Most of the Volume Never Reaches a Model
The LLM is the most expensive stage in a matching pipeline, not the interface to it. I built a 3-gate funnel where 95%+ of jobs are filtered by SQL and vector similarity before any model call. Here are the real numbers.