At Next ‘23, we launched AlloyDB AI, an integrated set of capabilities built into AlloyDB for building generative AI applications. One of those capabilities allows you to call a Vertex AI model directly from the database using SQL.
AlloyDB is a fully managed PostgreSQL-compatible database that offers superior performance, availability and scale. In our performance tests, AlloyDB delivers up to 100X faster analytical queries than standard PostgreSQL, and AlloyDB AI runs vector queries up to 10x faster compared to standard PostgreSQL when using the IVFFlat index. It also provides enhanced vector search and predictive machine learning (ML) capabilities.
With Vertex AI, Google’s end-to-end AI platform, you can upload and label your data and train and deploy your own ML models. You can also utilize Google, third-party, and open-source AI models through Model Garden on Vertex AI.
You can enable AlloyDB access to Vertex AI with AlloyDB AI using the google_ml_integration extension, which allows you to run predictions with your data in AlloyDB using custom models in Vertex AI or models from Model Garden. AlloyDB AI also integrates open-source tools like pgvector and LangChain, allowing you to use AlloyDB as a vector store to store embeddings and connect to your LangChain applications.
By combining Google Cloud products and open-source AI tools, AlloyDB AI enables you to enhance your applications by creating new user experiences using live data. In other words, you can create dynamic AI experiences that change in real time according to changes in your database.
In this post, we’ll explore five examples of using SQL to access models in Vertex AI or your own custom models for similarity search, sentiment analysis, bot detection, healthcare predictions, and risk prediction.
Similarity search with vector embeddings
Let’s say you own a store called South Bay Furnishers and you store your inventory information in a relational database. You might have a table called products to store information, such as product descriptions, inventory-related information, and more.
Traditionally, your workflow might look like this if you want to generate embeddings for your product descriptions.
