#Firrst of all, I'd say that if it meets
opal glacier
Thank you for the detailed answer. My main area of focus in asking that question was to see what fundamental natural language, semantic parsing and text generation advantages might be afforded by Vertex AI over the Gemini API.
For example, like if there were some truly powerful features Vertex AI had like "fine tuning" or "built-in advanced RAG" that Gemini AI did not posess. But your answer, if I read it right, indicates that the advantages of Vertex AI are mostly privacy and execution environment customization based, not fundamental NLP tech. Is that about right?
feral basalt
Yes and no.
Vertex AI, outside of Gemini, offers all of those tools. It has other embedding models. It has vector stores. It has embedding models for images (!). It has a built-in advanced RAG system that is tied into an advanced chatbot system.
Some of them use Gemini, some don't.
Plus it has piles of other models. Want a managed Claude instance? Sure. Want to fine-tune Gemma? Sure.
Vertex AI is a massive suite of AI tools, built on top of a massive suite of cloud tools. AI Studio has a very narrow focus.
And once you get past some of the simple elements of AI Studio, you start having the authentication complexity that Vertex AI has. So if you do AQA in AI Studio - you have to use OAuth. Fine tuning? OAuth. And no automated pipelines to do that fine tuning.
opal glacier
Thanks.
- What's AQA?
- Do you have a link to a good doc on this: "It has a built-in advanced RAG system that is tied into an advanced chatbot system."
That last RAG video I saw was from the Amazon Bedrock folk and it seemed like a pretty simple idea. Convert your source texts to a vector data base. Then, every time you do a query, embed encode the query, run it by the vector data base, take the top N matches, and feed those to the LLM in the prompt. So now I'm wondering if there are more advanced or esoteric RAG strategies than that? Obviously, to jump into Vertex/Rag like you just mentioned, I'd expect it to do a lot more for me than the simple RAG strategy I just described?
So the Anthropic models (Clause) are available via Vertex? I wouldn't have thought that Google would offer a competitor's models (in the LLM space).
feral basalt
Yes, Claude is available in Vertex AI.
In some ways, it is a competitor, yes. But in others it compliments it. The goal with Vertex AI is to provide a platform people will use AI on. Not just Google's AI.
(And at the moment, Claude is even cheaper than Google is for the base models.)
Gemini is good because it means that they're not subject to the whims of another company... but having more gives their customers one-stop-offers.
See https://ai.google.dev/docs/semantic_retriever for more on AQA. Attributed Quation and Answering
RAG, fundamentally, is pretty straightforward. It doens't even require a vector database, tho that's certainly one implementation.
Google Cloud Vertex AI Search and Conversation is the advanced RAG system.
https://cloud.google.com/vertex-ai-search-and-conversation?hl=en
Here's an article talking about it as well (tho it is a bit sales-y)
https://cloud.google.com/blog/products/ai-machine-learning/rags-powered-by-google-search-technology-part-1
