OpenRAG is a comprehensive, single package Retrieval-Augmented Generation platform built on Langflow, Docling, and Opensearch.
Python
fix: tune search relevance for partial and unique queries
Olfa Maslahdb88ef82d9be
diff --git a/frontend/app/api/queries/useGetSearchQuery.ts b/frontend/app/api/queries/useGetSearchQuery.tsindex 88bb9d45e..91504da78 100644--- a/frontend/app/api/queries/useGetSearchQuery.ts+++ b/frontend/app/api/queries/useGetSearchQuery.ts@@ -70,7 +70,6 @@ export const useGetSearchQuery = (options?: Omit<UseQueryOptions, "queryKey" | "queryFn">,) => {const queryClient = useQueryClient();-const getFileIdentity = (chunk: ChunkResult): string => {const normalizedFilename = chunk.filename?.trim();if (normalizedFilename) {@@ -87,6 +86,7 @@ export const useGetSearchQuery = (// Normalize the query to match what will actually be searchedconst effectiveQuery = query || queryData?.query || "*";+ const normalizedQuery = effectiveQuery.trim();async function getFiles(): Promise<File[]> {try {@@ -98,10 +98,21 @@ export const useGetSearchQuery = (? SEARCH_CONSTANTS.WILDCARD_QUERY_LIMIT: queryData?.limit || 100;+ const baseScoreThreshold =+ queryData?.scoreThreshold ?? SEARCH_CONSTANTS.DEFAULT_SCORE_THRESHOLD;+ const isShortSingleTokenQuery =+ normalizedQuery !== "*" &&+ normalizedQuery.length > 0 &&+ normalizedQuery.length <= 4 &&+ !normalizedQuery.includes(" ");+ const dynamicScoreThreshold = isShortSingleTokenQuery+ ? Math.min(baseScoreThreshold, 1.0)+ : baseScoreThreshold;+const searchPayload: SearchPayload = {query: effectiveQuery,limit: searchLimit,- scoreThreshold: queryData?.scoreThreshold || 0,+ scoreThreshold: dynamicScoreThreshold,};if (queryData?.filters) {searchPayload.filters =diff --git a/frontend/lib/constants.ts b/frontend/lib/constants.tsindex f0e1a70eb..cdc7b2ec3 100644--- a/frontend/lib/constants.ts+++ b/frontend/lib/constants.ts@@ -30,6 +30,7 @@ export const UI_CONSTANTS = {*/export const SEARCH_CONSTANTS = {WILDCARD_QUERY_LIMIT: 10000, // Maximum allowed limit for wildcard searches+ DEFAULT_SCORE_THRESHOLD: 1.25, // Default relevance threshold for knowledge search} as const;export const ANIMATION_DURATION = 0.4;diff --git a/src/services/search_service.py b/src/services/search_service.pyindex b1992efe0..0ac9c3e1b 100644--- a/src/services/search_service.py+++ b/src/services/search_service.py@@ -315,10 +315,21 @@ class SearchService:"query": query,"fields": ["text^2", "filename^1.5"],"type": "best_fields",- "fuzziness": "AUTO",+ "operator": "and",+ "fuzziness": "AUTO:4,7","boost": 0.3, # 30% weight for keyword search}},+ {+ # Help prefix/partial inputs (e.g. "vita" -> "vitamin")+ # without making the whole query fully fuzzy.+ "multi_match": {+ "query": query,+ "fields": ["text^1.5", "filename^2"],+ "type": "bool_prefix",+ "boost": 0.25,+ }+ },],"minimum_should_match": 1,"filter": all_filters,@@ -479,15 +490,35 @@ class SearchService:})+ # If query text appears verbatim in one subset of files, prefer those files+ # to avoid broad semantic spillover for unique lookups.+ normalized_query = query.strip().lower()+ if (+ normalized_query+ and not is_wildcard_match_all+ and len(normalized_query) >= 4+ ):+ exact_files = {+ filename+ for chunk in chunks+ for filename in [chunk.get("filename")]+ if isinstance(filename, str)+ and (+ normalized_query in filename.lower()+ or (+ isinstance(chunk.get("text"), str)+ and normalized_query in chunk.get("text", "").lower()+ )+ )+ }+ if exact_files:+ chunks = [chunk for chunk in chunks if chunk.get("filename") in exact_files]+# Return both transformed results and aggregationsreturn {"results": chunks,"aggregations": results.get("aggregations", {}),- "total": (- results.get("hits", {}).get("total", {}).get("value")- if isinstance(results.get("hits", {}).get("total"), dict)- else results.get("hits", {}).get("total")- ),+ "total": len(chunks),}async def search(