<- all tokdocs

NVIDIA's cuDF and cuVS: Making 90% of the World's Data Searchable

Watch on TikTok

View on TikTok ->

About 90% of the data generated every year -- PDFs, videos, speeches, images -- is unstructured. Until now, this data has been effectively useless for computation. You could read it, file it away, and that was it. You couldn't query it or search it meaningfully. NVIDIA's two new foundational libraries, cuDF and cuVS, are positioned to change that by making unstructured data indexable, searchable, and queryable at scale.

The Unstructured Data Problem

Structured data (rows and columns, SQL databases) has been searchable and queryable for decades. Unstructured data has not, because there is no natural indexing scheme. A PDF doesn't have columns. A video doesn't have rows. To search unstructured data, you first have to understand its meaning and purpose -- something that was computationally intractable until the arrival of multimodal AI.

The breakthrough is that the same AI capabilities that solved perception and understanding -- reading a PDF, watching a video, listening to speech -- can now be used to extract semantic meaning from unstructured data and embed it into searchable structures.

NVIDIA's Two Platforms

NVIDIA has created two foundational libraries that parallel what they did with RTX for 3D graphics:

Library Purpose Data Type
cuDF GPU-accelerated DataFrames Structured data (tables, CSVs, SQL-like operations)
cuVS GPU-accelerated vector search and clustering Unstructured/semantic data (embeddings from PDFs, videos, text)

cuDF: Structured Data at GPU Speed

cuDF accelerates familiar data processing libraries -- pandas, Polars, Apache Spark -- on NVIDIA GPUs. It handles the structured side: cleaning, joining, aggregating, and transforming tabular data.

cuVS: Vector Search for Semantic Data

cuVS is the more novel piece. It provides GPU-accelerated vector search and clustering for embeddings -- the numerical representations of meaning extracted from unstructured data by AI models. When an AI reads a PDF and understands its content, that understanding gets encoded as a vector. cuVS indexes those vectors (using algorithms like CAGRA and IVF-PQ) so they can be searched at massive scale and low latency.

cuVS integrates with existing vector database infrastructure including FAISS, Milvus, Lucene, and Kinetica.

Why This Matters

The combination of multimodal AI (to extract meaning) and GPU-accelerated vector search (to index and query that meaning) unlocks the 90% of world data that has been sitting in file systems, unqueryable. This is not incremental improvement -- it is making an entirely new category of data computationally useful.

Key Takeaways

  • 90% of data generated annually is unstructured and has been effectively unsearchable
  • Multimodal AI can now extract semantic meaning from PDFs, video, speech, and images
  • NVIDIA's cuDF handles structured data processing on GPUs; cuVS handles vector search for semantic data
  • cuVS integrates with existing vector databases (FAISS, Milvus, Lucene)
  • Together, these libraries form the infrastructure layer for making unstructured data queryable at scale

Resources

Published April 18, 2026. Writeup generated from a favorited TikTok.