Bypassing inference bottlenecks: Accelerating complex AI search with Retrieve-for-Train · Bharat Hunt