Abstract Hybrid vector–relational queries require the simultaneous evaluation of relational predicates and high-dimensional vector similarity scores. Traditional database engines typically decouple these operations or maintain separate physical layouts, incurring separate memory access overheads. This paper presents an experimental evaluation of an alternative approach: inline metadata embedding, where relational metadata tokens are injected directly into the least significant bits (LSBs) of the floating-point vector coordinates themselves. We prototype this approach across two environments: a high-level Python macro-sweep framework to validate algorithmic correctness and an optimized native C++17 implementation accelerated with explicit AVX2 SIMD intrinsics to evaluate bare-metal performance. Under the evaluated workloads sweeping 315 configurations in Python, we observed 100.00% top-10 neighbor recall retention under up to 8-bit allocation budgets and an isolated global latency reduction of 55.19% within the high-level runtime environment for high-dimensional vectors (≥1024 dimensions). However, our native C++ micro-benchmarks—which isolate allocation and sorting overheads—reveal a more nuanced reality: under the evaluated workloads, the measured bit-extraction overhead in a bare-metal environment often counteracts the layout benefits, resulting in an observed throughput decrease of up to 10.52% at lower selectivities, while a modest performance gain (+9.09%) was observed only under specific intermediate selectivity conditions.
Nabh Sanjay Mehta (2026) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: