This page describes AlloyDB AI vector search strategies and explains when to use each strategy. By default, AlloyDB for PostgreSQL uses k-nearest neighbors search (KNN) to find vectors that are similar to a query. Vector indexes implement a search strategy called Approximate Nearest Neighbor (ANN). When you create a vector index, AlloyDB AI uses ANN, which provides better performance than KNN. Keep in mind that, when you select a vector index, you need to balance query latency and recall.
Recall measures how effectively a search retrieves all relevant items for a given query. For example, imagine you have 100 embeddings, each one representing an entity in your database. You query your embeddings with a target vector and limit it to 10 results. A KNN vector search finds the 10 exact closest vectors using a brute force calculation method, which results in 100% recall. AlloyDB AI uses this method by default if no vector search index is created or chosen. When you create a vector index in AlloyDB, it typically uses ANN, which might partition vectors according to similarity to facilitate faster retrieval. As a result, using ANN, the 10 vectors returned in the earlier example might not be exactly the 10 vectors that are closest in distance. If 9 out of the 10 retrieved vectors are the closest in space to your query vector, then your recall is 90%. For more information, see Measure vector query recall .
Query latency defines how fast the search results are generated. For example, latency is calculated based on the time spent on a search to return the vectors after you submit a query.
Choose your search strategy
When you perform vector search in AlloyDB, choose one the following search strategies:
An algorithm that finds the k-nearest neighbors data points to a given query data point. When you perform a vector search without creating an index, a KNN search is performed by default.
To further improve the performance of KNN search, add your embedding column, and other columns related to your query, to the column store in the columnar engine . You can add the columns manually or add the columns using auto-columnarization .
- Your application is very sensitive to accuracy and you need the exact closest matches.
- You have tens of thousands of vectors
- Your application requires low latency.
- You have more than 100,000 vectors.
The two common ANN-based vector index types for AlloyDB are ScaNN and HNSW. Use the following guidelines to decide which index type would be best for your use case:
Usually well suited for low-dimensional data or massive datasets that exceed memory.
Scales well to 10B vectors.
Offers several benefits over HNSW on standard PostgreSQL:
- Up to 6 times faster vector queries and up to 10 times faster filtered vector search queries
- Up to 16 times faster index creation
- Up to 24 times higher write throughput
- Uses up to 4 times less memory
Usually well suited for higher-dimensional data that fits largely within the in-memory cache.
Scales well to 10-20 million vectors.
Google recommends that you create a vector index to optimize performance on your vector search queries. For more information about how the ANN index is used for similarity searches, see Create indexes using ScaNN .
To accelerate your filtered KNN search, use the columnar engine .
