Certified reuse, execution choices, and repeatability
· 8 updates Coverage: September 25, 2026–October 24, 2026
This issue covers selected developments from September 25 through October 24, 2026. The research raises three distinct questions: which previous decisions can be reused, which computation should implement an operator, and how reliably repeated runs reproduce an answer. Product releases address narrower parts of these questions through exact-call caching and extraction improvements.
Research summaries are based on publication records and abstracts; product summaries are based on official announcements. Reported performance is attributed to the source. These checks do not constitute full-paper review, implementation inspection or independent reproduction.
Research
Research
SemReuse: auditing reused filter results
SemReuse proposes storing semantic-filter outputs as materialized views and using inferred predicate relationships to provisionally accept or skip tuples. A pooled stratified audit checks the resulting answer and triggers additional evaluations when needed.
Its contribution is a certification procedure around fallible reuse decisions. The abstract reports up to 15.4× fewer evaluations in its experiments; this does not establish savings for every workload.
SemViews uses prior query labels, embeddings and helper scores to order tuples, then certifies the answer with fresh stratified oracle samples. The preprint reports 28–45% fewer oracle calls across four workloads.
Both SemViews and SemReuse study certified reuse, but their use of prior state and auditing procedures differ. Their reported savings should not be read as a head-to-head comparison.
NousDB’s October 21 revision describes a DataFusion-based engine with semantic algebra, token histograms and batching under context-window constraints. The first version appeared October 20.
Token volume and batching fidelity become planning concerns alongside relational rewrites. The profile retains the paper’s claims as author-reported results.
JEVDB uses typed decision models for filters, joins, classification and ranking, with selective escalation to generative models. Its join execution combines relational reduction with semantic candidate screening.
This explores a different execution choice from generating text for every decision. The reported evaluations use SemBench and the join-oriented Shelob workload.
Same Query, Different Answer studies repeated executions of semantic workloads at temperature zero. It distinguishes changes to individual operator decisions, complete answers and aggregate quality.
A stable average score need not mean a stable answer. Repeated execution and item-level reporting are relevant when comparing engines or validating a cached result.
The PACMMOD publication develops intent-specific task DAGs and task-extended relational algebra for semantic operators. It is associated with SIGMOD 2027; the preprint was already available in June 2026.
The newly added profile covers a previously cataloged source. Planning includes task decomposition and routing, rather than treating each semantic call as a single fixed implementation.
Databricks’ October 2 release notes describe a new standard-mode ai_extract checkpoint, with a vendor-reported 2.3× speed improvement and higher quality at unchanged cost. ai_parse_document also returns chart data as structured JSON.
These are changes to extraction functions. The announcement does not establish an end-to-end semantic-query optimization result.
These sources extend the site’s coverage but precede this issue’s reporting window. They are not counted among its new developments.
Research · Adjacent
IMLane: separating inference and database scheduling
IMLane uses process-level AI-function execution and separates inference scheduling from the database scheduler. Its publication and abstract describe integrations with OceanBase and DuckDB.
New to this catalog, but earlier than this issue’s reporting window. It concerns execution of AI functions, including predictive models, rather than a new semantic-operator language.
Featurized-Decomposition Join: pruning through features
FDJ extracts features and combines inexpensive comparisons to prune semantic-join candidates, with statistical quality guarantees described in the preprint.
This newly profiled paper was already in the bibliography. Its verified preprint date is December 2025, so it is background reading rather than a new October result.
Snowflake adds evaluation and configuration search
Two public previews support measuring AI-function quality, cost and consistency, and searching prompt/model configurations against a common dataset and metric. A selected configuration can become a reusable SQL AI function.
These September 21 announcements precede the reporting window. They supply context for function-level optimization; they do not establish joint optimization of an entire relational plan.
Google’s September 1 announcement describes native GQL graph traversal alongside SQL, with access to BigQuery ML and AI functions in the same query.
This is older, adjacent infrastructure for combining graph, relational and model-based computation. It is not evidence of a new semantic-operator optimization technique.
sqpe