Gemini Embeddings: Breakthrough or Just Buzz?

Key Takeaways

Google claims to have released the best embedding model to date. Let's break down the most recent AI headline, the release of Gemini embedding 001. First, it's quite promising. It ranks first on the massive text embedding benchmark before OpenAI and Coher embedding models. Based on the benchmark, Gemini is better than both, but I suspect the results will actually be quite similar. What's really cool is that it seems to be working much better in other languages and with code. Some other advantages is that it can premp compute and cache embeddings for much more efficient use and you can also batch embeddings so that if you have a lot of data to embed in your rack system, it will be two times cheaper. OpenAI's small embedding model is still the cheapest to date and in fact Gemini is the most expensive of all three companies, but it's still quite comparable. What's cool is that we've been using Gemini models in our courses mostly because of the free API tier and this model will be on the tier and we and our students will be able to use this model as well in the free tier. So that's quite cool. The main downside is that it's not multimodel. It doesn't process images, videos or audio. So the verdict, I think it's not hype, it's pure value. If you work with other languages than English or with code, try it out in your right system. I'm Luis Franis, co-founder and CTO of TORZI, and follow me for tomorrow's no BS roundup.