Retrieved passages are relevant but poorly ordered. What helps?
Why C is the answer
A reranker
A reranker scores candidates against the query.
In detail
- A reranker scores candidates against the query.
- It scores against the query.
Use case
An illustrative scenario, not a real incident.
The best passage is ranked seventh, and a reranker moves it to the top.
See it in different ways
flowchart LR A["Ingest and chunk"] --> B["Embed"] B --> C["Index"] D["Agent query"] --> C C --> E["Rerank"] E --> F["Context for the agent"]
stateDiagram-v2 [*] --> Parse: layout-aware Parse --> Index Index --> HybridSearch: keyword and vector HybridSearch --> Permissions: filter by user Permissions --> Rerank Rerank --> [*]
flowchart LR S0["Retrieve"] --> S1["Rerank"] --> S2["Keep the best"]
Key point: Rerankers improve ordering.
Why the other options are wrong
- A. A smaller model
- A smaller model does not reorder.
- B. More attention heads
- Heads are unrelated.
- D. A lower maximum output
- Output length is unrelated.