diff --git a/internal/application/repository/retriever/postgres/repository.go b/internal/application/repository/retriever/postgres/repository.go index cf357e08..3ff7796a 100644 --- a/internal/application/repository/retriever/postgres/repository.go +++ b/internal/application/repository/retriever/postgres/repository.go @@ -193,10 +193,13 @@ func (g *pgRepository) KeywordsRetrieve(ctx context.Context, Values: common.ToInterfaceSlice(params.TagIDs), }) } + + // Use ParadeDB's ||| operator for matching any token conds = append(conds, clause.Expr{ - SQL: "id @@@ paradedb.match(field => 'content', value => ?, distance => 1)", + SQL: "content ||| ?", Vars: []interface{}{params.Query}, }) + // Filter by is_enabled = true or NULL (NULL means enabled for historical data) conds = append(conds, clause.Expr{ SQL: "(is_enabled IS NULL OR is_enabled = ?)", diff --git a/internal/application/service/knowledgebase_search_fusion.go b/internal/application/service/knowledgebase_search_fusion.go index 0085a8bb..22c7adc3 100644 --- a/internal/application/service/knowledgebase_search_fusion.go +++ b/internal/application/service/knowledgebase_search_fusion.go @@ -3,9 +3,10 @@ package service import ( "context" + "slices" + "github.com/Tencent/WeKnora/internal/logger" "github.com/Tencent/WeKnora/internal/types" - "slices" ) // classifyRetrievalResults separates retrieval results by retriever type (vector vs keyword). @@ -35,6 +36,12 @@ func fuseOrDeduplicate(ctx context.Context, vectorResults, keywordResults []*typ logger.Infof(ctx, "Result count after deduplication: %d", len(result)) return result } + if len(vectorResults) == 0 { + // Keyword-only: keep original scores (important for FAQ) + result := deduplicateByScore(keywordResults) + logger.Infof(ctx, "Result count after deduplication: %d", len(result)) + return result + } // Hybrid: use RRF fusion to merge vector + keyword results result := fuseWithRRF(ctx, vectorResults, keywordResults) logger.Infof(ctx, "Result count after RRF fusion: %d", len(result))