[Feat]: Batch Prediction and Cached Prompt should work together

[LinusOstlund]

Is your feature request related to a problem? Please describe.

Batch predictions should be able to use a cached context.

Describe the solution you'd like

We are currently working on a classifying prompt which requires an extensive system prompt. We have been experimenting using batch processing together with context caching (leveraging the 50% discount from batch processing, with the 75% discount from context caching).

We have tried several approaches now, but the batch job fails. Here's and entry from the responding predictions.jsonl:

{"status":"Internal error occurred. Failed to get generateContentResponse: {\"error\": {\"code\": 404, \"message\": \"Not found: cached content metadata for 3814716010150232064.\", \"status\": \"NOT_FOUND\"}}"

• There is no official Google material (tutorials, documentation, model cards, etc) stating if batch predicitons and context caching is supported when used together.
• If it is supported, it would be lovely to see a tutorial.
• If it is NOT supported, please add a disclaimer to the batch / context cache docs.

Describe alternatives you've considered

It is supported on OpenAI.

Additional context

No response

Code of Conduct

• I agree to follow this project's Code of Conduct