--- title: "Scaling Test-Time Compute in Search Mode" url: https://daily.dev/posts/scaling-test-time-compute-in-search-mode-ctvttd5ti source_url: https://weaviate.io/blog/search-mode-effort type: article source: "Weaviate" published: 2026-08-11T15:14:43.053Z updated: 2026-08-11T15:15:20.352Z tags: ["rag", "vector-search", "weaviate"] reading_time: 8 upvotes: 2 comments: 2 language: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Scaling Test-Time Compute in Search Mode **[Weaviate](https://daily.dev/sources/weaviate)** · 8 min read · 2 upvotes · 2 comments ## Summary Weaviate has released an `effort` parameter for the Query Agent's Search Mode in `weaviate-agents 1.8.0` and `agents-typescript-client 1.7.0`. Three tiers — medium, high, and ultrahigh — control how much compute is spent on query decomposition and reranking at inference time. Benchmarks across 8 datasets (BRIGHT, IRPAPERS, WixQA, OBLIQ-Bench) show consistent accuracy gains over Hybrid Search at every tier, with the largest improvements on reasoning-intensive retrieval tasks. On BRIGHT Biology, ultrahigh effort lifts nDCG@10 from 13.0 to 57.5 — roughly a 4x improvement. On simpler domain-specific benchmarks like WixQA, medium effort already captures most of the gain. The tradeoff is latency and cost, making tier selection a per-request decision based on application needs. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### What nDCG@10 score does Weaviate Search Mode ultrahigh effort achieve on BRIGHT Biology compared to Hybrid Search? Ultrahigh effort Search Mode achieves nDCG@10 of 57.5 on BRIGHT Biology, compared to 13.0 with Hybrid Search alone — roughly a 4x improvement. The gains are largest on reasoning-intensive benchmarks like BRIGHT, where queries demand multi-step reasoning. On simpler domain-specific benchmarks like WixQA and IRPAPERS, medium effort already captures most of the gain over Hybrid Search. _Teams evaluating retrieval quality for agentic or RAG workloads track benchmark comparisons like these on daily.dev._ ### What versions of the Weaviate agents library introduced the effort parameter for Search Mode? The `effort` parameter for Search Mode was introduced in `weaviate-agents 1.8.0` for Python and `agents-typescript-client 1.7.0` for JavaScript/TypeScript. It accepts three tiers — medium, high, and ultrahigh — and can be set per request, letting applications balance accuracy against latency and cost depending on the query's importance. _Developers shipping Weaviate upgrades watch for version-specific feature releases like this on daily.dev._ ## Community discussion Top comments from developers on daily.dev. **@agustinbarrientos** · 0 upvotes > Ultrahigh belongs in the retry path, not the default. **@kartiknvj** · 0 upvotes > That nDCG@10 jump from 13 to 57.5 on BRIGHT Biology is a huge spread, though it makes me want the latency and cost numbers right next to it, since ultrahigh effort is presumably not free. An exposed effort dial is a nice pattern because it lets you spend compute only on the hard queries instead of flat-rating everything. Did the gains hold on easier datasets, or was BRIGHT the one where decomposition paid off most? ## Similar posts on daily.dev - [Search Mode Benchmarking](https://daily.dev/posts/search-mode-benchmarking-esibk4v40) · Weaviate · 0 upvotes · 0 comments - [Accelerating Data Workflows with Query Agent, now GA](https://daily.dev/posts/accelerating-data-workflows-with-query-agent-now-ga-yl8lcemzz) · Weaviate · 1 upvotes · 0 comments - [Weaviate 1.37 Release](https://daily.dev/posts/weaviate-1-37-release-nnexx0ui6) · Weaviate · 7 upvotes · 0 comments --- Tags: [#rag](https://daily.dev/tags/rag), [#vector-search](https://daily.dev/tags/vector-search), [#weaviate](https://daily.dev/tags/weaviate) [View this post on daily.dev](https://daily.dev/posts/scaling-test-time-compute-in-search-mode-ctvttd5ti)