Semantic queries stop throwing away old labels: reusing past answers as views cuts model calls by 28% to 45%

The cost of a semantic filter can come down while precision and recall targets still hold with a stated probability. Earlier methods forced a choice: cascades that trust a cheap proxy miss their targets on 30.8% to 65.4% of queries, while methods with finite-sample guarantees save only 2% to 10% of calls when the proxy is weak. Replayed on four workloads against 725,000 recorded oracle labels, SemViews meets both targets on 99.63% of queries, while the only prior method with the same guarantee needs 1.32 to 1.71 times as many calls.

Labels left behind by earlier queries become semantic views that order tuples for a new predicate instead of being served as the answer. Similarity caches and entailment-based reuse treat the relation between predicates as crisp and return a stored answer on a hit. Across four predicate pools, 56% of the implications an LLM judge asserts are violated by more than 10% of the antecedent's positives; similarity caches miss a target in 100% of hits, and SemWeave misses on 6.2% to 67.8% of queries.

Ordering and certification are separated: free signals decide where to look, and only fresh oracle labels enter a confidence bound. Per-query optimizers estimate threshold quality from samples of the whole relation and pay for every label of the current predicate. Certification draws uniformly without replacement from regions fixed before the draws and uses exact hypergeometric bounds; SemViews misses on at most 0.7% of queries per dataset, while every cheaper method misses on at least 32.4%.

When the same question returns in different words, the old view still removes calls. A reworded predicate agrees with its original on 76% to 99.8% of tuples, but the overlap of the two positive sets is only 0.57 to 0.95. On the reworded workload SemViews needs 47% to 53% of the calls of exact evaluation, 27% to 36% fewer than without views, and misses on at most 0.2% of queries, where similarity caches and SemWeave miss on 53% to 83%.

The study covers semantic filters over text; joins, ranking, and multimodal data are out of scope. The main relations hold 5,000 tuples, with 100,000-tuple real relations used to check the trend, and each workload issues all predicates of a pool in a random order, two of them generated from taxonomies. A reader can watch next for certified reuse extended to semantic joins and ranking, and for measurements of how often real analyst sessions revisit their questions.

The guarantee is per query, so a workload's miss rate is bounded by a concentration argument rather than fixed query by query. Certification carries a floor tied to a predicate's selectivity and not to the relation size: predicates with at most 2% selectivity cost 80% of exact evaluation. How much views help depends on how related a workload's predicates are, so the taxonomy-generated GoEmotions pool may be an upper end while the human-written PubMed pool gains little. Two datasets used to fix hyperparameters are not fully held out.