Multi-stage retrieval improves search precision by splitting retrieval and ranking into specialized steps. Hybrid dense and sparse retrieval can help find relevant documents, but relevant results may still be buried below less-relevant ones. section 4.3 section 4.3.3
A typical pipeline progressively narrows and ranks candidate documents:
- Broad retrieval casts a wide net to find potentially relevant documents. section 4.3.3
- Progressive filtering applies increasingly selective filters to narrow the candidates. section 4.3.3
- Precision reranking uses more sophisticated models to sort the remaining results. section 4.3.3
Bi-encoders encode queries and documents separately, allowing document vectors to be precomputed and compared with query vectors. Cross-encoders instead process a query and document together as an input pair. section 4.3.3
1. "Company holiday and leave policies for 2023" [marginally relevant]
2. "Benefits enrollment deadlines" [not relevant]
3. "Employees may request remote work accommodations under ADA" [highly
relevant!]
4. "General remote work guidelines" [somewhat relevant]
5. "New office opening announcement" [not relevant]
The ranked results for a remote-work accommodation query show the precision problem: a highly relevant policy appears below marginal or irrelevant documents. section 4.3.3
