AWS Adds TwelveLabs Marengo Embed 3.0 to Bedrock Knowledge Bases for Video and Image Search

What's new

Amazon Web Services announced general availability of TwelveLabs Marengo Embed 3.0 as an embedding model option inside Amazon Bedrock Knowledge Bases. The integration lets teams run natural language searches — for example, "show me the penalty kick in the second half" — directly against raw video, audio, and image files, without building a custom pipeline of transcription services, frame extractors, embedding models, and vector databases.

How it works

Marengo Embed 3.0 is a multimodal embedding model that jointly encodes video, audio, images, and text into a single 512-dimensional vector space. Once selected inside a Managed Knowledge Base (Managed MKB), the system automatically handles segmentation, frame sampling, transcription, and embedding generation — no pre-processing of uploaded files is required.

Supported file types include MP4 and MOV video, JPEG and PNG images, and audio tracks. Data can be ingested from Amazon S3, SharePoint, Confluence, and other native connectors. Default audio and video segmentation is set to four-second chunks, configurable in advanced settings.

In AWS's walkthrough, a 10-minute clip of the 2022 FIFA World Cup final was ingested into a knowledge base, then queried with natural language prompts. Results returned ranked chunks with metadata — start time, end time, source URI, and embedding type — allowing developers to pull the exact video segment matching a query.

Applications can query the knowledge base through the Amazon Bedrock Retrieve API via the Boto SDK, or as a target inside Amazon Bedrock AgentCore's Gateway.

Availability and pricing

Managed Knowledge Bases with Marengo Embed 3.0 support is live in two AWS Regions: US East (N. Virginia, us-east-1) and US West (N. California, us-west-1). AWS says a complete list of supported models and Regions is maintained in its documentation, implying broader rollout is likely but not yet confirmed.

Pricing follows a pay-for-storage-and-retrieval model. Embedding generation with Marengo Embed 3.0 is billed at AWS's standard Bedrock model invocation rate; AWS did not publish a specific per-token or per-minute figure in this announcement, directing users to its general pricing documentation instead. Exact cost per 1M tokens or per minute of video was not disclosed.

Use cases AWS highlights

AWS points to several target industries, according to the company: sports analytics teams searching for specific plays across a season of footage, media companies indexing streaming archives, security teams searching camera footage for incidents, educational platforms locating lecture segments by concept rather than keyword, and retailers searching product demo videos for specific features.

What this means

This is a managed-service integration, not a new foundation model. Marengo Embed 3.0 itself comes from TwelveLabs; AWS's contribution here is folding it into Bedrock's existing Knowledge Bases infrastructure so customers avoid assembling their own multimodal RAG stack. The practical value is in reduced engineering overhead — teams that previously needed separate transcription, frame-extraction, and vector-database components can now point at an S3 bucket and get semantic video search with a few console clicks.

The two-Region limit (both in the US) signals this is an early-stage rollout rather than a global release, and the absence of concrete pricing figures means cost-sensitive teams processing large video archives should test invocation costs carefully before committing to production workloads. Competitors offering multimodal RAG — including specialized vector database vendors and rival cloud platforms — will likely face pressure to match this level of managed simplicity for video and audio search.