Search interfaces stop being hand-fixed: SearchWeaver rewrites tools and instructions per corpus, lifting answer accuracy from 71.6% to 78.4%

The search interface becomes a resource that can be generated automatically for a corpus and reused, rather than a one-off hand-fixed design. Retrieval-augmented generation, direct corpus interaction and approaches that pull retrieved documents into a workspace all specify the interface in advance and apply it unchanged across corpora, even though the most effective design depends on corpus size and structure, query types and the language model used. Starting from direct corpus interaction, SearchWeaver raises BrowseComp-Plus answer accuracy from 71.6% to 78.4% and average nDCG@10 across four BRIGHT subsets from 45.7 to 55.0, outperforming direct corpus interaction on both benchmarks and all four subsets.

Optimization changes three things at once: the corpus artifacts prepared before queries, the operations exposed as tools, and the decisions left to the search agent. Prompt optimizers revise text only, and another line of work revises a model's outer harness code, leaving corpus artifacts and tools untouched. A coding agent acts as proposer, reading the corpus, previous candidates, and the per-query rewards and full trajectories from each execution to write a new interface; one candidate is evaluated per iteration for ten iterations, and the candidate with the highest mean reward on optimization queries is selected, with test queries kept out of proposer feedback.

The discovered interfaces use different strategies across corpora and remain usable with another language model. On BRIGHT, Biology restricts retrieval to the question's topic directory, Earth Science introduces retrieval only when building the final ranking, and Economics and Robotics incorporate agent judgments through document ordering and relevance grades respectively. Interfaces optimized with GPT-5.4-nano, handed to Gemini 3.5 Flash-Lite without re-optimization, still exceed direct corpus interaction on BrowseComp-Plus accuracy and average BRIGHT nDCG@10, and mean tool execution time on BRIGHT also falls.

Experiments target fixed corpora: BrowseComp-Plus uses 30 queries for interface optimization and 800 for testing, the four BRIGHT subsets total 213 optimization and 210 test queries, the search agent is GPT-5.4-nano and the proposer is Claude Opus 5.5, with each corpus starting from the default direct corpus interaction interface, one candidate evaluated per iteration for ten iterations. This scope means the conclusions apply to the tested fixed corpora and model combinations.

As documents are added or revised, corpus artifacts and the tools that use them may need updating; when such updates become necessary, and whether they can be done incrementally without repeating the full optimization, is not studied.