Joost v.

I use litellm, with the openinference instrumenter. Langfuse is more an example where they explain how they manage it.

So this is what I think most other providers do. E.g. weave or langfuse. It is explained like this in langfuse docs:

By default, base64 encoded data URIs are handled automatically by the Langfuse SDKs. They are extracted from the payloads commonly used in multi-modal LLMs, uploaded to Langfuse’s object storage, and linked to the trace.

&

File uniqueness determined by project, content type, and content SHA256 hash

https://langfuse.com/docs/observability/features/multi-modality Having something similar to this in Arize would make sense to me.