > 0 && (
             <MyTooltip label="查看引用">
@@ -83,7 +83,8 @@ const ResponseTags = ({
             </MyTooltip>
           )}
         </>
-      ) : (
+      )}
+      {chatAccount > 1 && (
         <Tag colorSchema="blue" {...TagStyles}>
           多组 AI 对话
         </Tag>

client/src/pages/api/system/getInitData.ts

@@ -5,7 +5,8 @@ import { readFileSync } from 'fs';
 import {
   type QAModelItemType,
   type ChatModelItemType,
-  type VectorModelItemType
+  type VectorModelItemType,
+  FunctionModelItemType
 } from '@/types/model';
 
 export type InitDateResponse = {
@@ -83,6 +84,22 @@ const defaultQAModel = {
   maxToken: 16000,
   price: 0
 };
+const defaultExtractModel: FunctionModelItemType = {
+  model: 'gpt-3.5-turbo-16k',
+  name: 'GPT35-16k',
+  maxToken: 16000,
+  price: 0,
+  prompt: '',
+  functionCall: true
+};
+const defaultCQModel: FunctionModelItemType = {
+  model: 'gpt-3.5-turbo-16k',
+  name: 'GPT35-16k',
+  maxToken: 16000,
+  price: 0,
+  prompt: '',
+  functionCall: true
+};
 
 const defaultVectorModels: VectorModelItemType[] = [
   {
@@ -114,6 +131,8 @@ export async function getInitConfig() {
     global.feConfigs = res.FeConfig ? { ...defaultFeConfigs, ...res.FeConfig } : defaultFeConfigs;
     global.chatModels = res.ChatModels || defaultChatModels;
     global.qaModel = res.QAModel || defaultQAModel;
+    global.extractModel = res.ExtractModel || defaultExtractModel;
+    global.cqModel = res.CQModel || defaultCQModel;
     global.vectorModels = res.VectorModels || defaultVectorModels;
   } catch (error) {
     setDefaultData();

client/src/prompts/core/agent.ts

@@ -14,3 +14,45 @@ A2:
 我的文本："""{{text}}"""`,
   defaultTheme: '它们可能包含多个主题内容'
 };
+
+export const Prompt_ExtractJson = `你可以从 "对话记录" 中提取指定信息，并返回一个 JSON 对象，JSON 对象要求：
+1. JSON 对象仅包含字段说明中的值。
+2. 字段说明中的 required 决定 JSON 对象是否必须存在该字段。
+3. 必须存在的字段，值可以为空字符串或根据提取要求来设置，不能随机生成值。
+
+提取要求:
+"""
+{{description}}
+"""
+
+字段说明: 
+"""
+{{json}}
+"""
+
+对话记录:
+"""
+{{text}}
+"""
+`;
+
+export const Prompt_CQJson = `我会给你几个问题类型，请参考额外的背景知识（可能为空）和对话内容，判断我本次的问题类型，并返回对应类型的 ID，格式为 JSON 字符串:
+"""
+'{"type":"问题类型的 ID"}'
+"""
+
+问题类型：
+"""
+{{typeList}}
+"""
+
+额外背景知识:
+"""
+{{systemPrompt}}
+"""
+
+对话内容：
+"""
+{{text}}
+"""
+`;

client/src/service/events/pushBill.ts

@@ -20,40 +20,44 @@ export const pushTaskBill = async ({
   shareId?: string;
   response: ChatHistoryItemResType[];
 }) => {
-  const total = response.reduce((sum, item) => sum + item.price, 0);
+  try {
+    const total = response.reduce((sum, item) => sum + item.price, 0);
 
-  await Promise.allSettled([
+    await Promise.allSettled([
-    Bill.create({
+      Bill.create({
-      userId,
+        userId,
-      appName,
+        appName,
-      appId,
+        appId,
-      total,
+        total,
+        source,
+        list: response.map((item) => ({
+          moduleName: item.moduleName,
+          amount: item.price || 0,
+          model: item.model,
+          tokenLen: item.tokens
+        }))
+      }),
+      User.findByIdAndUpdate(userId, {
+        $inc: { balance: -total }
+      }),
+      ...(shareId
+        ? [
+            updateShareChatBill({
+              shareId,
+              total
+            })
+          ]
+        : [])
+    ]);
+
+    addLog.info(`finish completions`, {
       source,
-      list: response.map((item) => ({
+      userId,
-        moduleType: item.moduleType,
+      price: formatPrice(total)
-        amount: item.price || 0,
+    });
-        model: item.model,
+  } catch (error) {
-        tokenLen: item.tokens
+    addLog.error(`pushTaskBill error`, error);
-      }))
+  }
-    }),
-    User.findByIdAndUpdate(userId, {
-      $inc: { balance: -total }
-    }),
-    ...(shareId
-      ? [
-          updateShareChatBill({
-            shareId,
-            total
-          })
-        ]
-      : [])
-  ]);
-
-  addLog.info(`finish completions`, {
-    source,
-    userId,
-    price: formatPrice(total)
-  });
 };
 
 export const updateShareChatBill = async ({

client/src/service/models/bill.ts

@@ -31,24 +31,7 @@ const BillSchema = new Schema({
     default: BillSourceEnum.fastgpt
   },
   list: {
-    type: [
+    type: Array,
-      {
-        moduleName: {
-          type: String,
-          required: true
-        },
-        amount: {
-          type: Number,
-          required: true
-        },
-        model: {
-          type: String
-        },
-        tokenLen: {
-          type: Number
-        }
-      }
-    ],
     default: []
   }
 });

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

@@ -4,40 +4,69 @@ import type { ChatHistoryItemResType, ChatItemType } from '@/types/chat';
 import { ChatRoleEnum, TaskResponseKeyEnum } from '@/constants/chat';
 import { getAIChatApi, axiosConfig } from '@/service/lib/openai';
 import type { ClassifyQuestionAgentItemType } from '@/types/app';
-import { countModelPrice } from '@/service/events/pushBill';
-import { getModel } from '@/service/utils/data';
 import { SystemInputEnum } from '@/constants/app';
 import { SpecialInputKeyEnum } from '@/constants/flow';
 import { FlowModuleTypeEnum } from '@/constants/flow';
 import { ModuleDispatchProps } from '@/types/core/modules';
+import { replaceVariable } from '@/utils/common/tools/text';
+import { Prompt_CQJson } from '@/prompts/core/agent';
 
-export type CQProps = ModuleDispatchProps<{
+type Props = ModuleDispatchProps<{
   systemPrompt?: string;
   history?: ChatItemType[];
   [SystemInputEnum.userChatInput]: string;
   [SpecialInputKeyEnum.agents]: ClassifyQuestionAgentItemType[];
 }>;
-export type CQResponse = {
+type CQResponse = {
   [TaskResponseKeyEnum.responseData]: ChatHistoryItemResType;
   [key: string]: any;
 };
 
-const agentModel = 'gpt-3.5-turbo';
 const agentFunName = 'agent_user_question';
-const maxTokens = 3000;
 
 /* request openai chat */
-export const dispatchClassifyQuestion = async (props: Record<string, any>): Promise<CQResponse> => {
+export const dispatchClassifyQuestion = async (props: Props): Promise<CQResponse> => {
   const {
     moduleName,
     userOpenaiAccount,
-    inputs: { agents, systemPrompt, history = [], userChatInput }
+    inputs: { agents, userChatInput }
-  } = props as CQProps;
+  } = props as Props;
 
   if (!userChatInput) {
     return Promise.reject('Input is empty');
   }
 
+  const cqModel = global.cqModel;
+
+  const { arg, tokens } = await (async () => {
+    if (cqModel.functionCall) {
+      return functionCall(props);
+    }
+    return completions(props);
+  })();
+
+  const result = agents.find((item) => item.key === arg?.type) || agents[0];
+
+  return {
+    [result.key]: 1,
+    [TaskResponseKeyEnum.responseData]: {
+      moduleType: FlowModuleTypeEnum.classifyQuestion,
+      moduleName,
+      price: userOpenaiAccount?.key ? 0 : cqModel.price * tokens,
+      model: cqModel.name || '',
+      tokens,
+      cqList: agents,
+      cqResult: result.value
+    }
+  };
+};
+
+async function functionCall({
+  userOpenaiAccount,
+  inputs: { agents, systemPrompt, history = [], userChatInput }
+}: Props) {
+  const cqModel = global.cqModel;
+
   const messages: ChatItemType[] = [
     ...(systemPrompt
       ? [
@@ -55,14 +84,14 @@ export const dispatchClassifyQuestion = async (props: Record<string, any>): Prom
   ];
   const filterMessages = ChatContextFilter({
     messages,
-    maxTokens
+    maxTokens: cqModel.maxToken
   });
   const adaptMessages = adaptChat2GptMessages({ messages: filterMessages, reserveId: false });
 
   //   function body
   const agentFunction = {
     name: agentFunName,
-    description: '判断用户问题的类型属于哪方面，返回对应的枚举字段',
+    description: '判断用户问题的类型属于哪方面，返回对应的字段',
     parameters: {
       type: 'object',
       properties: {
@@ -79,7 +108,7 @@ export const dispatchClassifyQuestion = async (props: Record<string, any>): Prom
 
   const response = await chatAPI.createChatCompletion(
     {
-      model: agentModel,
+      model: cqModel.model,
       temperature: 0,
       messages: [...adaptMessages],
       function_call: { name: agentFunName },
@@ -92,20 +121,51 @@ export const dispatchClassifyQuestion = async (props: Record<string, any>): Prom
 
   const arg = JSON.parse(response.data.choices?.[0]?.message?.function_call?.arguments || '');
 
-  const tokens = response.data.usage?.total_tokens || 0;
+  return {
+    arg,
+    tokens: response.data.usage?.total_tokens || 0
+  };
+}
 
-  const result = agents.find((item) => item.key === arg?.type) || agents[0];
+async function completions({
+  userOpenaiAccount,
+  inputs: { agents, systemPrompt = '', history = [], userChatInput }
+}: Props) {
+  const extractModel = global.extractModel;
+
+  const messages: ChatItemType[] = [
+    {
+      obj: ChatRoleEnum.Human,
+      value: replaceVariable(extractModel.prompt || Prompt_CQJson, {
+        systemPrompt,
+        typeList: agents.map((item) => `ID: "${item.key}", 问题类型:${item.value}`).join('\n'),
+        text: `${history.map((item) => `${item.obj}:${item.value}`).join('\n')}
+Human:${userChatInput}`
+      })
+    }
+  ];
+
+  const chatAPI = getAIChatApi(userOpenaiAccount);
+
+  const { data } = await chatAPI.createChatCompletion(
+    {
+      model: extractModel.model,
+      temperature: 0.01,
+      messages: adaptChat2GptMessages({ messages, reserveId: false }),
+      stream: false
+    },
+    {
+      timeout: 480000,
+      ...axiosConfig()
+    }
+  );
+  const answer = data.choices?.[0].message?.content || '';
+  const totalTokens = data.usage?.total_tokens || 0;
+
+  const id = agents.find((item) => answer.includes(item.key))?.key || '';
 
   return {
-    [result.key]: 1,
+    tokens: totalTokens,
-    [TaskResponseKeyEnum.responseData]: {
+    arg: { type: id }
-      moduleType: FlowModuleTypeEnum.classifyQuestion,
-      moduleName,
-      price: userOpenaiAccount?.key ? 0 : countModelPrice({ model: agentModel, tokens }),
-      model: getModel(agentModel)?.name || agentModel,
-      tokens,
-      cqList: agents,
-      cqResult: result.value
-    }
   };
-};
+}

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

@@ -5,36 +5,88 @@ import { ChatRoleEnum, TaskResponseKeyEnum } from '@/constants/chat';
 import { getAIChatApi, axiosConfig } from '@/service/lib/openai';
 import type { ContextExtractAgentItemType } from '@/types/app';
 import { ContextExtractEnum } from '@/constants/flow/flowField';
-import { countModelPrice } from '@/service/events/pushBill';
-import { getModel } from '@/service/utils/data';
 import { FlowModuleTypeEnum } from '@/constants/flow';
 import { ModuleDispatchProps } from '@/types/core/modules';
+import { Prompt_ExtractJson } from '@/prompts/core/agent';
+import { replaceVariable } from '@/utils/common/tools/text';
 
-export type Props = ModuleDispatchProps<{
+type Props = ModuleDispatchProps<{
   history?: ChatItemType[];
   [ContextExtractEnum.content]: string;
   [ContextExtractEnum.extractKeys]: ContextExtractAgentItemType[];
   [ContextExtractEnum.description]: string;
 }>;
-export type Response = {
+type Response = {
   [ContextExtractEnum.success]?: boolean;
   [ContextExtractEnum.failed]?: boolean;
   [ContextExtractEnum.fields]: string;
   [TaskResponseKeyEnum.responseData]: ChatHistoryItemResType;
 };
 
-const agentModel = 'gpt-3.5-turbo';
 const agentFunName = 'agent_extract_data';
-const maxTokens = 4000;
 
-export async function dispatchContentExtract({
+export async function dispatchContentExtract(props: Props): Promise<Response> {
-  moduleName,
+  const {
-  userOpenaiAccount,
+    moduleName,
-  inputs: { content, extractKeys, history = [], description }
+    userOpenaiAccount,
-}: Props): Promise<Response> {
+    inputs: { content, description, extractKeys }
+  } = props;
+
   if (!content) {
     return Promise.reject('Input is empty');
   }
+
+  const extractModel = global.extractModel;
+
+  const { arg, tokens } = await (async () => {
+    if (extractModel.functionCall) {
+      return functionCall(props);
+    }
+    return completions(props);
+  })();
+
+  // remove invalid key
+  for (let key in arg) {
+    if (!extractKeys.find((item) => item.key === key)) {
+      delete arg[key];
+    }
+  }
+
+  // auth fields
+  let success = !extractKeys.find((item) => !arg[item.key]);
+  // auth empty value
+  if (success) {
+    for (const key in arg) {
+      if (arg[key] === '') {
+        success = false;
+        break;
+      }
+    }
+  }
+
+  return {
+    [ContextExtractEnum.success]: success ? true : undefined,
+    [ContextExtractEnum.failed]: success ? undefined : true,
+    [ContextExtractEnum.fields]: JSON.stringify(arg),
+    ...arg,
+    [TaskResponseKeyEnum.responseData]: {
+      moduleType: FlowModuleTypeEnum.contentExtract,
+      moduleName,
+      price: userOpenaiAccount?.key ? 0 : extractModel.price * tokens,
+      model: extractModel.name || '',
+      tokens,
+      extractDescription: description,
+      extractResult: arg
+    }
+  };
+}
+
+async function functionCall({
+  userOpenaiAccount,
+  inputs: { history = [], content, extractKeys, description }
+}: Props) {
+  const extractModel = global.extractModel;
+
   const messages: ChatItemType[] = [
     ...history,
     {
@@ -44,7 +96,7 @@ export async function dispatchContentExtract({
   ];
   const filterMessages = ChatContextFilter({
     messages,
-    maxTokens
+    maxTokens: extractModel.maxToken
   });
   const adaptMessages = adaptChat2GptMessages({ messages: filterMessages, reserveId: false });
 
@@ -77,7 +129,7 @@ export async function dispatchContentExtract({
 
   const response = await chatAPI.createChatCompletion(
     {
-      model: agentModel,
+      model: extractModel.model,
       temperature: 0,
       messages: [...adaptMessages],
       function_call: { name: agentFunName },
@@ -96,33 +148,79 @@ export async function dispatchContentExtract({
     }
   })();
 
-  // auth fields
-  let success = !extractKeys.find((item) => !arg[item.key]);
-  // auth empty value
-  if (success) {
-    for (const key in arg) {
-      if (arg[key] === '') {
-        success = false;
-        break;
-      }
-    }
-  }
-
   const tokens = response.data.usage?.total_tokens || 0;
-
   return {
-    [ContextExtractEnum.success]: success ? true : undefined,
+    tokens,
-    [ContextExtractEnum.failed]: success ? undefined : true,
+    arg
-    [ContextExtractEnum.fields]: JSON.stringify(arg),
-    ...arg,
-    [TaskResponseKeyEnum.responseData]: {
-      moduleType: FlowModuleTypeEnum.contentExtract,
-      moduleName,
-      price: userOpenaiAccount?.key ? 0 : countModelPrice({ model: agentModel, tokens }),
-      model: getModel(agentModel)?.name || agentModel,
-      tokens,
-      extractDescription: description,
-      extractResult: arg
-    }
   };
 }
+
+async function completions({
+  userOpenaiAccount,
+  inputs: { history = [], content, extractKeys, description }
+}: Props) {
+  const extractModel = global.extractModel;
+
+  const messages: ChatItemType[] = [
+    {
+      obj: ChatRoleEnum.Human,
+      value: replaceVariable(extractModel.prompt || Prompt_ExtractJson, {
+        description,
+        json: extractKeys
+          .map(
+            (item) =>
+              `key="${item.key}"，描述="${item.desc}"，required="${
+                item.required ? 'true' : 'false'
+              }"`
+          )
+          .join('\n'),
+        text: `${history.map((item) => `${item.obj}:${item.value}`).join('\n')}
+Human: ${content}`
+      })
+    }
+  ];
+
+  const chatAPI = getAIChatApi(userOpenaiAccount);
+
+  const { data } = await chatAPI.createChatCompletion(
+    {
+      model: extractModel.model,
+      temperature: 0.01,
+      messages: adaptChat2GptMessages({ messages, reserveId: false }),
+      stream: false
+    },
+    {
+      timeout: 480000,
+      ...axiosConfig()
+    }
+  );
+  const answer = data.choices?.[0].message?.content || '';
+  const totalTokens = data.usage?.total_tokens || 0;
+
+  // parse response
+  const start = answer.indexOf('{');
+  const end = answer.lastIndexOf('}');
+
+  if (start === -1 || end === -1)
+    return {
+      tokens: totalTokens,
+      arg: {}
+    };
+
+  const jsonStr = answer
+    .substring(start, end + 1)
+    .replace(/(\\n|\\)/g, '')
+    .replace(/  /g, '');
+
+  try {
+    return {
+      tokens: totalTokens,
+      arg: JSON.parse(jsonStr) as Record<string, any>
+    };
+  } catch (error) {
+    return {
+      tokens: totalTokens,
+      arg: {}
+    };
+  }
+}

client/src/service/utils/data.ts

@@ -14,5 +14,11 @@ export const getVectorModel = (model?: string) => {
 };
 
 export const getModel = (model?: string) => {
-  return [...global.chatModels, ...global.vectorModels].find((item) => item.model === model);
+  return [
+    ...global.chatModels,
+    ...global.vectorModels,
+    global.qaModel,
+    global.extractModel,
+    global.cqModel
+  ].find((item) => item.model === model);
 };

client/src/types/index.d.ts

@@ -3,7 +3,12 @@ import type { Agent } from 'http';
 import type { Pool } from 'pg';
 import type { Tiktoken } from 'js-tiktoken';
 import type { Logger } from 'winston';
-import { ChatModelItemType, QAModelItemType, VectorModelItemType } from './model';
+import {
+  ChatModelItemType,
+  FunctionModelItemType,
+  QAModelItemType,
+  VectorModelItemType
+} from './model';
 import { TrackEventName } from '@/constants/common';
 
 export type PagingData<T> = {
@@ -62,6 +67,8 @@ declare global {
   var systemEnv: SystemEnvType;
   var chatModels: ChatModelItemType[];
   var qaModel: QAModelItemType;
+  var extractModel: FunctionModelItemType;
+  var cqModel: FunctionModelItemType;
   var vectorModels: VectorModelItemType[];
   var systemVersion: string;
 

client/src/types/model.d.ts

@@ -21,3 +21,11 @@ export type VectorModelItemType = {
   price: number;
   maxToken: number;
 };
+export type FunctionModelItemType = {
+  model: string;
+  name: string;
+  maxToken: number;
+  price: number;
+  prompt: string;
+  functionCall: boolean;
+};