Your team is split on how to cut grounding failures: on...

Before choosing, you must measure what fraction of queries fail grounding and why. Widening the beam only helps if failures are due to insufficient candidate diversity; if the edge is missing, the beam width is irrelevant. A wider beam multiplies downstream constrained-BFS traversals but adds no LLM cost. A GNN-based retriever shifts the cost to training and embedding maintenance and runs at fixed latency, but may generalize poorly. A staff-level approach measures both and considers if they can be combined within the 300ms latency budget.