When decisions can be reformulated as a nearest-neighbor retrieval problem, the action space is large, candidates are relatively stable, and real-time requirements are high, replacing step-by-step generation with contrastive learning plus vector caching can yield order-of-magnitude latency improvements.