search system developers

Abstract: Rerankers decide which documents users and LLMs see, yet their standard metric, nDCG, relies on human relevance labels that are costly, sparse, noisy, and discretely graded. As rerankers approach each other in quality, nDCG on these labels therefore increasingly fails to separate them. LLM judges could supply dense labels. Relative judgments within one query tell even close candidates apart, yet their scores share no scale across queries. Absolute grades share one scale but are too coarse to distinguish documents of similar relevance. We propose Rubric-Calibrated Preferences (RCP), which combine both kinds of judgment. A listwise Bradley-Terry tournament orders each query's documents, and a rubric of yes/no criteria of increasing stringency provides an absolute standard. Item Response Theory (IRT), which scores test-takers based on their answers to common questions, then uses the shared criteria to put all queries' tournament scores on one scale. RCP's retrieval metric, RCP-nDCG, replaces nDCG's discrete labels with the resulting calibrated relevance probabilities. Against blind grades from 46 external annotators, calibration raises the correlation between a query's mean score and its mean human grade from 0.538 to 0.795. The probabilities rank a useful document above a non-useful one with probability 0.910 (AUC, chance 0.5), versus 0.651 for the benchmark labels. When the annotators' grades prefer one of two rerankers and exactly one metric agrees, that metric is RCP-nDCG in 72.4% of 185 comparisons (chance about 53%). On TREC-DL, RCP-nDCG sides with NIST assessors' grades on every reranker pair that these grades separate significantly. RCP-nDCG also resolves many of nDCG's ties and separates 1.9 times as many reranker pairs on NanoBEIR. Rubric calibration thus turns relative LLM judgments into dense relevance labels that are comparable across queries and agree with human judgment.

Abstract: Large language model (LLM) agents can propose, implement, and evaluate model changes. Autoresearch loops demonstrate this capability through minutes-scale iterations on a self-contained program. Online autoresearch instead spans asynchronous systems, hours-long variants, and weeks-long campaigns that can influence a product. A completed run can still support an invalid conclusion when a code change is a no-op, data windows leak, evaluator semantics drift, or the two arms traverse different serving funnels. We present EvoPilot, a human-gated method for long-horizon online autoresearch. Role-specific agents execute each round through a versioned domain skill and typed adapter. Durable records preserve experiments and failures; deterministic checks enforce recorded lessons. We study a 37-day campaign for the retrieval system that powers Video Deep Dive (VDD), an online experience for discovering follow-on videos after a user opens a seed video. The campaign covered seven directions and used an hourly refreshed index of hundreds of millions of videos. Earlier manual experiments had not established a benefit from an interaction head. A primitive autoresearch attempt revisited the direction but incorrectly attributed an offline hit-rate decline of 22 percentage points to the head. We then introduced EvoPilot. Its human-gated verification traced the drop to a pre-existing evaluation defect that produced output depths of 3,000 and 600. After repair, a matched comparison measured an offline improvement of 3.20 percentage points. Post-study replay and mutation tests rejected invalid comparisons while admitting valid counterparts. Durable state recovered an interrupted round, and artifact reuse avoided approximately five GPU-hours. Separately, a seven-day randomized online evaluation estimated a 0.66% relative increase in the VDD slice of Good Search Result Rate for Retention (GSRR).

Abstract: LLM serving reuses KV cache by exact prefix match, so when a prompt is assembled from a set of reusable pieces -- retrieved passages, tool definitions, few-shot exemplars -- the order chosen for those pieces determines how much computation can be shared. Every deployed system fixes that order by a single global convention. We prove this is optimal only when requests contain at most two pieces, and asymptotically wrong in general. Our main result is a structure theorem: the minimum prefix-trie cost equals min_H sum_x w(x) t_x(H) over binary hierarchies H on the requests, where t_x(H) is the canonical decomposition size of the set of requests needing chunk x. Choosing chunk orders is therefore equivalent to choosing one hierarchy over requests. The identity yields an O(3^m) exact algorithm, identifies the two-chunk case as minimum vertex cover, and shows that on the leave-one-out family the optimum is the minimum external path length of a binary tree -- the merge-sort recursion -- so a global order pays Theta(n^2) against a true cost of Theta(n log n). Agglomerative clustering by common intersection is a tight 1/2-approximation for the achievable saving. On BM25 retrieval traces over three BEIR corpora the resulting layout reduces prefill by 17-36% against production RAG ordering, and the margin widens with retrieval depth as the theory predicts. Serving requests in the hierarchy's DFS order finally lets a cache holding one request's context attain the unbounded-cache optimum exactly, so cache capacity and reorder window act as substitutes.