Files

T

zhuyongxin f035538531 feat(rag): dense+BM25 hybrid on MilvusClientV2, drop SDK path

Replace legacy MilvusServiceClient knowledge search/write with a single
MilvusClientV2 hybrid store (BM25 function + dense ANN + RRFRanker).
Use collection biz_hybrid and require knowledge reindex.

2026-07-27 18:49:20 +08:00

2.0 KiB

Raw Blame History

Design: rag-bm25-hybrid-drop-sdk

Decision: single backend

KnowledgeSearchPort
  -> MilvusHybridKnowledgeStore (MilvusClientV2 only)
Write
  -> VectorIndexService -> MilvusHybridKnowledgeStore

No retrieval.vector-store.mode=sdk|spring|auto for knowledge lookup.

Schema (`milvus.collection`, default `biz_hybrid`)

field	type	notes
id	VarChar PK	chunk id
content	VarChar	original chunk body returned to agent
search_text	VarChar + analyzer	BM25 input (title/path/content)
sparse_vector	SparseFloatVector	BM25 function output
vector	FloatVector	dense embedding
metadata	JSON	docId, chunkIndex, category, kb_scope, ...

Function: BM25(search_text -> sparse_vector)

Indexes:

vector: IVF_FLAT / L2 (or COSINE if configured later)
sparse_vector: SPARSE_INVERTED_INDEX / BM25

Hybrid query

AnnSearchReq(vector, FloatVec(queryEmbedding), topK, filter?)
AnnSearchReq(sparse_vector, EmbeddedText(query), topK, filter?)
HybridSearchReq + RRFRanker(k)

Dense-compatible score for thresholds: if entity lacks dense distance, map fused score conservatively or re-use dense-only probe. Preferred: keep dense distance when available from a parallel dense search metadata; for hybrid hits use inverse-rank placeholder only in metadata and set score from optional dense sub-hit map.

Implementation approach for threshold stability:

Run hybridSearch for ordering/identity
Build map evidenceKey -> dense L2 from a concurrent dense search (same filter/topK)
Attach dense score onto fused hits when present; else maxL2 (low similarity) so weak BM25-only hits don't fake PRECISE

Migration

New collection name avoids mutating legacy biz
Operators re-run knowledge init / document reindex
Document in acceptance

Drop SDK search

Delete SDK search methods usage from knowledge path
MilvusServiceClient bean may remain temporarily only if other non-search utilities need it; prefer migrate write/delete fully to V2 and stop creating V1 client if unused

2.0 KiB Raw Blame History