Scaling Test-Time Compute in Search Mode | Weaviate

What this post added

Introduces the 'effort' parameter for the Query Agent's Search Mode, with three tiers: 'medium', 'high', and 'ultrahigh'. This parameter scales computation at the query writing and reranking stages. The post provides benchmarks comparing these tiers against Hybrid Search across 8 datasets, showing accuracy improvements with increased effort, especially on complex reasoning tasks. It also details the methodology, metrics (Success@K, Recall@K, nDCG@K), and glossaries for benchmarks and metrics.