Consider/Clarify RAG data as training data

[IchordeDionysos]

In the specification, the term training data is mentioned a couple of times.

However, especially for LLMs the vast majority of use-cases does not train LLMs on their own and rather enriches/improves the LLMs performance via various means, such as RAG (Retrieval-Augmented Generation).

With RAG, one would fetch additional data from various sources to enrich the prompt with additional context, improving response accuracy.

Given you might feed copyrighted material to a model and that model might output that data unaltered, RAG data should be held to the same standard as traditional training data.

---

I propose to specify that with training data, it also means RAG data.

e.g. for 2.12:

Before:
Check regulatory obligations to ensure compliance when handling and processing model training data.

After:
Check regulatory obligations to ensure compliance when handling and processing model training or RAG data.

Same for the other requirements.