Amey-Thakur Multi-hop retrieval
Single-shot retrieval answers questions whose answer sits in one place. Questions that require connecting two facts fail silently, returning passages relevant to the surface words and missing the connection entirely.
Method
- Detect when one hop is insufficient. Comparisons, causal chains, and questions containing an unresolved reference usually need decomposition rather than a bigger top-k.
- Decompose into answerable sub-questions. Each one retrievable on its own, with the second often depending on the first's answer (see agent-plan-execute-replan).
- Feed each answer into the next query. The point of hopping is that the intermediate result narrows the next search, which a parallel fan-out does not achieve.
- Cap the hops. Two or three, then answer with what you have, because unbounded chains accumulate error and cost quickly.
- Carry provenance through every hop. The final answer must cite the sources from each step, not only the last (see citation-grounding).
- Detect and stop on dead ends. A hop that retrieves nothing relevant should end the chain with an honest partial answer rather than continuing on a guess.
- Verify the assembled chain. The individual facts may be right and the connection wrong, which is a distinct failure worth checking (see agent-generate-and-verify).
Boundaries
Multi-hop multiplies latency and cost by the number of hops, so it suits questions that need it rather than being the default. Error compounds across hops, making chains beyond a few steps unreliable. Some connections are genuinely absent from the corpus and no amount of hopping finds them.