Prompt caching is one of the most testable ways to cut cost and latency. These notes cover what to cache, how checkpoints and the TTL behave, why a cache miss happens and how the token counts in the response add up.
Prompt caching reuses a long, repeated prompt prefix to cut latency and input token cost, so put static content first and dynamic content last. Cache checkpoints need a model-specific minimum (512 to 4,096 tokens) and the default 5-minute TTL resets on every hit.
Official source: docs.aws.amazon.com/bedrock/latest/userguide/prompt-caching.html
Cache checkpoints are processed in the order tools, system, messages, and changing an earlier section invalidates the later ones. Any change to the cached prefix is a cache miss; check cacheReadInputTokens and cacheWriteInputTokens in the response.
Official source: docs.aws.amazon.com/bedrock/latest/userguide/prompt-caching.html
With caching on, inputTokens only counts non-cached tokens. Total input tokens = inputTokens + cacheReadInputTokens + cacheWriteInputTokens. Cached reads are billed at the cache-read rate, and prompt caching is not supported with batch inference.
Official source: docs.aws.amazon.com/bedrock/latest/userguide/prompt-caching.html
Answer practice questions, then review your mistakes with spaced repetition.