zhuyongxin
376ad0c241
Add configurable hybrid mode on KnowledgeSearchPort that fuses dense unfiltered, dense filtered, and lexical ranks via RRF while preserving dense-compatible threshold scores. Archives Delivery 2 OpenSpec change.
59 lines
1.7 KiB
Markdown
59 lines
1.7 KiB
Markdown
# Design: rag-hybrid-search-rrf
|
|
|
|
## Context
|
|
|
|
- Delivery 1 archived: chunk evidenceKey, SearchPort, return caps.
|
|
- Legacy SDK path will be abandoned later; hybrid must live on SearchPort.
|
|
- Full sparse schema rebuild is operationally heavy; ship fusion first.
|
|
|
|
## Decisions
|
|
|
|
### D1. Hybrid = multi-path + RRF on SearchPort
|
|
|
|
When `retrieval.search.mode=hybrid`:
|
|
|
|
```text
|
|
paths:
|
|
1) dense(query, filter=null) # always
|
|
2) dense(query, filter=category) # if category present
|
|
3) lexical rank over union of dense hits # sparse-lite
|
|
fuse by evidenceKey using RRF(k)
|
|
return topK fused hits
|
|
```
|
|
|
|
### D2. RRF formula
|
|
|
|
```text
|
|
score(d) = Σ w_i / (k + rank_i(d))
|
|
```
|
|
|
|
Defaults: k=60, all w_i=1.0. Optional weights via config.
|
|
|
|
### D3. Score semantics
|
|
|
|
- `KnowledgeSearchHit.score` remains **compatible L2 distance from best dense hit** for post-process thresholds.
|
|
- Fused RRF score is carried in metadata (`fusedScore`, `fusionRanks`) for trace, not as L2.
|
|
|
|
### D4. Lexical path (sparse-lite)
|
|
|
|
Until BM25 schema:
|
|
|
|
- Tokenize query (simple whitespace / non-alnum split, lower-case)
|
|
- Score each candidate by term coverage over title+breadcrumb+content
|
|
- Rank candidates for RRF path only
|
|
- Not a replacement for true BM25 inverted index
|
|
|
|
### D5. Serial filter retry
|
|
|
|
Lookup tool keeps low-quality unfiltered retry as safety net even in hybrid, because hybrid already includes unfiltered dense; retry remains cheap no-op when already fused well.
|
|
|
|
## Non-goals now
|
|
|
|
- New Milvus collection fields
|
|
- Reindex jobs
|
|
- SDK hybridSearch API calls
|
|
|
|
## Risks
|
|
|
|
- Lexical path only ranks already-recalled dense candidates → does not expand pure-term misses outside dense topK. Acceptable intermediate; true BM25 later expands recall.
|