Foliant.Core 1.8.1

Foliant

Layout-aware PDF document AI for .NET — fully local, no Python sidecar, no cloud APIs.

Foliant extracts structured content (Markdown / JSON / typed objects) from PDFs the way commercial document-intelligence services do: layout detection, per-region OCR, table-structure recognition, and reading-order assembly — all running on ONNX Runtime, entirely on your machine. Documents never leave the host.

using Foliant.Pipeline;

// Models download once into a local cache (SHA-256 verified), then everything runs offline.
using var processor = await FoliantProcessor.CreateDefaultAsync();
var result = await processor.ProcessAsync(File.ReadAllBytes("document.pdf"));

Console.WriteLine(result.Markdown);          // layout-aware Markdown
string json = result.ToJson(indented: true); // structured regions, tables, bounds, confidence

Why Foliant

• Born-digital fast path — pages with an embedded text layer take characters verbatim (layout still analyzed from pixels); OCR runs only where needed. ~0.4 s/page at 300 DPI.
• Forms done right — recursive ruling-line decomposition keeps checkbox marks, field labels, and values associated on government-style forms; table grids come from a hybrid of Microsoft TableTransformer and rule-based ruling analysis, with a grid-fit guard that renders a mis-detected form block as ordered prose instead of a scrambled table. Opt-in Foliant.Forms.* profile packs add typed key-value extraction for known form families.
• Self-verifying — every page enforces a coverage invariant (every extracted line provably lands in the output or is intentional page furniture), reports word recall against the PDF's own text layer, and scores reading-order fidelity against that text layer's order — so a page that keeps every word but permutes it is caught (word recall alone is order-blind). On a 474-page federal-RFP reference corpus: 99.7% average word recall, 100% of pages ≥95%, zero text loss, zero fabricated form values.
• Pluggable — every stage (layout, OCR, tables, reading order) is an interface in Foliant.Core; swap any backend without forking.

Packages

Model weights (~280 MB total) are not inside the packages; they download on first use with checksum verification, or pre-fetch them yourself and point FoliantProcessor.CreateDefault at the directory.

Apache 2.0. English-language documents. Scanned-document support ships today — deskew, denoise, contrast normalization, and coarse page-orientation correction (0/90/180/270°); page dewarp and low-DPI super-resolution are on the roadmap.