# NVIDIA Cosmos Dataset Search (CDS)

A platform for semantic search across large video datasets, supporting text-to-video and video-to-video queries with GPU-accelerated embeddings and vector similarity search.

NVIDIA Cosmos Dataset Search (CDS) is a platform for semantic search across video collections. It supports both text and video queries against a common embedding space, using a GPU-accelerated embedding service and a vector similarity search backed by Milvus with NVIDIA's cuVS acceleration, and can search across multiple collections at once.

An ingestion pipeline extracts frames and metadata from incoming video automatically, storing media in S3-compatible object storage such as LocalStack, MinIO or AWS S3, with batch and incremental ingestion parallelized through Ray. A REST API and a dedicated CLI cover programmatic access, collection management and data ingestion, and search results include video metadata for downstream use.

It deploys through Docker Compose for local development or Helm charts for AWS EKS and generic Kubernetes, with monitoring, logging and independent scaling of each component. Its embedding service requires an NVIDIA GPU with hardware video decoding; unsupported video inputs are rejected with a re-encoding hint rather than falling back to a CPU decoder.

Website: https://github.com/NVIDIA-Omniverse-blueprints/cosmos-dataset-search