News
Amazon Nova Multimodal Embeddings Unifies Text, Image, Video and Audio Search in Bedrock
Amazon announced general availability of Amazon Nova Multimodal Embeddings in Amazon Bedrock, introducing a unified embedding model that processes text, documents, images, video, and audio through a single model for crossmodal retrieval. The model is designed for agentic retrieval-augmented generation (RAG) and semantic search, enabling developers to unlock insights from mixed-modality content without building separate pipelines.
The model converts multimodal inputs into numerical embeddings that capture meaning across modalities, supporting use cases such as searching videos with text queries, retrieving visual documents, and matching audio clips to written descriptions, a recent AWS blog post said. Benchmarks included in the announcement show Nova leading on multiple retrieval tasks, such as ActivityNet and TextCaps, compared with models from TwelveLabs, Google Vertex AI, Cohere, and Amazon Titan.
Nova Multimodal Embeddings supports context lengths up to 8,192 tokens and inputs in up to 200 languages. It accepts synchronous and asynchronous API calls and provides segmentation (chunking) for long text, video, or audio inputs. Developers can choose among four output embedding dimensions (3,072, 1,024, 384, or 256) trained using Matryoshka Representation Learning, allowing a balance between accuracy and cost. The model's embeddings can be stored in Amazon S3 Vectors or integrated with Amazon OpenSearch Service for large-scale semantic search applications.
Amazon emphasized Nova's crossmodal search capabilities as a key differentiator, allowing users to query across content types within a unified semantic space. Responsible AI measures include content safety filters and fairness features built into Amazon Bedrock. The model can be invoked via real-time or asynchronous APIs, depending on workload latency sensitivity.
Amazon Nova Multimodal Embeddings is available in the US East (N. Virginia) Region, with pricing listed on the Amazon Bedrock pricing page. Additional developer resources include the Amazon Nova User Guide and the Nova model cookbook on GitHub.
About the Author
David Ramel is an editor and writer at Converge 360.