Before you dive in
What you’ll find in the original
- Batch many questions only when they share one state; reranking different passages requires one state per candidate and concurrent requests instead.
- Use different confidence thresholds for branches with different consequences, keep escalation policy in ordinary code, and include a none-of-these option.
- Report rank bounds when a baseline ties: assigning an arbitrary sorted position to an unranked item can make the comparison look better than it is.
Worth knowing
The direct TypeSafe path follows documented interfaces but was not run; the author verified OpenRouter. The 100% reranking result covers six authored queries against 26 passages and is explicitly a demonstration, not a benchmark.