Use case · Related prototype
Rank search results
Put useful matches near the top.
- Query and first matches
- Score relevance
- Return a ranked list
A knowledge base can show useful documents without generating an answer. A second scorer can reorder the first matches when word overlap alone gives a poor ranking.
Where the small model fits
A small model reads each query-document pair and returns a relevance score. An adaptive version could reserve a larger scorer for close calls, provided that deciding when to stop costs less than it saves.
What would need to work
Use held-out relevance judgments, ranking quality and total query time. The current reranker experiment is not evidence that early stopping preserves those rankings.