gpu-document-processing

To extract tables from PDFs in bulk, this Skill converts PDF tables into structured CSV or DataFrames while maintaining column typing and handling merged or multi-row headers. It processes large document collections by offloading parsing to a GPU environment for faster results.

You can generate embeddings for large document collections by using this Skill to perform semantic chunking into 512-token segments and running embedding workflows via NVIDIA NeMo Retriever NIM. It keeps agent reasoning on CPU while sending embedding generation to a GPU.

Layout-preserving PDF extraction detects headers, paragraphs, lists, and tables while keeping page and section boundaries intact. This approach ensures structured outputs include per-section content and page-referenced results for accurate document chunking.

A GPU is required for this Skill because it uses sandbox-as-tool execution to send parsing and embedding workloads to a GPU-equipped environment. This setup is necessary when your document collections exceed typical CPU processing comfort zones for bulk text extraction.

CPU-only document processing becomes slow for large PDF collections because parsing layout, harvesting tables, and generating embeddings are computationally heavy. GPU-accelerated extraction offloads these intensive workloads to reduce the time and effort required for bulk processing.