Azure Cosmos DB Vector Indexing

Category: Multi-model & Hybrid Databases
Brand: Microsoft Azure
Slug: azure-cosmos-db-vector-indexing

Overview

Azure Cosmos DB Vector Indexing adds native vector storage and search to Azure Cosmos DB for NoSQL. It lets you store multi-modal, high-dimensional vector embeddings directly in documents alongside other JSON data and perform efficient vector similarity search at scale using built-in vector index types.

Features

Native vector storage

• Store vector embeddings as properties within the same documents that hold your application data.
• Supports multi-modal, high-dimensional vectors (e.g., text, images, audio embeddings).
• Co-locates vectors and associated data in a single logical unit for simpler data management and query patterns.

Vector index types

• Flat index (exact kNN / brute-force):

  • Performs exact k-nearest neighbors search.
  • Provides 100% recall.
  • Suited for smaller or more focused vector search scenarios.
  • Works well when combined with query filters and partition keys.

• Quantized flat index:

  • Compresses vectors using DiskANN-based quantization methods.
  • Enables more efficient brute-force kNN search over large datasets by reducing memory and storage footprint.

• DiskANN index:

  • Uses Microsoft Research’s DiskANN algorithms for approximate nearest neighbor (ANN) search.
  • Designed for efficient, high-accuracy multi-modal vector search at large scale.
  • Optimized for low latency and cost-efficient operation on large vector collections.

Integrated querying

• Vector search is exposed through the VectorDistance function in Azure Cosmos DB queries (NoSQL API).
• Vector similarity search can be combined with other query filters using WHERE clauses.
• Works with existing Cosmos DB indexing and partitioning, enabling:
  • Hybrid queries that filter on document fields and then apply vector similarity.
  • Scenario-specific filtering (e.g., by user, region, time range) plus vector search.

Performance and scale characteristics

• Designed for low-latency and high-throughput vector similarity search directly within Cosmos DB collections.
• Supports both exact brute-force search and approximate nearest neighbor search, so you can trade off accuracy vs performance and cost.
• Architecture supports scalable handling of large, high-dimensional vector datasets used in AI applications.