Semantic Caching and Model Routing Trade-offs

To cut cost, an engineer proposes two changes to a help-centre assistant: (1) a semantic cache that returns a stored answer whenever a new query's embedding has cosine similarity ≥ 0.85 with a previous query, and (2) routing every query to a small model first and only escalating when the small model says it is unsure.

1. What can go wrong with the 0.85 semantic cache, and how would you set the threshold and scope the cache safely?
2. What is the failure mode of "escalate when the small model says it is unsure", and what better routing signals exist?
3. How would you measure whether each change is a net win?