mirror of
https://github.com/labring/FastGPT.git
synced 2025-07-21 11:43:56 +00:00
42b2046f96
* agent search demo * edit form force close image select * feat: llm params and doubao1.5 * perf: model error tip * fix: template register path * package
278 lines
7.5 KiB
TypeScript
278 lines
7.5 KiB
TypeScript
import { chats2GPTMessages } from '@fastgpt/global/core/chat/adapt';
|
||
import { ChatItemType } from '@fastgpt/global/core/chat/type';
|
||
import { DatasetSearchModeEnum } from '@fastgpt/global/core/dataset/constants';
|
||
import { getLLMModel } from '../../ai/model';
|
||
import { filterGPTMessageByMaxContext } from '../../chat/utils';
|
||
import { replaceVariable } from '@fastgpt/global/common/string/tools';
|
||
import { createChatCompletion } from '../../ai/config';
|
||
import { llmCompletionsBodyFormat } from '../../ai/utils';
|
||
import { ChatCompletionMessageParam } from '@fastgpt/global/core/ai/type';
|
||
import { SearchDataResponseItemType } from '@fastgpt/global/core/dataset/type';
|
||
import { searchDatasetData } from './controller';
|
||
|
||
type SearchDatasetDataProps = {
|
||
queries: string[];
|
||
histories: ChatItemType[];
|
||
teamId: string;
|
||
model: string;
|
||
similarity?: number; // min distance
|
||
limit: number; // max Token limit
|
||
datasetIds: string[];
|
||
searchMode?: `${DatasetSearchModeEnum}`;
|
||
usingReRank?: boolean;
|
||
reRankQuery: string;
|
||
|
||
/*
|
||
{
|
||
tags: {
|
||
$and: ["str1","str2"],
|
||
$or: ["str1","str2",null] null means no tags
|
||
},
|
||
createTime: {
|
||
$gte: 'xx',
|
||
$lte: 'xxx'
|
||
}
|
||
}
|
||
*/
|
||
collectionFilterMatch?: string;
|
||
};
|
||
|
||
const analyzeQuery = async ({ query, histories }: { query: string; histories: ChatItemType[] }) => {
|
||
const modelData = getLLMModel('gpt-4o-mini');
|
||
|
||
const systemFewShot = `
|
||
## 知识背景
|
||
FastGPT 是低代码AI应用构建平台，支持通过语义相似度实现精准数据检索。用户正在利用该功能开发数据检索应用。
|
||
|
||
## 任务目标
|
||
基于用户历史对话和知识背景，生成多维度检索方案，确保覆盖核心语义及潜在关联维度。
|
||
|
||
## 工作流程
|
||
1. 问题解构阶段
|
||
[意图识别] 提取用户问题的核心实体和关系：
|
||
- 显性需求：直接提及的关键词
|
||
- 隐性需求：可能涉及的关联概念
|
||
[示例] 若问题为"推荐手机"，需考虑价格、品牌、使用场景等维度
|
||
|
||
2. 完整性校验阶段
|
||
[完整性评估] 检查是否缺失核心实体和关系：
|
||
- 主语完整
|
||
- 多实体关系准确
|
||
[维度扩展] 检查是否需要补充：
|
||
□ 时间范围 □ 地理限定 □ 比较维度
|
||
□ 专业术语 □ 同义词替换 □ 场景参数
|
||
|
||
3. 检索生成阶段
|
||
[组合策略] 生成包含以下要素的查询序列：
|
||
① 基础查询（核心关键词）
|
||
② 扩展查询（核心+同义词）
|
||
③ 场景查询（核心+场景限定词）
|
||
④ 逆向查询（相关技术/对比对象）
|
||
|
||
## 输出规范
|
||
格式要求：
|
||
1. 每个查询为完整陈述句
|
||
2. 包含至少1个核心词+1个扩展维度
|
||
3. 按查询范围从宽到窄排序
|
||
|
||
禁止项：
|
||
- 使用问句形式
|
||
- 包含解决方案描述
|
||
- 超出话题范围的假设
|
||
|
||
## 执行示例
|
||
用户问题："如何优化数据检索速度"
|
||
|
||
查询内容：
|
||
1. FastGPT 数据检索速度优化的常用方法
|
||
2. FastGPT 大数据量下的语义检索性能提升方案
|
||
3. FastGPT API 响应时间的优化指标
|
||
|
||
## 任务开始
|
||
`.trim();
|
||
const filterHistories = await filterGPTMessageByMaxContext({
|
||
messages: chats2GPTMessages({ messages: histories, reserveId: false }),
|
||
maxContext: modelData.maxContext - 1000
|
||
});
|
||
|
||
const messages = [
|
||
{
|
||
role: 'system',
|
||
content: systemFewShot
|
||
},
|
||
...filterHistories,
|
||
{
|
||
role: 'user',
|
||
content: query
|
||
}
|
||
] as any;
|
||
|
||
const { response: result } = await createChatCompletion({
|
||
body: llmCompletionsBodyFormat(
|
||
{
|
||
stream: false,
|
||
model: modelData.model,
|
||
temperature: 0.1,
|
||
messages
|
||
},
|
||
modelData
|
||
)
|
||
});
|
||
let answer = result.choices?.[0]?.message?.content || '';
|
||
|
||
// Extract queries from the answer by line number
|
||
const queries = answer
|
||
.split('\n')
|
||
.map((line) => {
|
||
const match = line.match(/^\d+\.\s*(.+)$/);
|
||
return match ? match[1].trim() : null;
|
||
})
|
||
.filter(Boolean) as string[];
|
||
|
||
if (queries.length === 0) {
|
||
return [answer];
|
||
}
|
||
|
||
return queries;
|
||
};
|
||
const checkQuery = async ({
|
||
queries,
|
||
histories,
|
||
searchResult
|
||
}: {
|
||
queries: string[];
|
||
histories: ChatItemType[];
|
||
searchResult: SearchDataResponseItemType[];
|
||
}) => {
|
||
const modelData = getLLMModel('gpt-4o-mini');
|
||
|
||
const systemFewShot = `
|
||
## 知识背景
|
||
FastGPT 是低代码AI应用构建平台，支持通过语义相似度实现精准数据检索。用户正在利用该功能开发数据检索应用。
|
||
|
||
## 查询结果
|
||
${searchResult.map((item) => item.q + item.a).join('---\n---')}
|
||
|
||
## 任务目标
|
||
检查"检索结果"是否覆盖用户的问题，如果无法覆盖用户问题，则再次生成检索方案。
|
||
|
||
## 工作流程
|
||
1. 检查检索结果是否覆盖用户的问题
|
||
2. 如果检索结果覆盖用户问题，则直接输出："Done"
|
||
3. 如果无法覆盖用户问题，则结合用户问题和检索结果，生成进一步的检索方案，进行深度检索
|
||
|
||
## 输出规范
|
||
|
||
1. 每个查询均为完整的查询语句
|
||
2. 通过序号来表示多个检索内容
|
||
|
||
## 输出示例1
|
||
Done
|
||
|
||
## 输出示例2
|
||
1. 环界云计算的办公地址
|
||
2. 环界云计算的注册地址在哪里
|
||
|
||
## 任务开始
|
||
`.trim();
|
||
const filterHistories = await filterGPTMessageByMaxContext({
|
||
messages: chats2GPTMessages({ messages: histories, reserveId: false }),
|
||
maxContext: modelData.maxContext - 1000
|
||
});
|
||
|
||
const messages = [
|
||
{
|
||
role: 'system',
|
||
content: systemFewShot
|
||
},
|
||
...filterHistories,
|
||
{
|
||
role: 'user',
|
||
content: queries.join('\n')
|
||
}
|
||
] as any;
|
||
console.log(messages);
|
||
const { response: result } = await createChatCompletion({
|
||
body: llmCompletionsBodyFormat(
|
||
{
|
||
stream: false,
|
||
model: modelData.model,
|
||
temperature: 0.1,
|
||
messages
|
||
},
|
||
modelData
|
||
)
|
||
});
|
||
let answer = result.choices?.[0]?.message?.content || '';
|
||
console.log(answer);
|
||
if (answer.includes('Done')) {
|
||
return [];
|
||
}
|
||
|
||
const nextQueries = answer
|
||
.split('\n')
|
||
.map((line) => {
|
||
const match = line.match(/^\d+\.\s*(.+)$/);
|
||
return match ? match[1].trim() : null;
|
||
})
|
||
.filter(Boolean) as string[];
|
||
|
||
return nextQueries;
|
||
};
|
||
export const agentSearchDatasetData = async ({
|
||
searchRes = [],
|
||
tokens = 0,
|
||
...props
|
||
}: SearchDatasetDataProps & {
|
||
searchRes?: SearchDataResponseItemType[];
|
||
tokens?: number;
|
||
}) => {
|
||
const query = props.queries[0];
|
||
|
||
const searchResultList: SearchDataResponseItemType[] = [];
|
||
let searchQueries: string[] = [];
|
||
|
||
// 1. agent 分析问题
|
||
searchQueries = await analyzeQuery({ query, histories: props.histories });
|
||
|
||
// 2. 检索内容 + 检查
|
||
let retryTimes = 3;
|
||
while (true) {
|
||
retryTimes--;
|
||
if (retryTimes < 0) break;
|
||
|
||
console.log(searchQueries, '--');
|
||
const { searchRes: searchRes2, tokens: tokens2 } = await searchDatasetData({
|
||
...props,
|
||
queries: searchQueries
|
||
});
|
||
// console.log(searchRes2.map((item) => item.q));
|
||
// deduplicate and merge search results
|
||
const uniqueResults = searchRes2.filter((item) => {
|
||
return !searchResultList.some((existingItem) => existingItem.id === item.id);
|
||
});
|
||
searchResultList.push(...uniqueResults);
|
||
if (uniqueResults.length === 0) break;
|
||
|
||
const checkResult = await checkQuery({
|
||
queries: searchQueries,
|
||
histories: props.histories,
|
||
searchResult: searchRes2
|
||
});
|
||
|
||
if (checkResult.length > 0) {
|
||
searchQueries = checkResult;
|
||
} else {
|
||
break;
|
||
}
|
||
}
|
||
|
||
console.log(searchResultList.length);
|
||
return {
|
||
searchRes: searchResultList,
|
||
tokens: 0,
|
||
usingSimilarityFilter: false,
|
||
usingReRank: false
|
||
};
|
||
};
|