perf: tool promot and reg slice;query extension prompt (#1576)

labring · May 23, 2024 · c4ce123 · c4ce123
1 parent 4eb2c9b
commit c4ce123
Show file tree

Hide file tree

Showing 6 changed files with 85 additions and 45 deletions.
diff --git a/packages/global/common/string/tools.ts b/packages/global/common/string/tools.ts
@@ -64,4 +64,22 @@ export const getNanoid = (size = 12) => {
   return `${firstChar}${randomsStr}`;
 };
 
+/* Custom text to reg, need to replace special chats */
 export const replaceRegChars = (text: string) => text.replace(/[.*+?^${}()|[\]\\]/g, '\\$&');
+
+/* slice json str */
+export const sliceJsonStr = (str: string) => {
+  str = str.replace(/(\\n|\\)/g, '').replace(/  /g, '');
+
+  const jsonRegex = /{[^{}]*}/g;
+  const matches = str.match(jsonRegex);
+
+  if (!matches) {
+    return '';
+  }
+
+  // 找到第一个完整的 JSON 字符串
+  const jsonStr = matches[0];
+
+  return jsonStr;
+};
diff --git a/packages/global/core/workflow/template/system/runApp.ts b/packages/global/core/workflow/template/system/runApp.ts
@@ -24,6 +24,7 @@ export const RunAppModule: FlowNodeTemplateType = {
   intro: '可以选择一个其他应用进行调用',
   showStatus: true,
   version: '481',
+  isTool: true,
   inputs: [
     {
       key: NodeInputKeyEnum.runAppSelectApp,

diff --git a/packages/service/core/ai/functions/queryExtension.ts b/packages/service/core/ai/functions/queryExtension.ts
@@ -65,7 +65,7 @@ Q: FastGPT 如何收费？
 A: FastGPT 收费可以参考……
 """
 原问题: 你知道 laf 么？
-检索词: ["laf是什么？","如何使用laf？","laf的介绍。"]
+检索词: ["laf 的官网地址是多少？","laf 的使用教程。","laf 有什么特点和优势。"]
 ----------------
 历史记录:
 """
@@ -75,7 +75,7 @@ A: 1. 开源
    3. 扩展性强
 """
 原问题: 介绍下第2点。
-检索词: ["介绍下 FastGPT 简便的优势", "FastGPT 为什么使用起来简便？","FastGPT的有哪些简便的功能？"]。
+检索词: ["介绍下 FastGPT 简便的优势"]。
 ----------------
 历史记录:
 """
@@ -85,7 +85,7 @@ Q: 什么是 Laf？
 A: Laf 是一个云函数开发平台。
 """
 原问题: 它们有什么关系？
-检索词: ["FastGPT和Laf有什么关系？","FastGPT的RAG是用Laf实现的么？"]
+检索词: ["FastGPT和Laf有什么关系？","介绍下FastGPT","介绍下Laf"]
 ----------------
 历史记录:
 """

diff --git a/packages/service/core/workflow/dispatch/agent/extract.ts b/packages/service/core/workflow/dispatch/agent/extract.ts
@@ -12,7 +12,7 @@ import { NodeInputKeyEnum, NodeOutputKeyEnum } from '@fastgpt/global/core/workfl
 import { DispatchNodeResponseKeyEnum } from '@fastgpt/global/core/workflow/runtime/constants';
 import type { ModuleDispatchProps } from '@fastgpt/global/core/workflow/type/index.d';
 import { Prompt_ExtractJson } from '@fastgpt/global/core/ai/prompt/agent';
-import { replaceVariable } from '@fastgpt/global/common/string/tools';
+import { replaceVariable, sliceJsonStr } from '@fastgpt/global/common/string/tools';
 import { LLMModelItemType } from '@fastgpt/global/core/ai/model.d';
 import { getHistories } from '../utils';
 import { ModelTypeEnum, getLLMModel } from '../../../ai/model';
@@ -348,22 +348,16 @@ Human: ${content}`
   const answer = data.choices?.[0].message?.content || '';
 
   // parse response
-  const start = answer.indexOf('{');
-  const end = answer.lastIndexOf('}');
+  const jsonStr = sliceJsonStr(answer);
 
-  if (start === -1 || end === -1) {
+  if (!jsonStr) {
     return {
       rawResponse: answer,
       tokens: await countMessagesTokens(messages),
       arg: {}
     };
   }
 
-  const jsonStr = answer
-    .substring(start, end + 1)
-    .replace(/(\\n|\\)/g, '')
-    .replace(/  /g, '');
-
   try {
     return {
       rawResponse: answer,

diff --git a/packages/service/core/workflow/dispatch/agent/runTool/constants.ts b/packages/service/core/workflow/dispatch/agent/runTool/constants.ts
@@ -3,7 +3,7 @@ export const Prompt_Tool_Call = `<Instruction>
 
 工具使用了 JSON Schema 的格式声明，其中 toolId 是工具的 description 是工具的描述，parameters 是工具的参数，包括参数的类型和描述，required 是必填参数的列表。
 
-请你根据工具描述，决定回答问题或是使用工具。在完成任务过程中，USER代表用户的输入，TOOL_RESPONSE代表工具运行结果。ASSISTANT 代表你的输出。
+请你根据工具描述，决定回答问题或是使用工具。在完成任务过程中，USER代表用户的输入，TOOL_RESPONSE代表工具运行结果，ANSWER 代表你的输出。
 你的每次输出都必须以0,1开头，代表是否需要调用工具：
 0: 不使用工具，直接回答内容。
 1: 使用工具，返回工具调用的参数。
@@ -12,14 +12,20 @@ export const Prompt_Tool_Call = `<Instruction>
 
 USER: 你好呀
 ANSWER: 0: 你好，有什么可以帮助你的么？
-USER: 今天杭州的天气如何
-ANSWER: 1: {"toolId":"testToolId",arguments:{"city": "杭州"}}
+USER: 现在几点了？
+ANSWER:  1: {"toolId":"timeToolId"}
+TOOL_RESPONSE: """
+2022/5/5 12:00 Thursday
+"""
+ANSWER: 0: 现在是2022年5月5日，星期四，中午12点。
+USER: 今天杭州的天气如何？
+ANSWER: 1: {"toolId":"testToolId","arguments":{"city": "杭州"}}
 TOOL_RESPONSE: """
 晴天......
 """
 ANSWER: 0: 今天杭州是晴天。
 USER: 今天杭州的天气适合去哪里玩？
-ANSWER: 1: {"toolId":"testToolId2",arguments:{"query": "杭州 天气 去哪里玩"}}
+ANSWER: 1: {"toolId":"testToolId2","arguments":{"query": "杭州 天气 去哪里玩"}}
 TOOL_RESPONSE: """
 晴天. 西湖、灵隐寺、千岛湖……
 """
@@ -35,5 +41,4 @@ ANSWER: 0: 今天杭州是晴天，适合去西湖、灵隐寺、千岛湖等地
 下面是正式的对话内容：
 
 USER: {{question}}
-ANSWER: 
-`;
+ANSWER: `;
diff --git a/packages/service/core/workflow/dispatch/agent/runTool/promptCall.ts b/packages/service/core/workflow/dispatch/agent/runTool/promptCall.ts
@@ -20,7 +20,7 @@ import { dispatchWorkFlow } from '../../index';
 import { DispatchToolModuleProps, RunToolResponse, ToolNodeItemType } from './type.d';
 import json5 from 'json5';
 import { countGptMessagesTokens } from '../../../../../common/string/tiktoken/index';
-import { getNanoid, replaceVariable } from '@fastgpt/global/common/string/tools';
+import { getNanoid, replaceVariable, sliceJsonStr } from '@fastgpt/global/common/string/tools';
 import { AIChatItemType } from '@fastgpt/global/core/chat/type';
 import { GPTMessages2Chats } from '@fastgpt/global/core/chat/adapt';
 import { updateToolInputValue } from './utils';
@@ -33,6 +33,8 @@ type FunctionCallCompletion = {
   toolAvatar?: string;
 };
 
+const ERROR_TEXT = 'Tool run error';
+
 export const runToolWithPromptCall = async (
   props: DispatchToolModuleProps & {
     messages: ChatCompletionMessageParam[];
@@ -122,14 +124,23 @@ export const runToolWithPromptCall = async (
     }
   })();
 
-  const parseAnswerResult = parseAnswer(answer);
+  const { answer: replaceAnswer, toolJson } = parseAnswer(answer);
   // console.log(parseAnswer, '==11==');
   // No tools
-  if (typeof parseAnswerResult === 'string') {
+  if (!toolJson) {
+    if (replaceAnswer === ERROR_TEXT && stream && detail) {
+      responseWrite({
+        res,
+        event: SseResponseEventEnum.answer,
+        data: textAdaptGptResponse({
+          text: replaceAnswer
+        })
+      });
+    }
     // No tool is invoked, indicating that the process is over
     const gptAssistantResponse: ChatCompletionAssistantMessageParam = {
       role: ChatCompletionRequestMessageRoleEnum.Assistant,
-      content: parseAnswerResult
+      content: replaceAnswer
     };
     const completeMessages = filterMessages.concat(gptAssistantResponse);
     const tokens = await countGptMessagesTokens(completeMessages, undefined);
@@ -148,18 +159,16 @@ export const runToolWithPromptCall = async (
 
   // Run the selected tool.
   const toolsRunResponse = await (async () => {
-    if (!parseAnswerResult) return Promise.reject('tool run error');
-
-    const toolNode = toolNodes.find((item) => item.nodeId === parseAnswerResult.name);
+    const toolNode = toolNodes.find((item) => item.nodeId === toolJson.name);
     if (!toolNode) return Promise.reject('tool not found');
 
-    parseAnswerResult.toolName = toolNode.name;
-    parseAnswerResult.toolAvatar = toolNode.avatar;
+    toolJson.toolName = toolNode.name;
+    toolJson.toolAvatar = toolNode.avatar;
 
     // run tool flow
     const startParams = (() => {
       try {
-        return json5.parse(parseAnswerResult.arguments);
+        return json5.parse(toolJson.arguments);
       } catch (error) {
         return {};
       }
@@ -172,11 +181,11 @@ export const runToolWithPromptCall = async (
         event: SseResponseEventEnum.toolCall,
         data: JSON.stringify({
           tool: {
-            id: parseAnswerResult.id,
+            id: toolJson.id,
             toolName: toolNode.name,
             toolAvatar: toolNode.avatar,
-            functionName: parseAnswerResult.name,
-            params: parseAnswerResult.arguments,
+            functionName: toolJson.name,
+            params: toolJson.arguments,
             response: ''
           }
         })
@@ -211,7 +220,7 @@ export const runToolWithPromptCall = async (
         event: SseResponseEventEnum.toolResponse,
         data: JSON.stringify({
           tool: {
-            id: parseAnswerResult.id,
+            id: toolJson.id,
             toolName: '',
             toolAvatar: '',
             params: '',
@@ -237,7 +246,7 @@ export const runToolWithPromptCall = async (
   // 合并工具调用的结果，使用 functionCall 格式存储。
   const assistantToolMsgParams: ChatCompletionAssistantMessageParam = {
     role: ChatCompletionRequestMessageRoleEnum.Assistant,
-    function_call: parseAnswerResult
+    function_call: toolJson
   };
   const concatToolMessages = [
     ...filterMessages,
@@ -248,7 +257,7 @@ export const runToolWithPromptCall = async (
     ...concatToolMessages,
     {
       role: ChatCompletionRequestMessageRoleEnum.Function,
-      name: parseAnswerResult.name,
+      name: toolJson.name,
       content: toolsRunResponse.toolResponsePrompt
     }
   ];
@@ -266,7 +275,7 @@ export const runToolWithPromptCall = async (
     : [toolsRunResponse.moduleRunResponse];
 
   // get the next user prompt
-  lastMessage.content += `${answer}
+  lastMessage.content += `${replaceAnswer}
 TOOL_RESPONSE: """
 ${toolsRunResponse.toolResponsePrompt}
 """
@@ -362,24 +371,37 @@ async function streamResponse({
   return { answer: textAnswer.trim() };
 }
 
-const parseAnswer = (str: string): FunctionCallCompletion | string => {
-  // 首先，使用正则表达式提取TOOL_ID和TOOL_ARGUMENTS
-  const prefix = '1:';
+const parseAnswer = (
+  str: string
+): {
+  answer: string;
+  toolJson?: FunctionCallCompletion;
+} => {
   str = str.trim();
-  if (str.startsWith(prefix)) {
-    const toolString = str.substring(prefix.length).trim();
+  // 首先，使用正则表达式提取TOOL_ID和TOOL_ARGUMENTS
+  const prefixReg = /^1(:|：)/;
+
+  if (prefixReg.test(str)) {
+    const toolString = sliceJsonStr(str);
 
     try {
       const toolCall = json5.parse(toolString);
       return {
-        id: getNanoid(),
-        name: toolCall.toolId,
-        arguments: JSON.stringify(toolCall.arguments || toolCall.parameters)
+        answer: `1: ${toolString}`,
+        toolJson: {
+          id: getNanoid(),
+          name: toolCall.toolId,
+          arguments: JSON.stringify(toolCall.arguments || toolCall.parameters)
+        }
       };
     } catch (error) {
-      return str;
+      return {
+        answer: ERROR_TEXT
+      };
     }
   } else {
-    return str;
+    return {
+      answer: str
+    };
   }
 };