diff --git a/server/core/retriever.go b/server/core/retriever.go
index 6094a02..679df0c 100644
--- a/server/core/retriever.go
+++ b/server/core/retriever.go
@@ -14,6 +14,7 @@ import (
 	"github.com/qdrant/go-client/qdrant"
 	"github.com/wangle201210/go-rag/server/core/common"
 	"github.com/wangle201210/go-rag/server/core/rerank"
+	"github.com/wangle201210/go-rag/server/core/retriever"
 	coretypes "github.com/wangle201210/go-rag/server/core/types"
 )
 
@@ -114,23 +115,34 @@ func (x *Rag) retrieveDoOnce(ctx context.Context, req *RetrieveReq) (relatedDocs
 		qaDocs []*schema.Document
 	)
 	g.Log().Infof(ctx, "query: %v", req.optQuery)
-	// 通过内容检索
+	// 向量路一：content_vector
 	docs, err = x.retrieve(ctx, req, false)
 	if err != nil {
 		g.Log().Errorf(ctx, "retrieve failed, err=%v", err)
 		return
 	}
-	// 通过qa检索
+	// 向量路二：qa_content_vector
 	qaDocs, err = x.retrieve(ctx, req, true)
 	if err != nil {
 		g.Log().Errorf(ctx, "qa retrieve failed, err=%v", err)
 		return
 	}
-	docs = append(docs, qaDocs...)
-	// 去重
-	docs = common.RemoveDuplicates(docs, func(doc *schema.Document) string {
-		return doc.ID
-	})
+
+	inputs := [][]*schema.Document{docs, qaDocs}
+
+	// 关键字路三：ES BM25（conf.Client != nil 时生效）
+	if x.conf.Client != nil {
+		bm25Docs, bm25Err := retriever.Bm25Retrieve(ctx, x.conf.Client, x.conf.IndexName, req.KnowledgeName, req.optQuery, req.excludeIDs, esTopK)
+		if bm25Err != nil {
+			g.Log().Warningf(ctx, "bm25 retrieve failed, skip: %v", bm25Err)
+		} else {
+			inputs = append(inputs, bm25Docs)
+		}
+	}
+
+	// RRF 融合多路结果（同时完成去重）
+	docs = retriever.RRFFusion(inputs)
+
 	// 重排
 	docs, err = rerank.NewRerank(ctx, req.optQuery, docs, req.TopK)
 	if err != nil {
diff --git a/server/core/retriever/bm25_retriever.go b/server/core/retriever/bm25_retriever.go
new file mode 100644
index 0000000..53b4e71
--- /dev/null
+++ b/server/core/retriever/bm25_retriever.go
@@ -0,0 +1,64 @@
+package retriever
+
+import (
+	"context"
+	"fmt"
+
+	"github.com/cloudwego/eino/schema"
+	"github.com/elastic/go-elasticsearch/v8"
+	"github.com/elastic/go-elasticsearch/v8/typedapi/core/search"
+	"github.com/elastic/go-elasticsearch/v8/typedapi/types"
+	coretypes "github.com/wangle201210/go-rag/server/core/types"
+)
+
+// Bm25Retrieve 使用 ES BM25 对 content 字段做全文检索。
+// 仅在 conf.ESClient != nil 时调用。
+func Bm25Retrieve(ctx context.Context, client *elasticsearch.Client, indexName, knowledgeName, query string, excludeIDs []string, topK int) ([]*schema.Document, error) {
+	must := []types.Query{
+		{Match: map[string]types.MatchQuery{
+			coretypes.FieldContent: {Query: query},
+		}},
+		{Bool: &types.BoolQuery{
+			Must: []types.Query{
+				{Match: map[string]types.MatchQuery{
+					coretypes.KnowledgeName: {Query: knowledgeName},
+				}},
+			},
+		}},
+	}
+
+	boolQuery := &types.BoolQuery{Must: must}
+
+	if len(excludeIDs) > 0 {
+		boolQuery.MustNot = []types.Query{
+			{Terms: &types.TermsQuery{
+				TermsQuery: map[string]types.TermsQueryField{
+					"_id": excludeIDs,
+				},
+			}},
+		}
+	}
+
+	sreq := search.NewRequest()
+	sreq.Query = &types.Query{Bool: boolQuery}
+	sreq.Size = &topK
+
+	resp, err := search.NewSearchFunc(client)().
+		Index(indexName).
+		Request(sreq).
+		Do(ctx)
+	if err != nil {
+		return nil, fmt.Errorf("bm25 search failed: %w", err)
+	}
+
+	var docs []*schema.Document
+	for _, hit := range resp.Hits.Hits {
+		doc, err := EsHit2Document(ctx, hit)
+		if err != nil {
+			continue
+		}
+		docs = append(docs, doc)
+	}
+
+	return docs, nil
+}
diff --git a/server/core/retriever/rrf.go b/server/core/retriever/rrf.go
new file mode 100644
index 0000000..b35db63
--- /dev/null
+++ b/server/core/retriever/rrf.go
@@ -0,0 +1,41 @@
+package retriever
+
+import (
+	"sort"
+
+	"github.com/cloudwego/eino/schema"
+)
+
+// RRFFusion 对多路召回结果执行 RRF（Reciprocal Rank Fusion）融合。
+// inputs 每个元素是一路召回的有序文档列表（按相关性降序）。
+func RRFFusion(inputs [][]*schema.Document) []*schema.Document {
+	const k = 60
+
+	docScores := make(map[string]float64)
+	docMap := make(map[string]*schema.Document)
+
+	for _, docs := range inputs {
+		for rank, doc := range docs {
+			if doc.ID == "" {
+				continue
+			}
+			docScores[doc.ID] += 1.0 / float64(k+rank+1)
+			if _, exists := docMap[doc.ID]; !exists {
+				docMap[doc.ID] = doc
+			}
+		}
+	}
+
+	result := make([]*schema.Document, 0, len(docMap))
+	for id, score := range docScores {
+		doc := docMap[id]
+		doc.WithScore(score)
+		result = append(result, doc)
+	}
+
+	sort.Slice(result, func(i, j int) bool {
+		return result[i].Score() > result[j].Score()
+	})
+
+	return result
+}
diff --git a/server/internal/logic/rag/retriever.go b/server/internal/logic/rag/retriever.go
index 49a0ca6..49c50e7 100644
--- a/server/internal/logic/rag/retriever.go
+++ b/server/internal/logic/rag/retriever.go
@@ -55,6 +55,19 @@ func init() {
 		client = esStore.GetClient()
 	} else if qdrantStore, ok := vectorStore.(*vector.QdrantVectorStore); ok {
 		qdrantClient = qdrantStore.GetClient()
+		// Qdrant 模式下，若配置了 ES 地址则初始化 ES 客户端用于 BM25 混合检索
+		esAddr := g.Cfg().MustGet(ctx, "vector.es.address").String()
+		if esAddr != "" {
+			client, err = elasticsearch.NewClient(elasticsearch.Config{
+				Addresses: []string{esAddr},
+				Username:  g.Cfg().MustGet(ctx, "vector.es.username").String(),
+				Password:  g.Cfg().MustGet(ctx, "vector.es.password").String(),
+			})
+			if err != nil {
+				g.Log().Warningf(ctx, "init ES client for BM25 failed, hybrid retrieval disabled: %v", err)
+				client = nil
+			}
+		}
 	}
 
 	ragSvr, err = core.New(ctx, &config.Config{