The useful choice in NoveltyScope: search for counterevidence before scoring novelty. The authors retrieve prior work with BM25 and dense search, then request paired passages for potential refutation. arxiv.org/abs/2610.22703?utm…

We take real queries into a search vendor call. A leaderboard score cannot tell us what fails on our corpus, what the bill looks like at our traffic shape, or how hard it is to leave.

We let some clinical-note queries take up to 30 seconds. That was a product choice tied to the work, not a missed latency target. Trio Health needed answers to questions such as why a patient stopped a medication. That reason can live in free-text notes rather than structured fields. We built a two-stage indexing path—extraction, then embedding—and retrieval that combines BM25, semantic search, and reranking. An agent harness uses the search API to produce structured results. The deployed index holds 500M chunks from 37M notes, stored in S3 via Turbopuffer. The early usage corrected a tidy "hybrid search" story. Both application and agent searches leaned mostly on lexical retrieval. Vector similarity helped more with query expansion than recall. Agent runs generated several orders of magnitude more search traffic than the early human beta. Those are observations from this system, not a claim about every clinical corpus. We chose the latency budget with those query shapes and costs in mind. Measure the actual workload before optimizing the wrong path.

We would test skill rerankers on near-duplicate instructions. SkillContrast's authors keep differing text and local context before reranking; query-only selection can erase what distinguishes two skills. arxiv.org/abs/2610.11650?utm…

Most search tools never say no: top-k comes back even when the index has nothing, so the agent reads noise as a hit. The authors test what frozen agents do when the tool refuses, across 257 NQ and 300 HotpotQA questions in a 21M-passage BM25 index. arxiv.org/abs/2610.05348