Gpu Document Processing

This handles large PDF processing and document collections by offloading the heavy lifting to a GPU sandbox while keeping the agent reasoning on CPU. It's built for the 50+ page PDFs and bulk extraction jobs where you need text chunking, table extraction, or embedding generation at scale. The sandbox-as-tool pattern means your API keys stay out of the GPU environment and you only pay for acceleration during actual processing, not during the agent's planning and synthesis work. Works with NVIDIA NeMo Retriever for embeddings and can process document batches in parallel. If you're dealing with large document sets that would otherwise crawl through CPU-only extraction, this is the play.