rrf_weight (#5551)

Co-authored-by: xxYyh <xxyyh@xxYyhdeMacBook-Pro.local>

.gitignore

@@ -37,4 +37,6 @@ files/helm/fastgpt/charts/*.tgz
 
 
 tmp/
 tmp/
 coverage
 coverage
-document/.source
+document/.source
\ No newline at end of file
+
+bun.lock
\ No newline at end of file

packages/global/core/dataset/search/utils.ts

@@ -3,7 +3,7 @@ import { type SearchDataResponseItemType } from '../type';
 
 
 /* dataset search result concat */
 /* dataset search result concat */
 export const datasetSearchResultConcat = (
 export const datasetSearchResultConcat = (
-  arr: { k: number; list: SearchDataResponseItemType[] }[]
+  arr: { weight: number; list: SearchDataResponseItemType[] }[]
 ): SearchDataResponseItemType[] => {
 ): SearchDataResponseItemType[] => {
   arr = arr.filter((item) => item.list.length > 0);
   arr = arr.filter((item) => item.list.length > 0);
 
 
@@ -14,12 +14,11 @@ export const datasetSearchResultConcat = (
 
 
   // rrf
   // rrf
   arr.forEach((item) => {
   arr.forEach((item) => {
-    const k = item.k;
+    const weight = item.weight;
 
 
     item.list.forEach((data, index) => {
     item.list.forEach((data, index) => {
       const rank = index + 1;
       const rank = index + 1;
-      const score = 1 / (k + 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

packages/service/core/dataset/search/controller.ts

@@ -784,10 +784,10 @@ export async function searchDatasetData(
 
 
     // rrf concat
     // rrf concat
     const rrfEmbRecall = datasetSearchResultConcat(
     const rrfEmbRecall = datasetSearchResultConcat(
-      embeddingRecallResults.map((list) => ({ k: 60, list }))
+      embeddingRecallResults.map((list) => ({ weight: 1, list }))
     ).slice(0, embeddingLimit);
     ).slice(0, embeddingLimit);
     const rrfFTRecall = datasetSearchResultConcat(
     const rrfFTRecall = datasetSearchResultConcat(
-      fullTextRecallResults.map((list) => ({ k: 60, list }))
+      fullTextRecallResults.map((list) => ({ weight: 1, list }))
     ).slice(0, fullTextLimit);
     ).slice(0, fullTextLimit);
 
 
     return {
     return {
@@ -850,24 +850,22 @@ export async function searchDatasetData(
   })();
   })();
 
 
   // embedding recall and fullText recall rrf concat
   // embedding recall and fullText recall rrf concat
-  const baseK = 120;
-  const embK = Math.round(baseK * (1 - embeddingWeight)); // 搜索结果的 k 值
-  const fullTextK = Math.round(baseK * embeddingWeight); // rerank 结果的 k 值
+  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([
-    { k: embK, list: embeddingRecallResults },
-    { k: fullTextK, list: fullTextRecallResults }
+    { weight: embWeight, list: embeddingRecallResults },
+    { weight: fullTextWeight, 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 searchK = Math.round(baseK * rerankWeight); // 搜索结果的 k 值
-    const rerankK = Math.round(baseK * (1 - rerankWeight)); // rerank 结果的 k 值
+    const searchWeight = 1 - rerankWeight; // 搜索结果的 weight 大小
 
 
     return datasetSearchResultConcat([
     return datasetSearchResultConcat([
-      { k: searchK, list: rrfSearchResult },
-      { k: rerankK, list: reRankResults }
+      { weight: searchWeight, list: rrfSearchResult },
+      { weight: rerankWeight, list: reRankResults }
     ]);
     ]);
   })();
   })();
 
 

packages/service/core/workflow/dispatch/dataset/concat.ts

@@ -29,7 +29,7 @@ export async function dispatchDatasetConcat(
 
 
   const rrfConcatResults = datasetSearchResultConcat(
   const rrfConcatResults = datasetSearchResultConcat(
     quoteList.map((list) => ({
     quoteList.map((list) => ({
-      k: 60,
+      weight: 1,
       list
       list
     }))
     }))
   );
   );