Video and image search in Amazon Bedrock Knowledge Base using Marengo 3.0

AWS says Marengo Embed 3.0 is now generally available as an embedding model inside Amazon Bedrock Knowledge Bases. The pitch is simple: video and other media have been hard to search by meaning, and now the same kind of natural-language lookup people use for text can reach into clips, images, and audio tracks too.

That matters because the old way is a mess. A team trying to find a single moment in hours of footage has usually had to wire together transcription, frame extraction, embeddings, vector storage, and synchronization logic. Bedrock Knowledge Bases bundles the plumbing into a managed RAG service that handles storage, ingestion, embedding, re-ranking, and retrieval.

Marengo Embed 3.0 is multimodal, so it encodes video, audio, images, and text into a 512-dimensional vector space. AWS says Managed Knowledge Bases automatically creates multimodal embeddings and captures visual, textual, speech, and audio signals in one representation. The service supports MP4, MOV, JPEG, PNG, and audio tracks, with connectors for Amazon S3, SharePoint, Confluence, and more.

The walkthrough in the post uses a 10-minute clip from the 2022 FIFA World Cup final. You upload the media to S3, create a managed knowledge base in the Bedrock console, pick TwelveLabs/Marengo Embed 3.0 instead of the default Amazon Titan Text embeddings, and leave the segmentation defaults at 4 seconds for audio and video. After syncing, the system extracts frames, transcribes audio, generates embeddings for each segment, and writes vectors to the index.

AWS then shows a test query such as “show me the penalty kicks from this soccer match.” The results come back ranked, with source chunk start and end times, the source URI, and the embedding type, so an app can jump straight to the relevant section. The service is available in us-east-1 and us-west-1, and AWS says you pay for what you store and retrieve, plus the standard model invocation rate for embeddings generation.