Scaling Test-Time Compute in Search Mode | Weaviate

This post details the implementation and benchmarking of the 'effort' parameter within Weaviate's Query Agent Search Mode. It explains how the 'effort' tiers (medium, high, ultrahigh) scale computation at query writing and reranking stages to balance accuracy, latency, and cost. The post provides detailed benchmark results across various datasets (BRIGHT, IRPAPERS, WixQA, OBLIQ-Bench) comparing different effort levels against Hybrid Search, highlighting the trade-offs and performance gains, especially on reasoning-intensive tasks. It also elaborates on the Search Mode's capabilities beyond document retrieval, including combining semantic meaning with structured constraints.