How are vectors stored in a database?

Vectors are stored in databases using specialized data structures and storage formats optimized for high-dimensional numerical data. In most cases, vectors—represented as arrays of floating-point numbers or integers—are stored in dedicated columns or fields designed to handle large numerical sequences. For example, PostgreSQL’s vector extension introduces a vector column type that directly stores arrays like [0.25, -0.12, 0.98]. NoSQL databases like MongoDB use BSON (Binary JSON) formats to serialize vector arrays into documents. These storage methods prioritize efficient serialization and deserialization to minimize overhead during read/write operations.

To enable fast similarity searches (e.g., finding vectors closest to a query), databases use indexing techniques tailored for vector data. Common approaches include hierarchical navigable small worlds (HNSW), inverted file (IVF) indexes, or tree-based structures like KD-trees. For instance, Milvus, a vector database, uses HNSW to create graph-like layers that allow approximate nearest neighbor (ANN) searches with sublinear time complexity. Indexes are often built during data insertion and stored separately from raw vectors to optimize search performance. Some systems also partition data into shards or clusters (e.g., using k-means in IVF) to narrow down search ranges, reducing computational overhead.

Optimizations like quantization and compression further enhance storage efficiency. Quantization reduces vector precision (e.g., converting 32-bit floats to 8-bit integers) to save space and accelerate computations. For example, Facebook’s FAISS library uses product quantization to split vectors into subvectors and compress them using codebooks. Databases may also employ memory-mapped files or GPU-accelerated storage for large datasets. Practical implementations often combine these techniques: a database might store raw vectors in a compressed binary format, maintain an HNSW index in memory, and use quantization for disk storage. This layered approach balances speed, accuracy, and resource usage for scalable vector management.