TwelveLabs Marengo in Amazon Bedrock Managed Knowledge Base: Video Search Without the Infrastructure Build

Getting real semantic video search out of an embedding model takes more than the model. Enterprises still need a multi-vector index across modalities, query routing, result fusion and reranking, and a vector database to run it on. That's a multi-month build customers had to take on themselves just to try Marengo on their own content.

Today, that project goes away for AWS customers. TwelveLabs' Marengo is now available as a managed embedding model inside Amazon Bedrock Managed Knowledge Base. Customers can select Marengo on Amazon Bedrock, connect a video data source, and get working multimodal semantic search, without provisioning a vector database, building a retrieval pipeline, or managing their own infrastructure.

The Gap This Closes

Marengo is our category-leading multimodal embedding model purpose-built for video, not an image model stretched to fit. Teams have gotten real value out of it as an inference endpoint on Amazon Bedrock alone. But an embedding model only answers the question of "how do I represent this video mathematically?" Not "how do I search it?"

That second question is where most customers got stuck: routing a query across visual and audio embeddings, fusing and reranking multiple result sets, keeping a vector store running at scale. Evaluating whether Marengo was even the right fit for their use case meant building most of a search architecture first.

That's true of most embedding models, not just Marengo, and it meant technical evaluations that should take weeks often stretched into months. That kind of delay is exactly the sort of thing that pushes a decision toward whatever’s fastest to get running, not necessarily what’s best.

What Changed

With Marengo in Amazon Bedrock Managed Knowledge Base, the search stack is no longer the customer's problem. Connect a data source, such as an Amazon Simple Storage Service (Amazon S3) bucket, and Amazon Bedrock Managed Knowledge Base handles ingestion, indexing, and query orchestration automatically. Marengo generates the embeddings, and TwelveLabs' retrieval logic, running inside AWS's managed infrastructure, handles routing, fusion, and ranking behind the scenes. The customer gets a search endpoint, not a set of parts to assemble.

Here’s what that looks like in practice:

• Multimodal search. Visual and audio understanding are combined automatically, with ranking handled behind the scenes instead of hand-tuned per query.

• Hybrid search. Combine semantic meaning with metadata filters like source, modality, or file type to narrow results precisely.

• Data stays in-account. All data stays inside the customer's AWS account throughout ingestion and search.

• Faster evaluation. Customers can go from connecting a data source to seeing real search results on their own footage almost immediately.

That last point is the piece that actually matters. Marengo has always been excellent at finding the right moment inside a video. What's different is how quickly someone can get from "we have access to Marengo" to "we have working semantic search on our own content," and Amazon Bedrock Managed Knowledge Base makes that almost immediate.

Why This Matters

The pattern shows up everywhere that video is piling up faster than anyone can watch it. Media and entertainment companies — studios, broadcasters, sports leagues, post-production houses — sit on archives too large to search by humans scrubbing timelines. Security and safety teams in video surveillance providers or smart home companies need to search footage for specific events or objects, but only within infrastructure that never leaves their own account. Everyone has been asking for the same thing: point us at our data and let us search it.

Underneath that request is a shift that's already underway. Over the last several years, text has become a first-class input to AI agents: large language models turned words into tokens, and those tokens became the semantic layer that agents reason and act on. Video hasn't had that moment yet. Most of it still sits in archives and camera systems, accessible only through filenames and human memory, despite it being the closest record we have to what actually happened.

That’s the frontier TwelveLabs has set out to build toward: making every second of video addressable, searchable, and usable by agents. Marengo in Amazon Bedrock Managed Knowledge Base is a concrete step in that direction, putting video on the same footing as text and structured data.

Built With AWS, Not Just For It

TwelveLabs designed the database schema and the retrieval logic that runs the search; AWS built the managed infrastructure, data connectors, and orchestration around it. Each team built the piece it's best positioned to build, which has always been our operating philosophy in bringing video intelligence to the market.

That kind of development work is also why TwelveLabs holds the AWS AI Competency, a designation AWS reserves for partners with demonstrated technical proficiency delivering AI solutions on its platform. Marengo in Amazon Bedrock Managed Knowledge Base is the newest result of that work. It’s also the clearest one yet for teams trying to search video at scale.

Try It Yourself

If you're evaluating video search on AWS, or have been putting it off because of the build it used to require, now's the time to take another look. Connect a data source in Amazon Bedrock Managed Knowledge Base and see what Marengo finds in your own content.

For a closer look at how it works under the hood, read the technical deep-dive, or reach out to our sales team.