Vector-Symbolic Hybrid Systems Bypass Single-Vector Embedding Limitations on Structured Queries

This technical report demonstrates how Vector Symbolic Architecture (VSA) hybrid systems bypass the theoretical limitations of single-vector embeddings established by Weller et al. (2025).The LIMIT dataset serves as a mathematical stress test where single-vector approaches fail dueto sign-rank constraints, achieving less than 20% recall@100 despite the simplicity of the queries.In contrast, our VSA hybrid approach achieves perfect performance by operating in discrete symbolic space rather than continuous vector space. Through role-filler binding operations that enable exact symbolic matching, the system achieves 100% Recall@5/10 and MRR=1.0 while maintaining constant-time retrieval independent of corpus size. We provide detailed latency measurements demonstrating sub-millisecond performance on CPU and show that the approach scales to 50,000 documents without degradation. These results suggest that compositional embedding methods, largely overlooked in recent years, deserve renewed attention as practical solutions to fundamental limitations in embedding-based retrieval systems.