Быстрый тест по кэшированию

Ответы и пояснения появляются по мере прохождения теста.

1. What can a provider-managed prompt cache let a later request avoid?
   • Recomputing some work for an eligible repeated prompt prefix
   • Running the model on any new user content
   • Checking provider-specific cache eligibility
   • Applying attention to the uncached part of a request
2. Why can a small prompt-prefix change affect cache eligibility?
   • Every provider guarantees cache reuse regardless of prompt changes
   • Provider-specific matching rules or breakpoints may make the changed prefix ineligible for reuse
   • Causal attention prevents any prefix from being reused
   • Changing later content always invalidates every earlier prefix
3. Какой сценарий БОЛЬШЕ всего выигрывает от оперативного кэширования?
   • Один разовый вопрос с короткой подсказкой
   • Подсказка, в которой каждое слово каждый раз меняется
   • Множество запросов, которые имеют одно и то же длинное и одинаковое системное приглашение или документ.
   • Генерация случайных чисел без ввода
4. How should a team estimate the price effect of a prompt-cache hit?
   • Assume every provider discounts cached tokens by 90%
   • Assume cache hits are always free
   • Use the number of prompt characters alone
   • Use the current provider and model pricing for cached versus uncached tokens
5. How long does a prompt cache remain available?
   • It depends on the provider, model, and cache-retention settings
   • Exactly one hour for every provider
   • Forever unless a user deletes it
   • Exactly five minutes in every API
6. How can an OpenAI API client identify how many input tokens were served from cache?
   • The number of characters in the system message
   • The response usage field input_tokens_details.cached_tokens
   • The cache key length
   • Количество параметров модели
7. In OpenAI implicit prompt-caching mode, where is a breakpoint placed by default?
   • It caches every prompt regardless of length
   • It disables prefix matching
   • It places a breakpoint at the end of the latest eligible message
   • It stores all user prompts permanently
8. What should a team check before caching sensitive prompt content?
   • Only the color of the API dashboard
   • Whether the prompt contains an odd number of tokens
   • Whether another provider offers the same discount
   • The provider’s retention, data-use, and cache-control terms for the selected model