Amazon Nova Embeddings Advance Multimodal Search in Manufacturing

Amazon Web Services has introduced Amazon Nova Multimodal Embeddings, a new capability aimed at enhancing information retrieval from complex manufacturing documentation. The technology maps text and images into a shared vector space, addressing challenges where critical data resides in diagrams, CAD files, or photographs rather than plain text.

Traditional text-only search systems often fail to locate information embedded within engineering diagrams, CAD drawings, or inspection photographs. Nova Multimodal Embeddings bridges this gap by allowing text queries to retrieve visual content and image queries to find relevant written specifications. This development is particularly relevant for industries like aerospace, automotive, and heavy manufacturing, where documentation is inherently multimodal.

To demonstrate its capabilities, AWS built a multimodal retrieval system for aerospace manufacturing documents using Amazon Nova Multimodal Embeddings, Amazon Bedrock, and Amazon S3 Vectors. The system was evaluated against 26 manufacturing-related queries, with its performance compared against a text-only retrieval pipeline.

The significance of multimodal search in manufacturing stems from the common practice of combining text with visual elements in crucial documents. Work orders, inspection reports, and material certifications frequently include assembly instructions alongside annotated photographs, or radiographic images of welds paired with detailed measurements. This new technology offers enterprises a more comprehensive way to access and utilize their technical data.