gpu-document-processing

GPU-accelerated document processing extracts text from large PDFs while preserving headers, paragraphs, lists, tables, and page references. A CPU-based reasoning layer orchestrates GPU sandbox tasks to parse layouts and write structured findings.

GPU-accelerated tabular data extraction converts PDF tables into structured formats like CSV or DataFrames. It performs type detection and handles multi-row headers, outputting the structured table data to a shared directory.

Document chunking and embedding generation split large PDFs into meaningful chunks and produce embeddings for search and analysis. GPU sandboxes handle the bulk processing in parallel, aggregating metadata and per-document summaries.

Bulk processing handles collections of 10 or more files in parallel using GPU acceleration. The CPU orchestrator manages the GPU sandbox tasks, extracts metadata, and synthesizes per-document summaries for the entire collection.

GPU acceleration is required for the sandbox processing tasks that handle text extraction, table extraction, chunking, and embedding generation. A CPU-based reasoning layer orchestrates these GPU tasks and aggregates the final structured results.

Extracted text, structured tables, chunks, and embeddings are written to the /shared/ directory. The CPU-based orchestrator aggregates the results from the GPU sandbox and saves the final structured findings there.