Databricks thankfully continues to de-hype long-context RAG with new experiments.

Key observations:

- at shorter* context lengths (<128k), o1-preview outperforms others.

- beyond 128k, only gemini-1.5pro/flash continue to give correct answers. other models give wrong/empty answers.

- gemini-1.5pro isn't as good as o1-preview for shorter contexts. there's room for improvement here.

gemini-1.5pro is the only true! long-context RAG model out there, although with less accuracy.

article: