Embed the world: Multimodal AI for searchable aerial imagery at scale
TL;DR
AWS outlines a search architecture for Vexcels aerial imagery: tiles are embedded via Amazon Bedrock, indexed in Amazon OpenSearch Serverless, and queried with natural language. The benchmark used OpenStreetMap as ground truth for Grant Park in Chicago and compared about 100 configurations across two query types: swimming pools as discrete objects and roads as distributed infrastructure.
Nauti's Take
The post is clearly AWS- and Vexcel-friendly, but the useful part sits in the uncomfortable details. Multimodal search does not work here because a big model magically understands geography.
It works because the team defined what counts as a hit, tested K values, checked which image views were worth paying for, and measured captions against visual embeddings. That is where demo AI starts becoming production AI.
Briefingshow
This is more than an AWS reference story because it breaks a real search problem into measurable choices: model selection, image views, captioning, K values, and evaluation metrics all change result quality. For teams working with satellite, drone, or inspection imagery, the lesson is blunt: build the evaluation harness before swapping models. Otherwise, tuning becomes guesswork.
Sources
