Cache points attached to individual content parts of user or assistant messages are now correctly forwarded to Amazon Bedrock Converse when using @ai-sdk/amazon-bedrock. Previously, only message-level cache points (set via providerOptions.bedrock.cachePoint) were preserved; part-level cache points were silently dropped, meaning Bedrock never created or read prompt cache entries for those parts. Requests still succeeded but prompt caching was effectively disabled for content-part-level annotations. This is now fixed — cache points are emitted immediately after the content part that specifies them.
- #14809 — fix(bedrock): part-level cache points