PDF ingestion remains one of the biggest practical bottlenecks for retrieval-augmented generation and large-scale accessibility remediation: text order is lost, tables break, and manual tagging doesn’t scale. This project focuses on producing deterministic, citation-friendly outputs (JSON with bounding boxes, Markdown) while using an optional hybrid AI path only for genuinely hard pages—so you get reproducible local parsing for most documents and high-accuracy fixes where needed.
What Sets It Apart
- Deterministic local extraction + hybrid AI fallback: simple pages are parsed locally very fast (low latency), while complex pages (borderless tables, low-quality scans, formulas, charts) are routed to a hybrid AI backend to boost accuracy—this reduces unnecessary cloud/AI usage and preserves reproducibility.
- Element-level provenance: JSON output includes bounding boxes and semantic types for every element (heading, paragraph, table, image), enabling precise source citation and clickable "jump-to-source" UX in RAG pipelines.
- Accessibility-first pipeline: layout analysis feeds an auto-tagging flow that can generate Tagged PDFs (previewed as coming Q2 2026) and integrates programmatic validation with veraPDF and PDF Association guidance—bridging document ingestion and remediation.
- Pragmatic performance/accuracy tradeoffs: benchmarks in the project show market-leading extraction accuracy in hybrid mode while keeping a very fast deterministic local mode for bulk processing.
Who It's For and Tradeoffs
Great fit if you build RAG/document search pipelines, need reproducible element-level citations, or must scale PDF accessibility remediation without proprietary SDKs. Look elsewhere if you require native Office (DOCX/XLSX/PPTX) processing (not supported), or if you need GPU-accelerated deep-vision models for bespoke VLM tasks—this project optimizes CPU-first deterministic extraction with optional lightweight VLM/AI enrichment. For full PDF/UA export and enterprise visual studio features, enterprise add-ons are offered.
