Query decomposition transforms a compound information need into focused atomic queries before retrieval. Each sub-query receives its own embedding, route classification, and retrieval call, improving precision across distinct information needs. section 7.4 section 7.4.3
A practical decomposer follows this flow:
- Preserve entities and time periods, then route and retrieve each sub-query against the appropriate collection. section 7.4.3
- Return a single-element array containing the original query when no decomposition is needed. section 7.4.3
- Combine the focused retrieval results during final response synthesis. section 7.4.3
Comparing Uber and Lyft
The compound question “What was Uber's revenue in 2021, and how does its gross bookings growth compare to Lyft's?” becomes three queries: Uber's 2021 revenue, Uber's 2021 gross bookings growth, and Lyft's 2021 gross bookings growth. Separate embeddings and retrieval calls target the relevant filing chunks instead of producing one averaged ride-share embedding. section 7.4.3


