Retrieved passages are relevant, but the best one is often ranked fifth and gets cut off. What helps?
Why D is the answer
A reranking step after retrieval
A reranker scores candidates more carefully than the first search, which lifts the best passage into what the model sees.
In detail
- A reranker scores candidates more carefully than the first search.
- That lifts the best passage into what the model sees.
- Rerank a wider candidate set, then pass the best few to the model.
Use case
An illustrative scenario, not a real incident.
The best passage ranks fifth and gets cut. A reranker scores the top candidates against the query and moves it to first place.
See it in different ways
flowchart LR A["Documents"] --> B["Chunk by structure"] B --> C["Index with metadata"] C --> D["Filter and search"] D --> E["Rerank"] E --> F["Answer with citations"]
mindmap
root((Retrieval))
Ingestion
Structure-aware chunks
Search
Keyword
Semantic
Ranking
Reranker
Access
Filter by userflowchart LR S0["Retrieve many"] --> S1["Rerank"] --> S2["Keep the best few"] --> S3["Send to the model"]
Key point: Retrieve broadly, then rerank.
Why the other options are wrong
- A. Fewer documents in the knowledge base, so that competing results disappear
- Removing documents loses knowledge, and may not fix the order.
- B. A system prompt asking the model to look harder at the passages it receives
- The model cannot look at a passage that was cut off.
- C. A higher temperature, so that the model considers more of the retrieved text
- Temperature does not change which passages arrive.