AWS adds TwelveLabs Marengo 3.0 to Bedrock for managed multimodal retrieval

AWS has added TwelveLabs Marengo 3.0 as an embedding model in Amazon Bedrock Managed Knowledge Base, giving developers a managed way to build natural-language search across video, audio and image archives without assembling separate pipelines for frames, transcripts, vector indexing and retrieval.¹

The move expands Bedrock Knowledge Bases beyond text-centered retrieval-augmented generation workflows. AWS said the service already supported media search by transcribing audio and video into text and generating text embeddings. Marengo 3.0 adds multimodal embeddings that encode visual scenes, speech and video cues directly, capturing signals that transcript-only systems can miss.¹

For AI application developers, the practical change is architectural. Instead of extracting frames, generating transcripts, choosing embedding models for each modality, storing vectors and mapping search results back to timestamps, teams can upload media assets from sources such as Amazon S3, sync them with Bedrock Managed Knowledge Base and query them in natural language.¹

From text RAG to media-native retrieval

Retrieval-augmented generation systems have typically centered on text: documents are chunked, embedded, stored in a vector index and retrieved to ground model responses. That pattern works well for manuals, web pages, support tickets and knowledge articles. It becomes more complex when the source material is a video library, classroom recording archive, surveillance repository or product imagery collection.

In those settings, developers often need to decide how frequently to sample video frames, how to align speech transcripts with visuals, how to represent images and audio in a shared embedding space, and how to return results that point users to the right moment rather than simply the right file. AWS is positioning the Marengo 3.0 integration as a managed alternative to that custom stack.

According to AWS, Marengo 3.0 generates compact 512-dimensional vectors and supports configurable segmentation options, allowing developers to tune how media is divided based on the structure of the content.¹ That matters for applications where retrieval quality depends not only on finding the right video, but also on identifying the relevant clip inside it.

Segment-level results

A key feature of the integration is segment-level retrieval. AWS said results include segment start and end times, enabling applications to send users directly to the relevant moment in a video.¹

That capability changes the user experience for media search. A security analyst could search for a specific visual event, an educator could locate the portion of a lecture that explains a concept, and a media editor could find a scene type or spoken reference across a large archive. Instead of returning a long file and leaving the user to scrub through it manually, an application can surface the relevant temporal span.

For developers building AI assistants over media assets, segment timestamps can also improve downstream generation. A model can cite or summarize the relevant clip while the application preserves a path back to the original evidence.

Use cases across industries

AWS cited use cases across sports analytics, media and entertainment, security, education and retail.¹ In media organizations, multimodal retrieval can help teams search broadcast footage, production archives or marketing assets by meaning rather than file names or tags. In education, lecture videos can become searchable by topic, example or concept, even when the exact keyword is not spoken.

In security, natural-language retrieval over video could help investigators find events that combine visual and temporal context. In retail, teams could search product imagery, store video or customer-experience recordings to locate scenes, displays or behaviors relevant to operations and merchandising.

The common engineering pattern is the same: large media repositories become searchable through embeddings that represent more than text. By packaging that capability inside Bedrock Managed Knowledge Base, AWS is treating multimodal retrieval as infrastructure rather than a bespoke application layer.

What developers still need to design

The announcement reduces systems work, but it does not remove application design choices. Developers still need to decide how to organize source media, set segmentation options, govern access to sensitive audio and video, and design user interfaces that clearly expose retrieved clips, confidence signals and source context.

They also need to evaluate retrieval quality for domain-specific content. A retail visual search workflow, a lecture archive and a security investigation tool may require different segmentation strategies, metadata filters and relevance thresholds.

Still, the direction is clear: cloud providers are moving retrieval systems from text-only knowledge bases toward managed multimodal infrastructure. With Marengo 3.0 in Bedrock Managed Knowledge Base, AWS is making video, audio and image search a more native part of the generative AI application stack.¹

AWS adds TwelveLabs Marengo 3.0 to Bedrock for managed multimodal retrieval

More from Siften Technology

Microsoft’s COPILOT() Retreat Shows the Limits of AI Inside Excel’s Grid

Things to know

Comments

Source Material

From text RAG to media-native retrieval

Segment-level results

Use cases across industries

What developers still need to design

Key topics covered

Extreme Networks Brings Agent ONE Coworker to General Availability as Network AI Moves Past Dashboard Chat

Automox Moves AI Vulnerability Workflows From Discovery to Mitigation

Google Adds Governed Drive, Gmail and Chat Automation to Workspace Studio