Show HN: DocSlicer – Structure-aware PDF/DOCX/PPTX parser, 31 pages/SEC on CPU (github.com)

DocSlicer, a new document parsing tool, has made waves in the AI/ML community by achieving an impressive processing speed of 31 pages per second without relying on large language models (LLMs) or heavy machine learning frameworks. Designed specifically for business documents such as PDFs, Word files, and PowerPoint presentations, DocSlicer efficiently converts these formats into structured content, including clean text chunks, tables, and navigable hierarchies. It scored high on the BizDocBench benchmark, outperforming existing tools with metrics like 0.98 content faithfulness and a unique capability to handle various document structures accurately. The significance of DocSlicer lies in its deterministic design, which allows it to function without any model weights, facilitating quick deployment that doesn't require GPU resources or extensive setup time. Its structure-aware chunking preserves semantic coherence by splitting text at heading boundaries, ensuring non-overlapping chunks that can be directly embedded into AI pipelines. The tool's integration potential is substantial; whether it's serving as a classic Retrieval-Augmented Generation (RAG) component or a vectorless approach for immediate responses from documents, DocSlicer promises to streamline workflows in sectors like legal, finance, and technical documentation, enhancing the efficiency of document management tasks.