Bounded Semantic Interpretation for Reliable Hybrid Product Search

Product search must reconcile noisy natural-language queries with structured catalogs, exact compatibility constraints, rapidly changing inventory, biased behavioral signals, and strict serving budgets. This systems perspective proposes Bounded Semantic Interpretation (BSI), a hybrid search architecture in which generative models may propose typed intent and attribute hypotheses but cannot directly create catalog facts or bypass deterministic validation. The design combines lexical, dense, learned-sparse, multimodal, and graph-based candidate generation with provenance-preserving fusion and staged re-ranking. It also defines explicit fallback behavior, constraint checks, observability requirements, and release gates. The contribution is architectural and methodological rather than empirical: no production performance claims are made. A reproducible public-benchmark protocol is specified using the Shopping Queries Dataset and the Shopping Queries Image Dataset, together with latency, constraint-violation, robustness, calibration, and subgroup-reliability measures. The framework is intended to help engineering teams test whether semantic improvements remain reliable under operational and governance constraints, and to make failure modes visible before online deployment.