Skip to content

赵月辉 / fastgpt-migrated

Go to a project
Commit a9525398 by Archer Committed by GitHub
Options

perf: rrf code (#5558)

parent e4756c76
Showing with 503 additions and 19 deletions
.gitignore
...@@ -38,5 +38,3 @@ files/helm/fastgpt/charts/*.tgz ...@@ -38,5 +38,3 @@ files/helm/fastgpt/charts/*.tgz
tmp/ tmp/
coverage coverage
document/.source document/.source
bun.lock
\ No newline at end of file
document/content/docs/toc.mdx
...@@ -101,6 +101,7 @@ description: FastGPT 文档目录 ...@@ -101,6 +101,7 @@ description: FastGPT 文档目录
- [/docs/upgrading/4-12/4120](/docs/upgrading/4-12/4120) - [/docs/upgrading/4-12/4120](/docs/upgrading/4-12/4120)
- [/docs/upgrading/4-12/4121](/docs/upgrading/4-12/4121) - [/docs/upgrading/4-12/4121](/docs/upgrading/4-12/4121)
- [/docs/upgrading/4-12/4122](/docs/upgrading/4-12/4122) - [/docs/upgrading/4-12/4122](/docs/upgrading/4-12/4122)
- [/docs/upgrading/4-12/4123](/docs/upgrading/4-12/4123)
- [/docs/upgrading/4-8/40](/docs/upgrading/4-8/40) - [/docs/upgrading/4-8/40](/docs/upgrading/4-8/40)
- [/docs/upgrading/4-8/41](/docs/upgrading/4-8/41) - [/docs/upgrading/4-8/41](/docs/upgrading/4-8/41)
- [/docs/upgrading/4-8/42](/docs/upgrading/4-8/42) - [/docs/upgrading/4-8/42](/docs/upgrading/4-8/42)
......
document/content/docs/upgrading/4-12/4123.mdx 0 → 100644
---
title: 'V4.12.3(进行中)'
description: 'FastGPT V4.12.3 更新说明'
---
## 🚀 新增内容
## ⚙️ 优化
1. 纠正 RRF 权重合并算法，使用标准 RRF 权重公式。
## 🐛 修复
## 🔨 工具更新
document/content/docs/upgrading/4-12/meta.json
{ {
"title": "4.12.x", "title": "4.12.x",
"description": "", "description": "",
"pages": ["4122", "4121", "4120"] "pages": ["4123", "4122", "4121", "4120"]
} }
document/data/doc-last-modified.json
...@@ -41,7 +41,7 @@ ...@@ -41,7 +41,7 @@
"document/content/docs/introduction/guide/DialogBoxes/htmlRendering.mdx": "2025-07-23T21:35:03+08:00", "document/content/docs/introduction/guide/DialogBoxes/htmlRendering.mdx": "2025-07-23T21:35:03+08:00",
"document/content/docs/introduction/guide/DialogBoxes/quoteList.mdx": "2025-07-23T21:35:03+08:00", "document/content/docs/introduction/guide/DialogBoxes/quoteList.mdx": "2025-07-23T21:35:03+08:00",
"document/content/docs/introduction/guide/admin/sso.mdx": "2025-07-24T13:00:27+08:00", "document/content/docs/introduction/guide/admin/sso.mdx": "2025-07-24T13:00:27+08:00",
"document/content/docs/introduction/guide/admin/teamMode.mdx": "2025-07-24T13:00:27+08:00", "document/content/docs/introduction/guide/admin/teamMode.mdx": "2025-08-27T16:59:57+08:00",
"document/content/docs/introduction/guide/course/ai_settings.mdx": "2025-07-24T13:00:27+08:00", "document/content/docs/introduction/guide/course/ai_settings.mdx": "2025-07-24T13:00:27+08:00",
"document/content/docs/introduction/guide/course/chat_input_guide.mdx": "2025-07-23T21:35:03+08:00", "document/content/docs/introduction/guide/course/chat_input_guide.mdx": "2025-07-23T21:35:03+08:00",
"document/content/docs/introduction/guide/course/fileInput.mdx": "2025-07-23T21:35:03+08:00", "document/content/docs/introduction/guide/course/fileInput.mdx": "2025-07-23T21:35:03+08:00",
...@@ -104,7 +104,7 @@ ...@@ -104,7 +104,7 @@
"document/content/docs/upgrading/4-11/4111.mdx": "2025-08-07T22:49:09+08:00", "document/content/docs/upgrading/4-11/4111.mdx": "2025-08-07T22:49:09+08:00",
"document/content/docs/upgrading/4-12/4120.mdx": "2025-08-12T22:45:19+08:00", "document/content/docs/upgrading/4-12/4120.mdx": "2025-08-12T22:45:19+08:00",
"document/content/docs/upgrading/4-12/4121.mdx": "2025-08-15T22:53:06+08:00", "document/content/docs/upgrading/4-12/4121.mdx": "2025-08-15T22:53:06+08:00",
"document/content/docs/upgrading/4-12/4122.mdx": "2025-08-26T23:51:54+08:00", "document/content/docs/upgrading/4-12/4122.mdx": "2025-08-27T00:31:33+08:00",
"document/content/docs/upgrading/4-8/40.mdx": "2025-08-02T19:38:37+08:00", "document/content/docs/upgrading/4-8/40.mdx": "2025-08-02T19:38:37+08:00",
"document/content/docs/upgrading/4-8/41.mdx": "2025-08-02T19:38:37+08:00", "document/content/docs/upgrading/4-8/41.mdx": "2025-08-02T19:38:37+08:00",
"document/content/docs/upgrading/4-8/42.mdx": "2025-08-02T19:38:37+08:00", "document/content/docs/upgrading/4-8/42.mdx": "2025-08-02T19:38:37+08:00",
......
packages/global/core/dataset/search/utils.ts
...@@ -18,7 +18,7 @@ export const datasetSearchResultConcat = ( ...@@ -18,7 +18,7 @@ export const datasetSearchResultConcat = (
item.list.forEach((data, index) => { item.list.forEach((data, index) => {
const rank = index + 1; const rank = index + 1;
const score = (weight * 1) / (60 + rank); const score = weight * (1 / (60 + rank));
const record = map.get(data.id); const record = map.get(data.id);
if (record) { if (record) {
// 合并两个score,有相同type的score,取最大值 // 合并两个score,有相同type的score,取最大值
...@@ -64,8 +64,9 @@ export const datasetSearchResultConcat = ( ...@@ -64,8 +64,9 @@ export const datasetSearchResultConcat = (
}); });
} }
// @ts-ignore return {
delete item.rrfScore; ...item,
return item; rrfScore: undefined
};
}); });
}; };
packages/service/core/dataset/search/controller.ts
...@@ -849,22 +849,16 @@ export async function searchDatasetData( ...@@ -849,22 +849,16 @@ export async function searchDatasetData(
} }
})(); })();
// embedding recall and fullText recall rrf concat
const embWeight = embeddingWeight; // 向量索引的 weight 大小
const fullTextWeight = 1 - embeddingWeight; // 全文索引的 weight 大小
const rrfSearchResult = datasetSearchResultConcat([ const rrfSearchResult = datasetSearchResultConcat([
{ weight: embWeight, list: embeddingRecallResults }, { weight: embeddingWeight, list: embeddingRecallResults },
{ weight: fullTextWeight, list: fullTextRecallResults } { weight: 1 - embeddingWeight, list: fullTextRecallResults }
]); ]);
const rrfConcatResults = (() => { const rrfConcatResults = (() => {
if (reRankResults.length === 0) return rrfSearchResult; if (reRankResults.length === 0) return rrfSearchResult;
if (rerankWeight === 1) return reRankResults; if (rerankWeight === 1) return reRankResults;
const searchWeight = 1 - rerankWeight; // 搜索结果的 weight 大小
return datasetSearchResultConcat([ return datasetSearchResultConcat([
{ weight: searchWeight, list: rrfSearchResult }, { weight: 1 - rerankWeight, list: rrfSearchResult },
{ weight: rerankWeight, list: reRankResults } { weight: rerankWeight, list: reRankResults }
]); ]);
})(); })();
......
test/cases/function/packages/global/core/dataset/search/utils.test.ts 0 → 100644
import { describe, it, expect } from 'vitest';
import { datasetSearchResultConcat } from '@fastgpt/global/core/dataset/search/utils';
import { SearchScoreTypeEnum } from '@fastgpt/global/core/dataset/constants';
import type { SearchDataResponseItemType } from '@fastgpt/global/core/dataset/type';
describe('datasetSearchResultConcat', () => {
// Helper function to create test data
const createSearchItem = (
id: string,
q: string,
scores: { type: `${SearchScoreTypeEnum}`; value: number; index: number }[] = []
): SearchDataResponseItemType => ({
id,
datasetId: 'dataset1',
collectionId: 'collection1',
sourceName: 'source1',
sourceId: 'source1',
q,
a: `Answer for ${q}`,
chunkIndex: 0,
updateTime: new Date(),
score: scores
});
describe('Edge cases', () => {
it('should handle empty array', () => {
const result = datasetSearchResultConcat([]);
expect(result).toEqual([]);
});
it('should handle all empty lists', () => {
const input = [
{ weight: 1.0, list: [] },
{ weight: 0.5, list: [] }
];
const result = datasetSearchResultConcat(input);
expect(result).toEqual([]);
});
it('should handle only one non-empty list', () => {
const items = [
createSearchItem('1', 'Question 1', [
{ type: SearchScoreTypeEnum.embedding, value: 0.9, index: 0 }
]),
createSearchItem('2', 'Question 2', [
{ type: SearchScoreTypeEnum.embedding, value: 0.8, index: 1 }
])
];
const input = [
{ weight: 1.0, list: items },
{ weight: 0.5, list: [] }
];
const result = datasetSearchResultConcat(input);
expect(result).toEqual(items);
});
});
describe('RRF algorithm tests', () => {
it('should calculate RRF scores correctly', () => {
const items1 = [
createSearchItem('1', 'Question 1', [
{ type: SearchScoreTypeEnum.embedding, value: 0.9, index: 0 }
]),
createSearchItem('2', 'Question 2', [
{ type: SearchScoreTypeEnum.embedding, value: 0.8, index: 1 }
])
];
const items2 = [
createSearchItem('2', 'Question 2', [
{ type: SearchScoreTypeEnum.fullText, value: 0.7, index: 0 }
]),
createSearchItem('1', 'Question 1', [
{ type: SearchScoreTypeEnum.fullText, value: 0.6, index: 1 }
])
];
const input = [
{ weight: 1.0, list: items1 },
{ weight: 1.0, list: items2 }
];
const result = datasetSearchResultConcat(input);
// Verify RRF score calculation
// item1: 1.0 * (1/(60+1)) + 1.0 * (1/(60+2)) = 1/61 + 1/62 ≈ 0.0163934 + 0.0161290 ≈ 0.0325224
// item2: 1.0 * (1/(60+2)) + 1.0 * (1/(60+1)) = 1/62 + 1/61 ≈ 0.0161290 + 0.0163934 ≈ 0.0325224
expect(result).toHaveLength(2);
// Verify RRF scores are added
result.forEach((item) => {
const rrfScore = item.score.find((s) => s.type === SearchScoreTypeEnum.rrf);
expect(rrfScore).toBeDefined();
expect(rrfScore!.value).toBeCloseTo(0.0325224, 6);
});
});
it('should weight RRF scores correctly', () => {
const items1 = [
createSearchItem('1', 'Question 1', [
{ type: SearchScoreTypeEnum.embedding, value: 0.9, index: 0 }
])
];
const items2 = [
createSearchItem('1', 'Question 1', [
{ type: SearchScoreTypeEnum.fullText, value: 0.7, index: 0 }
])
];
const input = [
{ weight: 2.0, list: items1 }, // Higher weight
{ weight: 1.0, list: items2 }
];
const result = datasetSearchResultConcat(input);
expect(result).toHaveLength(1);
const rrfScore = result[0].score.find((s) => s.type === SearchScoreTypeEnum.rrf);
expect(rrfScore).toBeDefined();
// Should be: 2.0 * (1/61) + 1.0 * (1/61) = 3.0 * (1/61) ≈ 0.0491803
expect(rrfScore!.value).toBeCloseTo(3.0 / 61, 6);
});
});
describe('Score merging tests', () => {
it('should merge different score types correctly', () => {
const items1 = [
createSearchItem('1', 'Question 1', [
{ type: SearchScoreTypeEnum.embedding, value: 0.9, index: 0 },
{ type: SearchScoreTypeEnum.reRank, value: 0.8, index: 0 }
])
];
const items2 = [
createSearchItem('1', 'Question 1', [
{ type: SearchScoreTypeEnum.fullText, value: 0.7, index: 0 }
])
];
const input = [
{ weight: 1.0, list: items1 },
{ weight: 1.0, list: items2 }
];
const result = datasetSearchResultConcat(input);
expect(result).toHaveLength(1);
expect(result[0].score).toHaveLength(4); // embedding, reRank, fullText, rrf
// Verify all score types exist
const scoreTypes = result[0].score.map((s) => s.type);
expect(scoreTypes).toContain(SearchScoreTypeEnum.embedding);
expect(scoreTypes).toContain(SearchScoreTypeEnum.fullText);
expect(scoreTypes).toContain(SearchScoreTypeEnum.reRank);
expect(scoreTypes).toContain(SearchScoreTypeEnum.rrf);
});
it('should take max value for same score types', () => {
const items1 = [
createSearchItem('1', 'Question 1', [
{ type: SearchScoreTypeEnum.embedding, value: 0.9, index: 0 }
])
];
const items2 = [
createSearchItem('1', 'Question 1', [
{ type: SearchScoreTypeEnum.embedding, value: 0.7, index: 0 } // Lower score
])
];
const input = [
{ weight: 1.0, list: items1 },
{ weight: 1.0, list: items2 }
];
const result = datasetSearchResultConcat(input);
expect(result).toHaveLength(1);
const embeddingScore = result[0].score.find((s) => s.type === SearchScoreTypeEnum.embedding);
expect(embeddingScore).toBeDefined();
expect(embeddingScore!.value).toBe(0.9); // Should take higher value
});
});
describe('Sorting tests', () => {
it('should sort by RRF score descending', () => {
const items1 = [
createSearchItem('1', 'Question 1', [
{ type: SearchScoreTypeEnum.embedding, value: 0.9, index: 0 }
]),
createSearchItem('2', 'Question 2', [
{ type: SearchScoreTypeEnum.embedding, value: 0.8, index: 1 }
]),
createSearchItem('3', 'Question 3', [
{ type: SearchScoreTypeEnum.embedding, value: 0.7, index: 2 }
])
];
const items2 = [
createSearchItem('3', 'Question 3', [
{ type: SearchScoreTypeEnum.fullText, value: 0.9, index: 0 }
]), // First position, higher RRF
createSearchItem('2', 'Question 2', [
{ type: SearchScoreTypeEnum.fullText, value: 0.8, index: 1 }
]),
createSearchItem('1', 'Question 1', [
{ type: SearchScoreTypeEnum.fullText, value: 0.7, index: 2 }
]) // Third position, lower RRF
];
const input = [
{ weight: 1.0, list: items1 },
{ weight: 1.0, list: items2 }
];
const result = datasetSearchResultConcat(input);
expect(result).toHaveLength(3);
// Verify descending RRF score order
for (let i = 0; i < result.length - 1; i++) {
const currentRrf = result[i].score.find((s) => s.type === SearchScoreTypeEnum.rrf)!.value;
const nextRrf = result[i + 1].score.find((s) => s.type === SearchScoreTypeEnum.rrf)!.value;
expect(currentRrf).toBeGreaterThanOrEqual(nextRrf);
}
// item1 and item3 have same RRF score, but item1 should be first due to stable sort order
expect(['1', '3']).toContain(result[0].id);
});
});
describe('RRF score update tests', () => {
it('should update existing RRF scores when multiple lists', () => {
const items1 = [
createSearchItem('1', 'Question 1', [
{ type: SearchScoreTypeEnum.embedding, value: 0.9, index: 0 },
{ type: SearchScoreTypeEnum.rrf, value: 0.5, index: 0 } // Existing RRF score
])
];
const items2 = [
createSearchItem('1', 'Question 1', [
{ type: SearchScoreTypeEnum.fullText, value: 0.7, index: 0 }
])
];
const input = [
{ weight: 1.0, list: items1 },
{ weight: 1.0, list: items2 }
];
const result = datasetSearchResultConcat(input);
expect(result).toHaveLength(1);
const rrfScores = result[0].score.filter((s) => s.type === SearchScoreTypeEnum.rrf);
expect(rrfScores).toHaveLength(1); // Should only have one RRF score
// RRF score should be updated to calculated value, not the original 0.5
expect(rrfScores[0].value).not.toBe(0.5);
expect(rrfScores[0].value).toBeCloseTo(1.0 / 61 + 1.0 / 61, 6);
expect(rrfScores[0].index).toBe(0); // Index after sorting
});
it('should add RRF score for items without one when multiple lists', () => {
const items1 = [
createSearchItem('1', 'Question 1', [
{ type: SearchScoreTypeEnum.embedding, value: 0.9, index: 0 }
// No RRF score
])
];
const items2 = [
createSearchItem('1', 'Question 1', [
{ type: SearchScoreTypeEnum.fullText, value: 0.7, index: 0 }
])
];
const input = [
{ weight: 1.0, list: items1 },
{ weight: 1.0, list: items2 }
];
const result = datasetSearchResultConcat(input);
expect(result).toHaveLength(1);
const rrfScore = result[0].score.find((s) => s.type === SearchScoreTypeEnum.rrf);
expect(rrfScore).toBeDefined();
expect(rrfScore!.value).toBeCloseTo(1.0 / 61 + 1.0 / 61, 6);
expect(rrfScore!.index).toBe(0);
});
it('should not modify single list (direct return)', () => {
const items1 = [
createSearchItem('1', 'Question 1', [
{ type: SearchScoreTypeEnum.embedding, value: 0.9, index: 0 }
])
];
const input = [{ weight: 1.0, list: items1 }];
const result = datasetSearchResultConcat(input);
expect(result).toHaveLength(1);
expect(result).toEqual(items1); // Should be exactly the same as input
// Should not have RRF score because single list is returned directly
const rrfScore = result[0].score.find((s) => s.type === SearchScoreTypeEnum.rrf);
expect(rrfScore).toBeUndefined();
});
});
describe('Complex scenario tests', () => {
it('should handle complex multi-source merging', () => {
const embeddingResults = [
createSearchItem('doc1', 'AI Introduction', [
{ type: SearchScoreTypeEnum.embedding, value: 0.95, index: 0 }
]),
createSearchItem('doc2', 'Machine Learning Basics', [
{ type: SearchScoreTypeEnum.embedding, value: 0.9, index: 1 }
]),
createSearchItem('doc3', 'Deep Learning Principles', [
{ type: SearchScoreTypeEnum.embedding, value: 0.85, index: 2 }
])
];
const fullTextResults = [
createSearchItem('doc2', 'Machine Learning Basics', [
{ type: SearchScoreTypeEnum.fullText, value: 0.88, index: 0 }
]),
createSearchItem('doc4', 'Neural Network Applications', [
{ type: SearchScoreTypeEnum.fullText, value: 0.82, index: 1 }
]),
createSearchItem('doc1', 'AI Introduction', [
{ type: SearchScoreTypeEnum.fullText, value: 0.78, index: 2 }
])
];
const reRankResults = [
createSearchItem('doc3', 'Deep Learning Principles', [
{ type: SearchScoreTypeEnum.reRank, value: 0.92, index: 0 }
]),
createSearchItem('doc1', 'AI Introduction', [
{ type: SearchScoreTypeEnum.reRank, value: 0.89, index: 1 }
])
];
const input = [
{ weight: 1.0, list: embeddingResults },
{ weight: 0.8, list: fullTextResults },
{ weight: 1.2, list: reRankResults }
];
const result = datasetSearchResultConcat(input);
// Should have 4 unique documents
expect(result).toHaveLength(4);
// Verify all documents have RRF scores
result.forEach((item) => {
const rrfScore = item.score.find((s) => s.type === SearchScoreTypeEnum.rrf);
expect(rrfScore).toBeDefined();
expect(rrfScore!.value).toBeGreaterThan(0);
});
// Verify merged scores
const doc1 = result.find((item) => item.id === 'doc1')!;
const doc1ScoreTypes = doc1.score.map((s) => s.type);
expect(doc1ScoreTypes).toContain(SearchScoreTypeEnum.embedding);
expect(doc1ScoreTypes).toContain(SearchScoreTypeEnum.fullText);
expect(doc1ScoreTypes).toContain(SearchScoreTypeEnum.reRank);
expect(doc1ScoreTypes).toContain(SearchScoreTypeEnum.rrf);
// Verify sorting by RRF score descending
for (let i = 0; i < result.length - 1; i++) {
const currentRrf = result[i].score.find((s) => s.type === SearchScoreTypeEnum.rrf)!.value;
const nextRrf = result[i + 1].score.find((s) => s.type === SearchScoreTypeEnum.rrf)!.value;
expect(currentRrf).toBeGreaterThanOrEqual(nextRrf);
}
});
});
describe('Edge weight tests', () => {
it('should handle zero weight', () => {
const items1 = [
createSearchItem('1', 'Question 1', [
{ type: SearchScoreTypeEnum.embedding, value: 0.9, index: 0 }
])
];
const items2 = [
createSearchItem('1', 'Question 1', [
{ type: SearchScoreTypeEnum.fullText, value: 0.7, index: 0 }
])
];
const input = [
{ weight: 0, list: items1 },
{ weight: 1.0, list: items2 }
];
const result = datasetSearchResultConcat(input);
expect(result).toHaveLength(1);
const rrfScore = result[0].score.find((s) => s.type === SearchScoreTypeEnum.rrf);
expect(rrfScore).toBeDefined();
expect(rrfScore!.value).toBeCloseTo(1.0 / 61, 6); // Only from second list
});
it('should handle negative weight', () => {
const items1 = [
createSearchItem('1', 'Question 1', [
{ type: SearchScoreTypeEnum.embedding, value: 0.9, index: 0 }
])
];
const items2 = [
createSearchItem('1', 'Question 1', [
{ type: SearchScoreTypeEnum.fullText, value: 0.7, index: 0 }
])
];
const input = [
{ weight: -1.0, list: items1 },
{ weight: 1.0, list: items2 }
];
const result = datasetSearchResultConcat(input);
expect(result).toHaveLength(1);
const rrfScore = result[0].score.find((s) => s.type === SearchScoreTypeEnum.rrf);
expect(rrfScore).toBeDefined();
// Should be: -1.0 * (1/61) + 1.0 * (1/61) = 0
expect(rrfScore!.value).toBeCloseTo(0, 6);
});
it('should handle very small weight', () => {
const items1 = [
createSearchItem('1', 'Question 1', [
{ type: SearchScoreTypeEnum.embedding, value: 0.9, index: 0 }
])
];
const items2 = [
createSearchItem('1', 'Question 1', [
{ type: SearchScoreTypeEnum.fullText, value: 0.7, index: 0 }
])
];
const input = [
{ weight: 0.001, list: items1 },
{ weight: 1.0, list: items2 }
];
const result = datasetSearchResultConcat(input);
expect(result).toHaveLength(1);
const rrfScore = result[0].score.find((s) => s.type === SearchScoreTypeEnum.rrf);
expect(rrfScore).toBeDefined();
expect(rrfScore!.value).toBeCloseTo(0.001 / 61 + 1.0 / 61, 6);
});
});
});
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or sign in to comment