vertexai: Add context caching to VertexAI class

[kardiff18]

PR Description

• Add support for context caching to VertexAI class (currently it is only supported by ChatVertexAI)
• Refactor is_gemini_advanced into utils (to be used by both VertexAI and ChatVertexAI)

Type

🆕 New Feature
🧹 Refactoring
✅ Test

[lkuligin]

I'm sorry, but how would it be different from

langchain-google/libs/vertexai/langchain_google_vertexai/chat_models.py

Line 1045 in 8c569ed

cached_content: Optional[str] = None

?

[kardiff18]

The issue today is that you cannot use VertexAI when using context caching, you can only use ChatVertexAI. If you try to use VertexAI with the cache it does not work.

So, the former PR that you linked it adds the parameter to the ChatVertexAI class, whereas this PR is adding it to VertexAI and therefore had to add it to the base model. Does that help?

[lkuligin]

The issue today is that you cannot use VertexAI when using context caching, you can only use ChatVertexAI. If you try to use VertexAI with the cache it does not work.

So, the former PR that you linked it adds the parameter to the ChatVertexAI class, whereas this PR is adding it to VertexAI and therefore had to add it to the base model. Does that help?

I'm wondering whether we should keep VertexAI at all. It's pretty outdated already and the gap will be increasing. Why do you need this class instead of ChatVertexAI?

[kardiff18]

That's valid! My customer had attempted to use VertexAI at first since there was no multi-turn for the use case. After realizing context caching wasn't being properly used (there was no error message, we just knew the cache wasn't being invoked since the LLM response clearly was not using the cached instructions), I realized the issue was that they used VertexAI and not ChatVertexAI. I got them to switch over easily to ChatVertexAI after finding the issue.

That being said, I didn't know if anyone else was trying to use VertexAI without knowing it wasn't working properly due to a lack of error message, so I implemented it. We don't need this merge request, but I was just trying to help out others. We also mainly use ChatVertexAI for everything else. So in sum, I don't really think VertexAI is needed as well, so you're welcome to just ignore this PR if you think you'll be deprecating the class entirely.