Use JSON extraction for content extract node (#7117)

38b26a3b · YeYuheng · GitHub · 7d5b6a5c · 38b26a3b · 38b26a3b
Commit 38b26a3b authored Jun 15, 2026 by YeYuheng Committed by GitHub Jun 15, 2026
4 changed files
--- a/packages/global/core/ai/prompt/agent.ts
+++ b/packages/global/core/ai/prompt/agent.ts
@@ -41,13 +41,16 @@ export const getExtractJsonPrompt = ({
    memory ? '【历史提取结果】' : ''
  ].filter(Boolean);
  const prompt = `## 背景
-用户需要执行一个函数，该函数需要一些参数，需要你结合${list.join('、')}，来生成对应的参数
+用户需要从文本中提取结构化信息，需要你结合${list.join('、')}，按 JSON Schema 生成对应字段值
 ## 基本要求
- 严格根据 JSON Schema 的描述来生成参数。
+- 严格根据 JSON Schema 的字段名、字段描述、字段类型和枚举值提取字段值。
- 不是每个参数都是必须生成的，如果没有合适的参数值，不要生成该参数，或返回空字符串。
+- 输出字段只能来自 JSON Schema，不要新增字段，不要修改字段名。
- 需要结合历史记录，一起生成合适的参数。
+- JSON Schema required 中声明的字段必须生成；如果没有可靠字段值，返回空字符串。
+- 非 required 字段如果没有可靠字段值，可以省略该字段或返回空字符串。
+- 不要编造用户输入、历史记录、背景知识和历史提取结果中无法支持的信息。
+- 需要结合历史记录、用户输入、背景知识和历史提取结果提取合适的字段值，最新用户输入优先。
 ${
  systemPrompt
@@ -69,8 +72,9 @@ ${schema}
 ## 输出要求
- 严格输出 json 字符串。
+- 严格只输出一个 JSON object 字符串。
- 不要回答问题。`.replace(/\n{3,}/g, '\n\n');
+- 不要输出 Markdown、代码块、解释、前后缀文本或回答问题。
+- JSON 必须可以被标准 JSON/JSON5 解析。`.replace(/\n{3,}/g, '\n\n');
  return prompt;
 };

--- a/packages/global/test/core/ai/prompt/agent.test.ts
+++ b/packages/global/test/core/ai/prompt/agent.test.ts
@@ -44,6 +44,14 @@ describe('getExtractJsonPrompt', () => {
    expect(result).toContain('## JSON Schema');
  });
+  it('should describe text extraction instead of function calling', () => {
+    const result = getExtractJsonPrompt({});
+    expect(result).toContain('从文本中提取结构化信息');
+    expect(result).toContain('按 JSON Schema 生成对应字段值');
+    expect(result).not.toContain('执行一个函数');
+    expect(result).not.toContain('生成对应的参数');
+  });
  it('should include systemPrompt section when systemPrompt is provided', () => {
    const result = getExtractJsonPrompt({ systemPrompt: '请提取用户姓名' });
    expect(result).toContain('【背景知识】');
@@ -95,8 +103,23 @@ describe('getExtractJsonPrompt', () => {
  it('should always contain output requirements section', () => {
    const result = getExtractJsonPrompt({});
-    expect(result).toContain('严格输出 json 字符串');
+    expect(result).toContain('严格只输出一个 JSON object 字符串');
-    expect(result).toContain('不要回答问题');
+    expect(result).toContain('不要输出 Markdown、代码块、解释、前后缀文本或回答问题');
+  });
+  it('should constrain plain JSON extraction format and fields', () => {
+    const result = getExtractJsonPrompt({
+      schema: '{"type":"object","properties":{"name":{"type":"string"}}}'
+    });
+    expect(result).toContain('输出字段只能来自 JSON Schema');
+    expect(result).toContain('不要新增字段');
+    expect(result).toContain('不要修改字段名');
+    expect(result).toContain('JSON Schema required 中声明的字段必须生成');
+    expect(result).toContain('非 required 字段如果没有可靠字段值，可以省略');
+    expect(result).toContain('字段类型和枚举值');
+    expect(result).toContain('不要编造');
+    expect(result).toContain('最新用户输入优先');
  });
 });

--- a/packages/service/core/workflow/dispatch/ai/extract.ts
+++ b/packages/service/core/workflow/dispatch/ai/extract.ts
 import { chats2GPTMessages } from '@fastgpt/global/core/chat/adapt';
-import { filterGPTMessageByMaxContext } from '../../../ai/llm/utils';
 import type { ChatItemMiniType } from '@fastgpt/global/core/chat/type';
 import { ChatRoleEnum } from '@fastgpt/global/core/chat/constants';
 import type { ContextExtractAgentItemType } from '@fastgpt/global/core/workflow/template/system/contextExtract/type';
@@ -20,12 +19,8 @@ import json5 from 'json5';
 import { getLogger, LogCategories } from '../../../../common/logger';
 const logger = getLogger(LogCategories.MODULE.WORKFLOW.AI);
-import { type ChatCompletionTool } from '@fastgpt/global/core/ai/llm/type';
 import { type DispatchNodeResultType } from '@fastgpt/global/core/workflow/runtime/type';
-import {
+import { getExtractJsonPrompt } from '@fastgpt/global/core/ai/prompt/agent';
-  getExtractJsonPrompt,
-  getExtractJsonToolPrompt
-} from '@fastgpt/global/core/ai/prompt/agent';
 import { createLLMResponse } from '../../../ai/llm/request';
 import type { JsonSchemaPropertiesItemType } from '@fastgpt/global/core/app/jsonschema';
@@ -44,8 +39,6 @@ type Response = DispatchNodeResultType<{
 type ActionProps = Props & { extractModel: LLMModelItemType; lastMemory?: Record<string, any> };
-const agentFunName = 'request_function';
 export async function dispatchContentExtract(props: Props): Promise<Response> {
  const {
    runningAppInfo,
@@ -69,22 +62,12 @@ export async function dispatchContentExtract(props: Props): Promise<Response> {
  >;
  try {
-    const { arg, inputTokens, outputTokens, usedUserOpenAIKey } = await (async () => {
+    const { arg, inputTokens, outputTokens, usedUserOpenAIKey } = await completions({
-      if (extractModel.toolChoice) {
-        return toolChoice({
      ...props,
      histories: chatHistories,
      extractModel,
      lastMemory
    });
-      }
-      return completions({
-        ...props,
-        histories: chatHistories,
-        extractModel,
-        lastMemory
-      });
-    })();
    // remove invalid key
    for (const key in arg) {
@@ -174,109 +157,6 @@ const getJsonSchema = ({ params: { extractKeys } }: ActionProps) => {
  return properties;
 };
-const toolChoice = async (props: ActionProps) => {
-  const {
-    externalProvider,
-    extractModel,
-    histories,
-    params: { content, description },
-    lastMemory
-  } = props;
-  const messages: ChatItemMiniType[] = [
-    {
-      obj: ChatRoleEnum.System,
-      value: [
-        {
-          text: {
-            content: getExtractJsonToolPrompt({
-              systemPrompt: description,
-              memory: lastMemory ? JSON.stringify(lastMemory) : undefined
-            })
-          }
-        }
-      ]
-    },
-    ...histories,
-    {
-      obj: ChatRoleEnum.Human,
-      value: [
-        {
-          text: {
-            content
-          }
-        }
-      ]
-    }
-  ];
-  const adaptMessages = chats2GPTMessages({
-    messages,
-    reserveId: false,
-    reserveReason: false
-  });
-  const filterMessages = await filterGPTMessageByMaxContext({
-    messages: adaptMessages,
-    maxContext: extractModel.maxContext
-  });
-  const schema = getJsonSchema(props);
-  const tools: ChatCompletionTool[] = [
-    {
-      type: 'function',
-      function: {
-        name: agentFunName,
-        description: '需要执行的函数',
-        parameters: {
-          type: 'object',
-          properties: schema,
-          required: []
-        }
-      }
-    }
-  ];
-  const body = {
-    stream: true,
-    model: extractModel.model,
-    messages: filterMessages,
-    tools,
-    tool_choice: { type: 'function', function: { name: agentFunName } },
-    toolCallMode: 'toolChoice',
-    ...(extractModel.reasoning ? { reasoning_effort: 'none' as const } : {})
-  } as const;
-  const {
-    answerText: text,
-    toolCalls,
-    usage: { inputTokens, outputTokens, usedUserOpenAIKey }
-  } = await createLLMResponse({
-    body,
-    userKey: externalProvider.openaiAccount
-  });
-  const arg: Record<string, any> = (() => {
-    try {
-      return json5.parse(toolCalls?.[0]?.function?.arguments || text || '');
-    } catch (error) {
-      logger.warn('Failed to parse tool call arguments', {
-        body,
-        responseText: text,
-        toolCall: toolCalls?.[0]?.function,
-        error
-      });
-      return {};
-    }
-  })();
-  return {
-    inputTokens,
-    outputTokens,
-    usedUserOpenAIKey,
-    arg
-  };
-};
 const completions = async (props: ActionProps) => {
  const {
    extractModel,
@@ -315,6 +195,8 @@ const completions = async (props: ActionProps) => {
  ];
  const {
+    requestId,
+    finish_reason: finishReason,
    answerText: answer,
    usage: { inputTokens, outputTokens, usedUserOpenAIKey }
  } = await createLLMResponse({
@@ -329,7 +211,26 @@ const completions = async (props: ActionProps) => {
  // parse response
  const jsonStr = sliceJsonStr(answer);
+  logger.debug('Content extract LLM response received', {
+    requestId,
+    model: extractModel.model,
+    finishReason,
+    answerLength: answer.length,
+    jsonLength: jsonStr.length,
+    inputTokens,
+    outputTokens
+  });
  if (!jsonStr) {
+    logger.warn('Content extract result has no JSON content', {
+      requestId,
+      model: extractModel.model,
+      finishReason,
+      answerLength: answer.length,
+      inputTokens,
+      outputTokens
+    });
    return {
      rawResponse: answer,
      inputTokens,
@@ -348,7 +249,14 @@ const completions = async (props: ActionProps) => {
      arg: json5.parse(jsonStr) as Record<string, any>
    };
  } catch (error) {
-    logger.warn('Failed to parse extract result', { answer, error });
+    logger.warn('Failed to parse extract result', {
+      requestId,
+      model: extractModel.model,
+      finishReason,
+      answerLength: answer.length,
+      jsonLength: jsonStr.length,
+      error
+    });
    return {
      rawResponse: answer,
      inputTokens,

--- a/packages/service/test/core/workflow/dispatch/ai/extract.test.ts
+++ b/packages/service/test/core/workflow/dispatch/ai/extract.test.ts
 import { beforeEach, describe, expect, it, vi } from 'vitest';
 import { ChatRoleEnum } from '@fastgpt/global/core/chat/constants';
-import { NodeInputKeyEnum } from '@fastgpt/global/core/workflow/constants';
+import { NodeInputKeyEnum, NodeOutputKeyEnum } from '@fastgpt/global/core/workflow/constants';
 import { FlowNodeTypeEnum } from '@fastgpt/global/core/workflow/node/constant';
-const {
+const { createLLMResponseMock, getLLMModelMock, formatModelChars2PointsMock } = vi.hoisted(() => ({
-  createLLMResponseMock,
-  filterGPTMessageByMaxContextMock,
-  getLLMModelMock,
-  formatModelChars2PointsMock
-} = vi.hoisted(() => ({
  createLLMResponseMock: vi.fn(),
-  filterGPTMessageByMaxContextMock: vi.fn(),
  getLLMModelMock: vi.fn(),
  formatModelChars2PointsMock: vi.fn()
 }));
@@ -19,10 +13,6 @@ vi.mock('@fastgpt/service/core/ai/llm/request', () => ({
  createLLMResponse: createLLMResponseMock
 }));
-vi.mock('@fastgpt/service/core/ai/llm/utils', () => ({
-  filterGPTMessageByMaxContext: filterGPTMessageByMaxContextMock
-}));
 vi.mock('@fastgpt/service/core/ai/model', () => ({
  getLLMModel: getLLMModelMock
 }));
@@ -72,17 +62,59 @@ const createProps = () =>
    }
  }) as any;
+const createAfterSaleProps = () =>
+  ({
+    ...createProps(),
+    params: {
+      ...createProps().params,
+      content: '客户李四反馈商品包装破损，但没有提供订单号和联系电话。他希望平台尽快处理。',
+      description: '从售后反馈中提取客户、订单、商品和问题信息',
+      extractKeys: [
+        {
+          key: 'customerName',
+          desc: '客户姓名',
+          required: true
+        },
+        {
+          key: 'orderNo',
+          desc: '订单号',
+          required: true
+        },
+        {
+          key: 'productName',
+          desc: '商品名称',
+          required: true
+        },
+        {
+          key: 'issue',
+          desc: '客户反馈的问题',
+          required: true
+        }
+      ]
+    }
+  }) as any;
+const mockLLMResponse = (answerText: string) => ({
+  requestId: 'request_1',
+  finish_reason: 'stop',
+  answerText,
+  usage: {
+    inputTokens: 10,
+    outputTokens: 5,
+    usedUserOpenAIKey: false
+  }
+});
 describe('dispatchContentExtract', () => {
  beforeEach(() => {
    vi.clearAllMocks();
-    filterGPTMessageByMaxContextMock.mockImplementation(async ({ messages }) => messages);
    formatModelChars2PointsMock.mockReturnValue({
      totalPoints: 1,
      modelName: 'DeepSeek R1'
    });
  });
-  it('forces reasoning models to disable reasoning in tool choice extraction', async () => {
+  it('uses plain JSON extraction even when the model supports tool choice', async () => {
    getLLMModelMock.mockReturnValue({
      model: 'deepseek-r1',
      name: 'DeepSeek R1',
@@ -90,28 +122,21 @@ describe('dispatchContentExtract', () => {
      reasoning: true,
      toolChoice: true
    });
-    createLLMResponseMock.mockResolvedValue({
+    createLLMResponseMock.mockResolvedValue(mockLLMResponse('{"name":"张三"}'));
-      answerText: '',
-      toolCalls: [
-        {
-          function: {
-            arguments: '{"name":"张三"}'
-          }
-        }
-      ],
-      usage: {
-        inputTokens: 10,
-        outputTokens: 5,
-        usedUserOpenAIKey: false
-      }
-    });
-    await dispatchContentExtract(createProps());
+    const result = await dispatchContentExtract(createProps());
    expect(createLLMResponseMock.mock.calls[0][0].body).toMatchObject({
      model: 'deepseek-r1',
-      reasoning_effort: 'none',
+      stream: true
-      toolCallMode: 'toolChoice'
+    });
+    expect(createLLMResponseMock.mock.calls[0][0].body).not.toHaveProperty('tools');
+    expect(createLLMResponseMock.mock.calls[0][0].body).not.toHaveProperty('tool_choice');
+    expect(createLLMResponseMock.mock.calls[0][0].body).not.toHaveProperty('toolCallMode');
+    expect(createLLMResponseMock.mock.calls[0][0].body).not.toHaveProperty('reasoning_effort');
+    expect(result.data).toMatchObject({
+      success: true,
+      name: '张三'
    });
  });
@@ -123,14 +148,7 @@ describe('dispatchContentExtract', () => {
      reasoning: true,
      toolChoice: false
    });
-    createLLMResponseMock.mockResolvedValue({
+    createLLMResponseMock.mockResolvedValue(mockLLMResponse('{"name":"张三"}'));
-      answerText: '{"name":"张三"}',
-      usage: {
-        inputTokens: 10,
-        outputTokens: 5,
-        usedUserOpenAIKey: false
-      }
-    });
    await dispatchContentExtract(createProps());
@@ -148,17 +166,86 @@ describe('dispatchContentExtract', () => {
      reasoning: false,
      toolChoice: false
    });
-    createLLMResponseMock.mockResolvedValue({
+    createLLMResponseMock.mockResolvedValue(mockLLMResponse('{"name":"张三"}'));
-      answerText: '{"name":"张三"}',
-      usage: {
-        inputTokens: 10,
-        outputTokens: 5,
-        usedUserOpenAIKey: false
-      }
-    });
    await dispatchContentExtract(createProps());
    expect(createLLMResponseMock.mock.calls[0][0].body).not.toHaveProperty('reasoning_effort');
  });
+  it('keeps the existing plain JSON slicing behavior', async () => {
+    getLLMModelMock.mockReturnValue({
+      model: 'gpt-4o',
+      name: 'GPT-4o',
+      maxContext: 128000,
+      reasoning: false,
+      toolChoice: false
+    });
+    createLLMResponseMock.mockResolvedValue(mockLLMResponse('提取结果：{"name":"张三"}'));
+    const result = await dispatchContentExtract(createProps());
+    expect(result.data).toMatchObject({
+      success: true,
+      name: '张三'
+    });
+  });
+  it('returns default required fields when the model response has no JSON content', async () => {
+    getLLMModelMock.mockReturnValue({
+      model: 'gpt-4o',
+      name: 'GPT-4o',
+      maxContext: 128000,
+      reasoning: false,
+      toolChoice: false
+    });
+    createLLMResponseMock.mockResolvedValue(mockLLMResponse('没有可提取的信息'));
+    const result = await dispatchContentExtract(createProps());
+    expect(result.data).toMatchObject({
+      success: true,
+      name: ''
+    });
+  });
+  it('extracts multiple fields and removes fields outside the schema', async () => {
+    getLLMModelMock.mockReturnValue({
+      model: 'gpt-4o',
+      name: 'GPT-4o',
+      maxContext: 128000,
+      reasoning: false,
+      toolChoice: false
+    });
+    createLLMResponseMock.mockResolvedValue(
+      mockLLMResponse(
+        JSON.stringify({
+          customerName: '李四',
+          orderNo: '',
+          productName: '',
+          issue: '商品包装破损，希望平台尽快处理',
+          phone: '13800138000'
+        })
+      )
+    );
+    const result = await dispatchContentExtract(createAfterSaleProps());
+    expect(result.data).toMatchObject({
+      success: true,
+      customerName: '李四',
+      orderNo: '',
+      productName: '',
+      issue: '商品包装破损，希望平台尽快处理'
+    });
+    expect(result.data).not.toHaveProperty('phone');
+    expect(result.data?.[NodeOutputKeyEnum.contextExtractFields]).toBe(
+      JSON.stringify({
+        customerName: '李四',
+        issue: '商品包装破损，希望平台尽快处理',
+        orderNo: '',
+        productName: ''
+      })
+    );
+  });
 });