V4.7-alpha (#985)

Co-authored-by: heheer <[email protected]>

Author: c121914yu

docSite/content/docs/development/configuration.md

@@ -20,14 +20,13 @@ llm模型全部合并
 ```json
 {
   "systemEnv": {
-    "openapiPrefix": "fastgpt",
     "vectorMaxProcess": 15,
     "qaMaxProcess": 15,
-    "pgHNSWEfSearch": 100
+    "pgHNSWEfSearch": 100 // 向量搜索参数。越大，搜索越精确，但是速度越慢。设置为100，有99%+精度。
   },
   "llmModels": [
     {
-      "model": "gpt-3.5-turbo-1106", // 模型名
+      "model": "gpt-3.5-turbo", // 模型名
       "name": "gpt-3.5-turbo", // 别名
       "maxContext": 16000, // 最大上下文
       "maxResponse": 4000, // 最大回复
@@ -37,12 +36,16 @@ llm模型全部合并
       "censor": false,
       "vision": false, // 是否支持图片输入
       "datasetProcess": false, // 是否设置为知识库处理模型（QA），务必保证至少有一个为true，否则知识库会报错
-      "toolChoice": true, // 是否支持工具选择
-      "functionCall": false, // 是否支持函数调用
+      "usedInClassify": true, // 是否用于问题分类（务必保证至少有一个为true）
+      "usedInExtractFields": true, // 是否用于内容提取（务必保证至少有一个为true）
+      "useInToolCall": true, // 是否用于工具调用（务必保证至少有一个为true）
+      "usedInQueryExtension": true, // 是否用于问题优化（务必保证至少有一个为true）
+      "toolChoice": true, // 是否支持工具选择（务必保证至少有一个为true）
+      "functionCall": false, // 是否支持函数调用（特殊功能，会优先使用 toolChoice，如果为false，则使用 functionCall，如果仍为 false，则使用提示词模式）
       "customCQPrompt": "", // 自定义文本分类提示词（不支持工具和函数调用的模型
       "customExtractPrompt": "", // 自定义内容提取提示词
       "defaultSystemChatPrompt": "", // 对话默认携带的系统提示词
-      "defaultConfig":{}  // 对话默认配置（比如 GLM4 的 top_p
+      "defaultConfig":{}  // LLM默认配置，可以针对不同模型设置特殊值（比如 GLM4 的 top_p
     },
     {
       "model": "gpt-3.5-turbo-16k",
@@ -55,6 +58,10 @@ llm模型全部合并
       "censor": false,
       "vision": false,
       "datasetProcess": true,
+      "usedInClassify": true,
+      "usedInExtractFields": true,
+      "useInToolCall": true,
+      "usedInQueryExtension": true,
       "toolChoice": true,
       "functionCall": false,
       "customCQPrompt": "",
@@ -73,6 +80,10 @@ llm模型全部合并
       "censor": false,
       "vision": false,
       "datasetProcess": false,
+      "usedInClassify": true,
+      "usedInExtractFields": true,
+      "useInToolCall": true,
+      "usedInQueryExtension": true,
       "toolChoice": true,
       "functionCall": false,
       "customCQPrompt": "",
@@ -91,6 +102,10 @@ llm模型全部合并
       "censor": false,
       "vision": true,
       "datasetProcess": false,
+      "usedInClassify": false,
+      "usedInExtractFields": false,
+      "useInToolCall": false,
+      "usedInQueryExtension": false,
       "toolChoice": true,
       "functionCall": false,
       "customCQPrompt": "",

docSite/content/docs/development/one-api.md

@@ -120,6 +120,10 @@ CHAT_API_KEY=sk-xxxxxx
       "censor": false,
       "vision": false, // 是否支持图片输入
       "datasetProcess": false, // 是否设置为知识库处理模型
+      "usedInClassify": true, // 是否用于问题分类
+      "usedInExtractFields": true, // 是否用于字段提取
+      "useInToolCall": true, // 是否用于工具调用
+      "usedInQueryExtension": true, // 是否用于问题优化
       "toolChoice": true, // 是否支持工具选择
       "functionCall": false, // 是否支持函数调用
       "customCQPrompt": "", // 自定义文本分类提示词（不支持工具和函数调用的模型

docSite/content/docs/development/upgrading/47.md

@@ -0,0 +1,19 @@
+---
+title: 'V4.7(进行中)'
+description: 'FastGPT V4.7更新说明'
+icon: 'upgrade'
+draft: false
+toc: true
+weight: 826
+---
+
+## 修改配置文件
+
+增加一些 Boolean 值，用于决定不同功能块可以使用哪些模型：[点击查看最新的配置文件](/docs/development/configuration/)
+
+
+## V4.7 更新说明
+
+1. 新增 - 工具调用模块，可以让LLM模型根据用户意图，动态的选择其他模型或插件执行。
+2. 新增 - 分类和内容提取支持 functionCall 模式。部分模型支持 functionCall 不支持 ToolCall，也可以使用了。需要把 LLM 模型配置文件里的 `functionCall` 设置为 `true`， `toolChoice`设置为 `false`。如果 `toolChoice` 为 true，会走 tool 模式。
+3. 优化 - 高级编排性能

package.json

@@ -6,7 +6,7 @@
     "prepare": "husky install",
     "format-code": "prettier --config \"./.prettierrc.js\" --write \"./**/src/**/*.{ts,tsx,scss}\"",
     "format-doc": "zhlint --dir ./docSite *.md --fix",
-    "gen:theme-typings": "chakra-cli tokens projects/app/src/web/styles/theme.ts --out node_modules/.pnpm/node_modules/@chakra-ui/styled-system/dist/theming.types.d.ts",
+    "gen:theme-typings": "chakra-cli tokens packages/web/styles/theme.ts --out node_modules/.pnpm/node_modules/@chakra-ui/styled-system/dist/theming.types.d.ts",
     "postinstall": "sh ./scripts/postinstall.sh",
     "initIcon": "node ./scripts/icon/init.js",
     "previewIcon": "node ./scripts/icon/index.js"

packages/global/common/string/tiktoken/index.ts

@@ -1,10 +1,15 @@
 /* Only the token of gpt-3.5-turbo is used */
 import type { ChatItemType } from '../../../core/chat/type';
 import { Tiktoken } from 'js-tiktoken/lite';
-import { adaptChat2GptMessages } from '../../../core/chat/adapt';
-import { ChatCompletionRequestMessageRoleEnum } from '../../../core/ai/constant';
+import { chats2GPTMessages } from '../../../core/chat/adapt';
 import encodingJson from './cl100k_base.json';
-import { ChatMessageItemType } from '../../../core/ai/type';
+import {
+  ChatCompletionMessageParam,
+  ChatCompletionContentPart,
+  ChatCompletionCreateParams,
+  ChatCompletionTool
+} from '../../../core/ai/type';
+import { ChatCompletionRequestMessageRoleEnum } from '../../../core/ai/constants';
 
 /* init tikToken obj */
 export function getTikTokenEnc() {
@@ -29,18 +34,25 @@ export function getTikTokenEnc() {
 
 /* count one prompt tokens */
 export function countPromptTokens(
-  prompt = '',
-  role: '' | `${ChatCompletionRequestMessageRoleEnum}` = '',
-  tools?: any
+  prompt: string | ChatCompletionContentPart[] | null | undefined = '',
+  role: '' | `${ChatCompletionRequestMessageRoleEnum}` = ''
 ) {
   const enc = getTikTokenEnc();
-  const toolText = tools
-    ? JSON.stringify(tools)
-        .replace('"', '')
-        .replace('\n', '')
-        .replace(/( ){2,}/g, ' ')
-    : '';
-  const text = `${role}\n${prompt}\n${toolText}`.trim();
+  const promptText = (() => {
+    if (!prompt) return '';
+    if (typeof prompt === 'string') return prompt;
+    let promptText = '';
+    prompt.forEach((item) => {
+      if (item.type === 'text') {
+        promptText += item.text;
+      } else if (item.type === 'image_url') {
+        promptText += item.image_url.url;
+      }
+    });
+    return promptText;
+  })();
+
+  const text = `${role}\n${promptText}`.trim();
 
   try {
     const encodeText = enc.encode(text);
@@ -50,15 +62,66 @@ export function countPromptTokens(
     return text.length;
   }
 }
+export const countToolsTokens = (
+  tools?: ChatCompletionTool[] | ChatCompletionCreateParams.Function[]
+) => {
+  if (!tools || tools.length === 0) return 0;
+
+  const enc = getTikTokenEnc();
+
+  const toolText = tools
+    ? JSON.stringify(tools)
+        .replace('"', '')
+        .replace('\n', '')
+        .replace(/( ){2,}/g, ' ')
+    : '';
+
+  return enc.encode(toolText).length;
+};
 
 /* count messages tokens */
-export const countMessagesTokens = (messages: ChatItemType[], tools?: any) => {
-  const adaptMessages = adaptChat2GptMessages({ messages, reserveId: true });
+export const countMessagesTokens = (messages: ChatItemType[]) => {
+  const adaptMessages = chats2GPTMessages({ messages, reserveId: true });
 
-  return countGptMessagesTokens(adaptMessages, tools);
+  return countGptMessagesTokens(adaptMessages);
 };
-export const countGptMessagesTokens = (messages: ChatMessageItemType[], tools?: any) =>
-  messages.reduce((sum, item) => sum + countPromptTokens(item.content, item.role, tools), 0);
+export const countGptMessagesTokens = (
+  messages: ChatCompletionMessageParam[],
+  tools?: ChatCompletionTool[],
+  functionCall?: ChatCompletionCreateParams.Function[]
+) =>
+  messages.reduce((sum, item) => {
+    // Evaluates the text of toolcall and functioncall
+    const functionCallPrompt = (() => {
+      let prompt = '';
+      if (item.role === ChatCompletionRequestMessageRoleEnum.Assistant) {
+        const toolCalls = item.tool_calls;
+        prompt +=
+          toolCalls
+            ?.map((item) => `${item?.function?.name} ${item?.function?.arguments}`.trim())
+            ?.join('') || '';
+
+        const functionCall = item.function_call;
+        prompt += `${functionCall?.name} ${functionCall?.arguments}`.trim();
+      }
+      return prompt;
+    })();
+
+    const contentPrompt = (() => {
+      if (!item.content) return '';
+      if (typeof item.content === 'string') return item.content;
+      return item.content
+        .map((item) => {
+          if (item.type === 'text') return item.text;
+          return '';
+        })
+        .join('');
+    })();
+
+    return sum + countPromptTokens(`${contentPrompt}${functionCallPrompt}`, item.role);
+  }, 0) +
+  countToolsTokens(tools) +
+  countToolsTokens(functionCall);
 
 /* slice messages from top to bottom by maxTokens */
 export function sliceMessagesTB({
@@ -68,7 +131,7 @@ export function sliceMessagesTB({
   messages: ChatItemType[];
   maxTokens: number;
 }) {
-  const adaptMessages = adaptChat2GptMessages({ messages, reserveId: true });
+  const adaptMessages = chats2GPTMessages({ messages, reserveId: true });
   let reduceTokens = maxTokens;
   let result: ChatItemType[] = [];
 

packages/global/core/ai/constant.ts

@@ -0,0 +1,27 @@
+export enum ChatCompletionRequestMessageRoleEnum {
+  'System' = 'system',
+  'User' = 'user',
+  'Assistant' = 'assistant',
+  'Function' = 'function',
+  'Tool' = 'tool'
+}
+
+export enum ChatMessageTypeEnum {
+  text = 'text',
+  image_url = 'image_url'
+}
+
+export enum LLMModelTypeEnum {
+  all = 'all',
+  classify = 'classify',
+  extractFields = 'extractFields',
+  toolCall = 'toolCall',
+  queryExtension = 'queryExtension'
+}
+export const llmModelTypeFilterMap = {
+  [LLMModelTypeEnum.all]: 'model',
+  [LLMModelTypeEnum.classify]: 'usedInClassify',
+  [LLMModelTypeEnum.extractFields]: 'usedInExtractFields',
+  [LLMModelTypeEnum.toolCall]: 'usedInToolCall',
+  [LLMModelTypeEnum.queryExtension]: 'usedInQueryExtension'
+};

packages/global/core/ai/constants.ts

@@ -10,7 +10,13 @@ export type LLMModelItemType = {
 
   censor?: boolean;
   vision?: boolean;
-  datasetProcess?: boolean;
+
+  // diff function model
+  datasetProcess?: boolean; // dataset
+  usedInClassify?: boolean; // classify
+  usedInExtractFields?: boolean; // extract fields
+  usedInToolCall?: boolean; // tool call
+  usedInQueryExtension?: boolean; // query extension
 
   functionCall: boolean;
   toolChoice: boolean;

packages/global/core/ai/model.d.ts

@@ -1,24 +1,40 @@
+import openai from 'openai';
 import type {
-  ChatCompletion,
-  ChatCompletionCreateParams,
+  ChatCompletionMessageToolCall,
   ChatCompletionChunk,
   ChatCompletionMessageParam,
-  ChatCompletionContentPart
+  ChatCompletionToolMessageParam,
+  ChatCompletionAssistantMessageParam
 } from 'openai/resources';
+import { ChatMessageTypeEnum } from './constants';
 
-export type ChatCompletionContentPart = ChatCompletionContentPart;
-export type ChatCompletionCreateParams = ChatCompletionCreateParams;
-export type ChatMessageItemType = Omit<ChatCompletionMessageParam, 'name'> & {
-  name?: any;
+export * from 'openai/resources';
+
+export type ChatCompletionMessageParam = ChatCompletionMessageParam & {
   dataId?: string;
-  content: any;
-} & any;
+};
+export type ChatCompletionToolMessageParam = ChatCompletionToolMessageParam & { name: string };
+export type ChatCompletionAssistantToolParam = {
+  role: 'assistant';
+  tool_calls: ChatCompletionMessageToolCall[];
+};
 
-export type ChatCompletion = ChatCompletion;
+export type ChatCompletionMessageToolCall = ChatCompletionMessageToolCall & {
+  toolName?: string;
+  toolAvatar?: string;
+};
+export type ChatCompletionMessageFunctionCall = ChatCompletionAssistantMessageParam.FunctionCall & {
+  id?: string;
+  toolName?: string;
+  toolAvatar?: string;
+};
 export type StreamChatType = Stream<ChatCompletionChunk>;
 
 export type PromptTemplateItem = {
   title: string;
   desc: string;
   value: string;
 };
+
+export default openai;
+export * from 'openai';