Bedrock Knowledge Base can now understand video, audio, and images directly!

Hi, I'm Shii-chan! Today I found news that's pretty exciting if you work with video or audio content, so let me tell you about it!

AWS What's New aws.amazon.com

What was announced?

According to AWS What's New, Amazon Bedrock's Managed Knowledge Base now supports a new embedding model: TwelveLabs' "Marengo 3.0." This lets you create multimodal embeddings for video, audio, and image content.

The story so far

Bedrock's Managed Knowledge Base already supported searching audio and video, but it worked by transcribing the audio or video into text first, then generating text-based embeddings from that transcript. That meant spoken content came through fine, but visual scenes and non-verbal audio cues that don't translate into words were hard to capture in the embeddings.

What changes

Marengo 3.0 encodes visual scenes, speech, and video-specific cues directly into multimodal embeddings, without going through a transcription step. In other words, it can now capture meaning that transcription alone would miss.

The workflow stays simple: upload your media assets from a data source like Amazon S3, sync, and then search using natural language, with no infrastructure to manage.

Dive Deep

Here are the concrete technical details from the announcement.

• Marengo 3.0 produces compact 512-dimensional vectors
• Search results include the start and end times of the matching segment, so you can jump directly to the relevant moment in a video
• Segmentation options are configurable to match your content structure
• Data sources include Amazon S3; the flow is upload, sync, then search in natural language

The announcement calls out these use cases:

• Sports analytics (e.g., finding specific plays across seasons of game footage)
• Media and entertainment
• Security
• Education (locating lecture segments by concept rather than keywords)
• Retail

For the full integration details, AWS points to the Amazon Bedrock Knowledge Base User Guide.

Wrap-up

• TwelveLabs Marengo 3.0 is now available as an embedding model in Bedrock's Managed Knowledge Base
• It embeds video, audio, and images directly as multimodal vectors, skipping the transcription step
• Vectors are a compact 512 dimensions, with segment-level timestamps for precise search
• Just upload to S3 and sync, no infrastructure to manage

If you're sitting on a big archive of video or audio and want to find "that one scene" using plain language, this is definitely worth a look!