Condé Nast cuts video search from 250 minutes to under 2 with Amazon Bedrock

The problem: search tools that cannot look inside video

Condé Nast's editorial teams had no fast way to search their video archive, which holds more than 140,000 videos. According to AWS, editors relied on titles and descriptions and manually scrubbed through footage, spending an average of 250 minutes per content discovery task. The effect reached brands such as Vogue, GQ, Vanity Fair and Wired.

AWS describes the core issue as structural: existing search tools cannot see inside video content. Teams depended on institutional knowledge to locate assets, which created single points of failure when specific people were unavailable. Some content also stayed undiscoverable because no keyword in a title or description matched the queries editors were running.

How the solution works

The system runs intent-based semantic search across video transcripts, visual elements and audio. The team chose the TwelveLabs Marengo embedding model, which AWS says can jointly encode visual, audio and transcript signals. Marengo is accessed through Amazon Bedrock, and the embeddings are indexed in Amazon OpenSearch Service, which provides managed k-nearest neighbor (k-NN) search.

Editors can describe what they need in natural language, such as "beginner yoga content" or "celebrity interview about sustainability", and receive clips with precise timestamps. The listed capabilities are:

• Multimodal understanding: a clip can be found by what is said, shown or heard.
• Image-based queries: users can upload a reference image to find visually similar content.
• Typo tolerance and intent resolution: the search focuses on meaning rather than exact keywords.
• Timestamp precision: results point to exact moments, so users do not need to watch entire clips.

AWS says using Bedrock let the team adopt a specialized embedding model without building or operating its own model-serving infrastructure, while applying AWS governance controls such as IAM, Amazon VPC and AWS CloudTrail.

Architecture: two decoupled planes

Because of the library's size, the team separated the compute-heavy work of generating embeddings from the low-latency work of serving queries. According to AWS, this lets each plane scale, fail and evolve independently, and it proved essential during backfill processing.

The ingestion plane is asynchronous. A video uploaded to Amazon S3 triggers an event, the ingestion service validates it and extracts metadata, and it is split into segments on Amazon ECS with AWS Fargate. Each chunk is processed by Marengo on Amazon Bedrock using asynchronous invocation. The embeddings are stored in S3 for durability and indexed in OpenSearch. AWS says the event-driven design gives per-step retry logic, parallel chunk processing and full traceability, which the team relied on heavily during the initial backfill of 140,000 videos.

The serving plane handles queries through load balancers and services on ECS with Fargate. The search service turns a query into a vector embedding, runs a k-NN search against the OpenSearch index, and enriches results with titles, thumbnails and references from Amazon DocumentDB.

The solution is deployed across multiple Availability Zones. OpenSearch replicates synchronously across three zones (two active, one standby), DocumentDB uses a primary node and a standby replica across two zones, and compute and data workloads sit in private subnets.

Results and what they mean

Condé Nast ran a benchmarking workshop in May 2026 to measure the impact. According to the company's measurements, content discovery time fell by 99.2 percent, from 250 minutes to approximately 2 minutes per task. Manual video review effort dropped by more than 90 percent, since teams receive targeted clips with timestamps instead of scrubbing through individual videos. AWS also cites approximately $800,000 in estimated annual operational savings based on productivity gains.

For media companies with large video archives, the case shows how embeddings that span visual, audio and transcript signals can make older content searchable without relying on manual metadata or individual staff knowledge. The figures come from Condé Nast's own workshop and are reported by AWS.