I send several requests to the DeepSeek API with the same long system prompt and shared reference material, followed by a different user question each time. I’m wondering how reliably the API can reuse that repeated prefix rather than process it from scratch.

For anyone who has tested DeepSeek API prompt caching, does the shared text need to match exactly, and is there a way to tell from usage data when a cache hit occurred? I’d like to know what to measure before changing how I structure requests.