Condé Nast says its AWS-built multimodal search reduced video discovery from 250 minutes to about two minutes per task across a 140,000-plus-video archive.
Condé Nast has deployed a multimodal search system that helps its editorial teams find relevant moments across a video archive of more than 140,000 items. AWS described the system on September 29, 2026, reporting that a May benchmarking workshop measured discovery time falling from an average of 250 minutes to approximately two minutes per task.
The system was developed with the AWS Generative AI Innovation Center and combines Amazon Bedrock, Amazon OpenSearch Service and TwelveLabs’ Marengo embedding model. Instead of relying only on video titles and descriptions, it searches across transcripts, visuals and audio. Editors can enter natural-language requests, submit a reference image, and receive relevant clips with precise timestamps; the search is also designed to handle typos and imprecise wording.
The architecture separates the work of preparing videos for search from responding to user queries. Ingestion processes uploaded videos, splits them into segments and generates multimodal embeddings asynchronously. Those embeddings are stored in Amazon S3 and indexed in OpenSearch. A separate search tier converts queries into vectors, retrieves similar segments and combines results with metadata such as titles and thumbnails from Amazon DocumentDB.
Condé Nast’s workshop results also showed a reduction of more than 90 percent in manual video-review effort. The company estimated annual operational savings at approximately $800,000, attributing the estimate to productivity gains from reducing repetitive manual searches. These figures are Condé Nast’s measurements and estimate, as reported by AWS; the article does not present them as independently audited results.
The system uses multiple Availability Zones for resilience, while AWS identity, network and audit controls support access management and oversight. AWS says the solution has been running in production for six months and remained available while the ingestion pipeline reprocessed archive content. For Condé Nast’s teams, the change means searching by what appears or occurs in a clip—not just by the words previously attached to its file.