Cross-modal necessary conditions for Semantic Bloom Filters

Identify which cross-modal necessary conditions for predicates over text, images, and other modalities can be extracted cheaply, reused safely, and incorporated into Semantic Bloom Filter candidate pruning.

Background

The paper proposes extending JEVDB’s typed decision interface beyond text to multimodal predicates, such as matching textual reports with inspection images, while retaining non-generative decision heads. Such an extension would require candidate-reduction mechanisms that operate across heterogeneous modalities.

The authors explicitly identify an unresolved question for Semantic Bloom Filters: determining which cross-modal conditions are sufficiently inexpensive to extract, reliable enough to reuse, and safe to apply for pruning without removing true semantic matches.

References

This also raises new questions for SBFs: which cross-modal necessary conditions can be extracted cheaply, reused safely, and pushed into candidate pruning?

— Prune First, Decide Fast: Scalable Semantic Query Processing with JEVDB  (2610.02046 - Wang et al., 1 Oct 2026) in Section 6, paragraph “Future Work,” bullet “Multi-modal semantic queries”