ChatGPT/FastGPT
1
0
Fork 0
mirror of https://github.com/labring/FastGPT.git synced 2025-07-23 13:03:50 +00:00
Code Issues Packages Projects Releases Wiki Activity

4.6.5- CoreferenceResolution Module (#631)

Browse Source
This commit is contained in:
Archer
2023-12-22 10:47:31 +08:00
committed by GitHub
parent 41115a96c0
commit cd682d4275
112 changed files with 4163 additions and 2700 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
docSite/content/docs/development/configuration.md
View File

@@ -92,7 +92,7 @@ weight: 708
"maxContext": 16000,
"maxResponse": 4000,
"price": 0,
"functionCall": true, // 是否支持function call， 不支持的模型需要设置为 false，会走提示词生成
"toolChoice": true, // 是否支持openai的 toolChoice， 不支持的模型需要设置为 false，会走提示词生成
"functionPrompt": ""
},
{
@@ -101,7 +101,7 @@ weight: 708
"maxContext": 8000,
"maxResponse": 8000,
"price": 0,
"functionCall": true,
"toolChoice": true,
"functionPrompt": ""
}
],
@@ -112,7 +112,7 @@ weight: 708
"maxContext": 16000,
"maxResponse": 4000,
"price": 0,
"functionCall": true,
"toolChoice": true,
"functionPrompt": ""
}
],

31
docSite/content/docs/development/upgrading/465.md Normal file
View File

@@ -0,0 +1,31 @@
---
title: 'V4.6.5（需要改配置文件）'
description: 'FastGPT V4.6.5'
icon: 'upgrade'
draft: false
toc: true
weight: 831
---
## 配置文件变更
由于 openai 已开始启用 function call，改为 toolChoice。FastGPT 同步的修改了对于的配置和调用方式，需要对配置文件做一些修改：
[点击查看最新的配置文件](/docs/development/configuration/)
主要是修改模型的`functionCall`字段，改成`toolChoice`即可。设置为`true`的模型，会默认走 openai 的 tools 模式；未设置或设置为`false`的，会走提示词生成模式。
问题补全模型与内容提取模型使用同一组配置。
## V4.6.5 功能介绍
1. 新增 - [问题补全模块](/docs/workflow/modules/coreferenceresolution/)
2. 新增 - [文本编辑模块](/docs/workflow/modules/text_editor/)
3. 新增 - [判断器模块](/docs/workflow/modules/tfswitch/)
4. 新增 - [自定义反馈模块](/docs/workflow/modules/custom_feedback/)
5. 新增 - 【内容提取】模块支持选择模型，以及字段枚举
6. 优化 - docx读取，兼容表格（表格转markdown）
7. 优化 - 高级编排连接线交互
8. 优化 - 由于 html2md 导致的 cpu密集计算，阻断线程问题
9. 修复 - 高级编排提示词提取描述

969
docSite/content/docs/workflow/examples/google_search.md
View File

File diff suppressed because it is too large Load Diff

1268
docSite/content/docs/workflow/examples/lab_appointment.md
View File

File diff suppressed because it is too large Load Diff

512
docSite/content/docs/workflow/examples/op_question.md
View File

@@ -1,512 +0,0 @@
---
title: '优化知识库搜索词'
description: '利用 GPT 优化和完善知识库搜索词，实现上下文关联搜索'
icon: 'search'
draft: false
toc: true
weight: 404
---
![](/imgs/demo_op_question1.png)
| 优化前 | 优化后 |
| --------------------- | --------------------- |
| ![](/imgs/demo_op_question3.png) | ![](/imgs/demo_op_question2.png) |
如上图，优化后的搜索可以针对【自动数据预处理】进行搜索，从而找到其相关的内容，一定程度上弥补了向量搜索的上下文缺失问题。
## 模块编排
复制下面配置，点击「高级编排」右上角的导入按键，导入该配置。
{{% details title="编排配置" closed="true" %}}
```json
[
{
"moduleId": "userChatInput",
"name": "用户问题(对话入口)",
"flowType": "questionInput",
"position": {
"x": 585.750318069507,
"y": 1597.4127130315183
},
"inputs": [
{
"key": "userChatInput",
"type": "systemInput",
"label": "用户问题",
"connected": true
}
],
"outputs": [
{
"key": "userChatInput",
"label": "用户问题",
"type": "source",
"valueType": "string",
"targets": [
{
"moduleId": "ssdd86",
"key": "content"
}
]
}
]
},
{
"moduleId": "history",
"name": "聊天记录",
"flowType": "historyNode",
"position": {
"x": 567.49877916803,
"y": 1289.3453864378014
},
"inputs": [
{
"key": "maxContext",
"type": "numberInput",
"label": "最长记录数",
"value": 6,
"min": 0,
"max": 50,
"connected": true
},
{
"key": "history",
"type": "hidden",
"label": "聊天记录",
"connected": true
}
],
"outputs": [
{
"key": "history",
"label": "聊天记录",
"valueType": "chatHistory",
"type": "source",
"targets": [
{
"moduleId": "ssdd86",
"key": "history"
}
]
}
]
},
{
"moduleId": "nkxlso",
"name": "知识库搜索",
"flowType": "datasetSearchNode",
"showStatus": true,
"position": {
"x": 1542.6434554710224,
"y": 1153.7853815737192
},
"inputs": [
{
"key": "kbList",
"type": "custom",
"label": "关联的知识库",
"value": [],
"list": [],
"connected": true
},
{
"key": "similarity",
"type": "slider",
"label": "相似度",
"value": 0.8,
"min": 0,
"max": 1,
"step": 0.01,
"markList": [
{
"label": "100",
"value": 100
},
{
"label": "1",
"value": 1
}
],
"connected": true
},
{
"key": "limit",
"type": "slider",
"label": "单次搜索上限",
"description": "最多取 n 条记录作为本次问题引用",
"value": 7,
"min": 1,
"max": 20,
"step": 1,
"markList": [
{
"label": "1",
"value": 1
},
{
"label": "20",
"value": 20
}
],
"connected": true
},
{
"key": "switch",
"type": "target",
"label": "触发器",
"valueType": "any",
"connected": false
},
{
"key": "userChatInput",
"type": "target",
"label": "用户问题",
"required": true,
"valueType": "string",
"connected": true
}
],
"outputs": [
{
"key": "isEmpty",
"label": "搜索结果为空",
"type": "source",
"valueType": "boolean",
"targets": []
},
{
"key": "unEmpty",
"label": "搜索结果不为空",
"type": "source",
"valueType": "boolean",
"targets": []
},
{
"key": "quoteQA",
"label": "引用内容",
"description": "始终返回数组，如果希望搜索结果为空时执行额外操作，需要用到上面的两个输入以及目标模块的触发器",
"type": "source",
"valueType": "datasetQuote",
"targets": [
{
"moduleId": "ol82hp",
"key": "quoteQA"
}
]
}
]
},
{
"moduleId": "ol82hp",
"name": "AI 对话",
"flowType": "chatNode",
"showStatus": true,
"position": {
"x": 2207.4577044902126,
"y": 1079.6308003796544
},
"inputs": [
{
"key": "model",
"type": "custom",
"label": "对话模型",
"value": "gpt-3.5-turbo",
"list": [],
"connected": true
},
{
"key": "temperature",
"type": "slider",
"label": "温度",
"value": 0,
"min": 0,
"max": 10,
"step": 1,
"markList": [
{
"label": "严谨",
"value": 0
},
{
"label": "发散",
"value": 10
}
],
"connected": true
},
{
"key": "maxToken",
"type": "custom",
"label": "回复上限",
"value": 2000,
"min": 100,
"max": 4000,
"step": 50,
"markList": [
{
"label": "100",
"value": 100
},
{
"label": "4000",
"value": 4000
}
],
"connected": true
},
{
"key": "systemPrompt",
"type": "textarea",
"label": "系统提示词",
"max": 300,
"valueType": "string",
"description": "模型固定的引导词，通过调整该内容，可以引导模型聊天方向。该内容会被固定在上下文的开头。可使用变量，例如 {{language}}",
"placeholder": "模型固定的引导词，通过调整该内容，可以引导模型聊天方向。该内容会被固定在上下文的开头。可使用变量，例如 {{language}}",
"value": "我会向你询问三引号引用中提及的内容，你仅使用提供的引用内容来回答我的问题，不要做额外的扩展补充。",
"connected": true
},
{
"key": "limitPrompt",
"type": "textarea",
"valueType": "string",
"label": "限定词",
"max": 500,
"description": "限定模型对话范围，会被放置在本次提问前，拥有强引导和限定性。不建议内容太长，会影响上下文，可使用变量，例如 {{language}}。可在文档中找到对应的限定例子",
"placeholder": "限定模型对话范围，会被放置在本次提问前，拥有强引导和限定性。不建议内容太长，会影响上下文，可使用变量，例如 {{language}}。可在文档中找到对应的限定例子",
"value": "",
"connected": true
},
{
"key": "switch",
"type": "target",
"label": "触发器",
"valueType": "any",
"connected": false
},
{
"key": "quoteQA",
"type": "target",
"label": "引用内容",
"description": "对象数组格式，结构：\n [{q:'问题',a:'回答'}]",
"valueType": "datasetQuote",
"connected": true
},
{
"key": "history",
"type": "target",
"label": "聊天记录",
"valueType": "chatHistory",
"connected": true
},
{
"key": "userChatInput",
"type": "target",
"label": "用户问题",
"required": true,
"valueType": "string",
"connected": true
}
],
"outputs": [
{
"key": "answerText",
"label": "AI回复",
"description": "将在 stream 回复完毕后触发",
"valueType": "string",
"type": "source",
"targets": []
},
{
"key": "finish",
"label": "回复结束",
"description": "AI 回复完成后触发",
"valueType": "boolean",
"type": "source",
"targets": []
}
]
},
{
"moduleId": "o62kns",
"name": "用户问题(对话入口)",
"flowType": "questionInput",
"position": {
"x": 1696.5940057372968,
"y": 2270.5070479742435
},
"inputs": [
{
"key": "userChatInput",
"type": "systemInput",
"label": "用户问题",
"connected": true
}
],
"outputs": [
{
"key": "userChatInput",
"label": "用户问题",
"type": "source",
"valueType": "string",
"targets": [
{
"moduleId": "ol82hp",
"key": "userChatInput"
}
]
}
]
},
{
"moduleId": "he7013",
"name": "聊天记录",
"flowType": "historyNode",
"position": {
"x": 1636.793907221069,
"y": 1952.7122387165764
},
"inputs": [
{
"key": "maxContext",
"type": "numberInput",
"label": "最长记录数",
"value": 6,
"min": 0,
"max": 50,
"connected": true
},
{
"key": "history",
"type": "hidden",
"label": "聊天记录",
"connected": true
}
],
"outputs": [
{
"key": "history",
"label": "聊天记录",
"valueType": "chatHistory",
"type": "source",
"targets": [
{
"moduleId": "ol82hp",
"key": "history"
}
]
}
]
},
{
"moduleId": "ssdd86",
"name": "文本内容提取",
"flowType": "contentExtract",
"showStatus": true,
"position": {
"x": 1031.822028231947,
"y": 1231.9793566344022
},
"inputs": [
{
"key": "switch",
"type": "target",
"label": "触发器",
"valueType": "any",
"connected": false
},
{
"key": "description",
"type": "textarea",
"valueType": "string",
"value": "结合上下文，优化用户的问题，要求不能包含\"它\"、\"第几个\"等代名词，需将他们替换成具体的名词。",
"label": "提取要求描述",
"description": "写一段提取要求，告诉 AI 需要提取哪些内容",
"required": true,
"placeholder": "例如: \n1. 你是一个实验室预约助手。根据用户问题，提取出姓名、实验室号和预约时间",
"connected": true
},
{
"key": "history",
"type": "target",
"label": "聊天记录",
"valueType": "chatHistory",
"connected": true
},
{
"key": "content",
"type": "target",
"label": "需要提取的文本",
"required": true,
"valueType": "string",
"connected": true
},
{
"key": "extractKeys",
"type": "custom",
"label": "目标字段",
"description": "由 '描述' 和 'key' 组成一个目标字段，可提取多个目标字段",
"value": [
{
"desc": "优化后的问题",
"key": "q",
"required": true
}
],
"connected": true
}
],
"outputs": [
{
"key": "success",
"label": "字段完全提取",
"valueType": "boolean",
"type": "source",
"targets": []
},
{
"key": "failed",
"label": "提取字段缺失",
"valueType": "boolean",
"type": "source",
"targets": []
},
{
"key": "fields",
"label": "完整提取结果",
"description": "一个 JSON 字符串，例如：{\"name:\":\"YY\",\"Time\":\"2023/7/2 18:00\"}",
"valueType": "string",
"type": "source",
"targets": []
},
{
"key": "q",
"label": "提取结果-优化后的问题",
"description": "无法提取时不会返回",
"valueType": "string",
"type": "source",
"targets": [
{
"moduleId": "nkxlso",
"key": "userChatInput"
}
]
}
]
}
]
```
{{% /details %}}
## 流程说明
1. 利用内容提取模块，将用户的问题进行优化。
2. 将优化后的问题传递到知识库搜索模块进行搜索。
3. 搜索内容传递到 AI 对话模块，进行回答。
## Tips
内容提取模块可以将自然语言提取成结构化数据，可以使用其进行一些神奇的操作。

4
docSite/content/docs/workflow/modules/_index.md
View File

@@ -5,4 +5,6 @@ description: "介绍 FastGPT 的常用模块"
icon: "apps"
draft: false
images: []
---
---
<!-- 350 ~ 400 -->

39
docSite/content/docs/workflow/modules/coreferenceResolution.md Normal file
View File

@@ -0,0 +1,39 @@
---
title: "问题补全"
description: "问题补全模块介绍和使用"
icon: "input"
draft: false
toc: true
weight: 364
---
## 特点
- 可重复添加
- 有外部输入
- 触发执行
![](/imgs/coreferenceResolution1.png)
## 背景
在 RAG 中，我们需要根据输入的问题去数据库里执行 embedding 搜索，查找相关的内容，从而查找到相似的内容（简称知识库搜索）。
在搜索的过程中，尤其是连续对话的搜索，我们通常会发现后续的问题难以搜索到合适的内容，其中一个原因是知识库搜索只会使用“当前”的问题去执行。看下面的例子：
![](/imgs/coreferenceResolution2.png)
用户在提问“第二点是什么”的时候，只会去知识库里查找“第二点是什么”，压根查不到内容。实际上需要查询的是“QA结构是什么”。因此我们需要引入一个【问题补全】模块，来对用户当前的问题进行补全，从而使得知识库搜索能够搜索到合适的内容。使用补全后效果如下：
![](/imgs/coreferenceResolution3.png)
## 功能
调用 AI 去对用户当前的问题进行补全。目前主要是补全“指代”词，使得检索词更加的完善可靠，从而增强上下文连续对话的知识库搜索能力。
遇到最大的难题在于：模型对于【补全】的概念可能不清晰，且对于长上下文往往无法准确的知道应该如何补全。
## 示例
- [接入谷歌搜索](/docs/workflow/examples/google_search/)

0
docSite/content/docs/workflow/modules/kb_search.md → docSite/content/docs/workflow/modules/dataset_search.md
View File

0
docSite/content/docs/workflow/modules/question.md → docSite/content/docs/workflow/modules/question_classify.md
View File

4
docSite/content/docs/workflow/modules/string.md → docSite/content/docs/workflow/modules/text_editor.md
View File

@@ -26,3 +26,7 @@ weight: 363
## 作用
给任意模块输入自定格式文本，或处理 AI 模块系统提示词。
## 示例
- [接入谷歌搜索](/docs/workflow/examples/google_search/)

3
docSite/content/docs/workflow/modules/judgement.md → docSite/content/docs/workflow/modules/tfswitch.md
View File

@@ -25,4 +25,5 @@ weight: 362
## 作用
适用场景有：让大模型做判断后输出固定内容，根据大模型回复内容判断是否触发后续模块。
适用场景有：让大模型做判断后输出固定内容，根据大模型回复内容判断是否触发后续模块。