Semantic search menggunakan meaning (embedding) bukan keyword matching.
Pipeline
Query -> Embedding -> Vector Search -> Rerank -> Response
BM25 (keyword fallback)
Components
| Component | Tool | Purpose |
|---|---|---|
| Embedding | text-embedding-3-small | Convert query to vector |
| Vector Search | sqlite-vec / FAISS | ANN nearest neighbor |
| Keyword Search | BM25 / Tantivy | Exact match fallback |
| Fusion | RRF | Hybrid ranking merge |
| Rerank | Cohere/cross-encoder | Re-rank top-K |