By name: set-latency-budget description: "Set a p95 latency budget, decompose per-stage latency, and tune relevance within it. Reach for this on a speed/capacity question."
Skill: Set latency budget
Don't chase NDCG into a timeout (§3 #4).
Step 1 — Set the p95 budget
A target the search must hold via search_relevance_engineering_calc.py latency-budget (§3 #4).
Step 2 — Decompose per stage
Query, fetch, rescore — find the slow stage (§3 #4).
Step 3 — Size the index
Shards/replicas/storage via search_relevance_engineering_calc.py index-sizing (§3 #7).
Step 4 — Tune relevance within budget
Buy relevance with rescoring only inside the headroom (§3 #4 #5).
Output
A p95 latency budget with per-stage decomposition and the relevance headroom. Traverse Tree 3 in the decision-trees file.