How Nearmap built continental-scale aerial search using Amazon S3 Vectors

This article describes how Nearmap built Zero-shot Search, a semantic search technology for its 100+ petabyte aerial imagery archive, using Amazon S3 Vectors to enable text, image, and location-based queries without training new detectors.

• Converts aerial images and search terms into vector embeddings for semantic similarity matching across 1+ billion vectors from 107 surveys
• Uses S3 Vectors for cost-efficient, serverless vector storage with native similarity search and metadata filtering capabilities
• Implements separate embedding specifications for map tiles, top-down property imagery, and oblique property imagery at different scales
• Employs hash sharding across multiple S3 Vectors indexes to enable parallel ingestion and retrieval at scale
• Runs batch ingestion on Amazon EKS with GPU nodes for embedding generation and CPU nodes for vector loading, using Spot Instances for cost efficiency
• Supports heatmap queries via deterministic keys for direct tile retrieval and property searches combining multiple view types
• Stores filterable metadata (survey ID, date, address) and non-filterable metadata (geometry, image references) to support diverse query patterns

S3 Vectors enables Nearmap to retain a growing corpus of historical embeddings without managing search infrastructure, allowing customers to find aerial features using natural language descriptions.