Skip to content
BranchesTags

Repository files navigation

FieldCure.DocumentParsers

Core Imaging Ocr Audio License: MIT

Lightweight document-to-text extraction library for .NET. Converts DOCX, HWPX, XLSX, PPTX, HTML, and PDF files into structured Markdown with heading detection and table support — designed for LLM/RAG pipelines.

Features

  • DOCX — headings, paragraphs, tables → markdown, math equations → LaTeX, metadata → YAML front matter, footnotes, endnotes, comments, headers/footers
  • HWPX — Korean standard format (KS X 6101/OWPML), headings, tables → markdown, equations → LaTeX, metadata → YAML front matter, footnotes, endnotes, memos, headers/footers
  • XLSX — sheets → markdown tables (multi-sheet support), metadata → YAML front matter
  • PPTX — slide text, speaker notes, slide tables → markdown, grouped shapes, metadata → YAML front matter
  • HTML — readable content extraction (SmartReader) → GitHub-flavored Markdown (ReverseMarkdown)
  • PDF — text extraction (PdfPig, pure managed). Page image rendering and OCR are separate opt-in packages.
  • Audio — MP3/WAV/M4A/OGG/FLAC/WebM → timestamped Markdown transcripts via Whisper.net (separate opt-in package).

Packages

Package Description Native deps
FieldCure.DocumentParsers DOCX, HWPX, XLSX, PPTX, HTML, PDF (text) —
FieldCure.DocumentParsers.Imaging PDF → page images (adds IMediaDocumentParser) PDFium
FieldCure.DocumentParsers.Ocr Tesseract OCR fallback for scanned PDFs — Windows (x64 + arm64) PDFium + Tesseract
FieldCure.DocumentParsers.Audio Audio → timestamped transcripts via Whisper.net — Windows only Whisper.net + NAudio

The core package is pure managed — no native binaries are pulled in unless you opt into Imaging, Ocr, or Audio.

The Ocr package is currently Windows only, but ships natives for both win-x64 (Tesseract 5.0 redistributed from the upstream NuGet) and win-arm64 (Tesseract 5.5.2 built from source via vcpkg, Authenticode-signed by FieldCure). The assembly carries [SupportedOSPlatform("windows")], so non-Windows consumers will see CA1416 warnings at compile time. Linux / macOS OCR is on the roadmap; in the meantime use the core package directly for PDFs that have an embedded text layer (works everywhere).

Deprecated (v2.0): FieldCure.DocumentParsers.Pdf (replaced by core + Imaging) and FieldCure.DocumentParsers.Pdf.Ocr (renamed to .Ocr).

Installation

# Core (DOCX, HWPX, XLSX, PPTX, HTML, PDF text)
dotnet add package FieldCure.DocumentParsers

# PDF page rendering (optional, pulls PDFium)
dotnet add package FieldCure.DocumentParsers.Imaging

# OCR fallback for scanned PDFs (optional, pulls Tesseract + PDFium)
dotnet add package FieldCure.DocumentParsers.Ocr

# Audio transcription (optional, pulls Whisper.net runtimes + NAudio)
dotnet add package FieldCure.DocumentParsers.Audio

Quick Start

using FieldCure.DocumentParsers;

// PDF is now registered automatically — no AddPdfSupport() call needed.
var parser = DocumentParserFactory.GetParser(".pdf");
var text = parser!.ExtractText(File.ReadAllBytes("document.pdf"));

// Same API for all formats
foreach (var ext in DocumentParserFactory.SupportedExtensions)
    Console.WriteLine(ext);
// .docx, .hwpx, .xlsx, .pptx, .html, .htm, .pdf
// Opt-out control for metadata, footnotes, etc.
var parser = new DocxParser();
var options = new ExtractionOptions
{
    IncludeMetadata = false,
    IncludeFootnotes = false
};
var text = parser.ExtractText(File.ReadAllBytes("report.docx"), options);
using FieldCure.DocumentParsers;
using FieldCure.DocumentParsers.Imaging;

// Upgrade the factory's .pdf entry to IMediaDocumentParser (text + images).
DocumentParserFactoryImagingExtensions.AddImagingSupport();
var pdf = (IMediaDocumentParser)DocumentParserFactory.GetParser(".pdf")!;
var images = pdf.ExtractImages(File.ReadAllBytes("document.pdf"), dpi: 150);
using FieldCure.DocumentParsers;
using FieldCure.DocumentParsers.Ocr;

// Register an OCR-augmented PDF parser. Dispose the engine at shutdown.
using var ocr = DocumentParserFactoryOcrExtensions.AddOcrSupport();

// Scanned pages are OCR'd; pages with an embedded text layer go through PdfPig.
var parser = DocumentParserFactory.GetParser(".pdf")!;
var text = parser.ExtractText(File.ReadAllBytes("scanned.pdf"));
using FieldCure.DocumentParsers;
using FieldCure.DocumentParsers.Audio;

// Register audio support. Dispose the transcriber at shutdown.
await using var transcriber = DocumentParserFactoryAudioExtensions.AddAudioSupport();

var parser = DocumentParserFactory.GetParser(".mp3")!;
var transcript = parser.ExtractText(File.ReadAllBytes("meeting.mp3"));
// Let the library pick a Whisper model size based on detected GPU/RAM/cores.
// QualityBias.Accuracy (default) shifts up one tier — suitable for batch indexing.
using FieldCure.DocumentParsers.Audio;

var recommended = WhisperEnvironment.RecommendModelSize(); // e.g. WhisperModelSize.Large
var options = AudioExtractionOptions.Default.WithModelSize(recommended);

var probe = WhisperEnvironment.Probe();
Console.Error.WriteLine(
    $"[Audio] CUDA={probe.CudaAvailable} Vulkan={probe.VulkanAvailable} " +
    $"RAM={probe.SystemRamBytes / (1024L * 1024 * 1024)}GB → {recommended}");

Custom Parser

Implement IDocumentParser to add support for any format:

public class MyParser : IDocumentParser
{
    public IReadOnlyList<string> SupportedExtensions => [".xyz"];

    public string ExtractText(byte[] data)
    {
        // Your extraction logic
        return "extracted text";
    }
}

// Register
DocumentParserFactory.Register(new MyParser());

Table Output Format

All parsers convert tables to markdown format for LLM comprehension:

| Name | Age | City |
| --- | --- | --- |
| Alice | 30 | Seoul |
| Bob | 25 | Busan |

Pipe characters inside cells are escaped as \| to preserve table structure.

Limitations

DOCX

Supported Not Yet Supported
Paragraph text Charts / SmartArt
Headings (Heading1–9 style + OutlineLevel) Images (embedded) — no OCR
Tables → markdown (including nested) Tracked changes
Hyperlink text Text boxes / shapes
Math equations (OMML → LaTeX) Legacy .doc format (use LibreOffice to convert)
Numbered / bulleted lists (as text)
Multi-section documents
Metadata → YAML front matter
Footnotes / Endnotes
Comments → inline blockquote
Headers / Footers

HWPX

Supported Not Yet Supported
Paragraph text (hp:p) Form fields
Headings (header.xml outline levels) Legacy .hwp format (binary, not XML)
Standalone tables (hp:tbl)
Embedded tables (hp:p > hp:run > hp:tbl)
Math equations (hp:equation → LaTeX)
Drawing text (hp:drawText)
Multi-section documents
Table cell merging
Metadata → YAML front matter
Footnotes / Endnotes
Memos → inline blockquote
Headers / Footers

XLSX

Supported Not Yet Supported
Cell text values Charts
SharedString references Pivot tables
Multi-sheet (separated by headings) Formula evaluation (values only)
Empty row/cell handling Conditional formatting info
Pipe character escaping Merged cells (partial support)

PPTX

Supported Not Yet Supported
Slide text (all shapes) SmartArt
Title / body separation Charts
Speaker notes Animations / transitions info
Slide tables → markdown Audio / video references
Grouped shapes (text extraction) Math equations
Slide ordering
Field elements (slide numbers, dates)

HTML

Supported Not Yet Supported
Readable article extraction (SmartReader) JavaScript-rendered content (SPA)
GitHub-flavored Markdown output Login-required pages
Tables, headings, links preserved Embedded media extraction
Nav / ads / footer auto-removal Non-UTF-8 encodings

PDF

Supported Not Yet Supported
Text extraction (text-based PDF) — core package Form field extraction
Page image rendering — Imaging package Digital signature info
OCR fallback for scanned PDFs — Ocr package PDF/A validation
Multi-page documents, Unicode text
English + Korean OCR (tessdata_fast) — Ocr

Audio

Supported Not Yet Supported
MP3, WAV, M4A, OGG, FLAC, WebM — Audio package Real-time microphone input
Timestamped Markdown transcript Speaker diarization
Whisper ggml model cache Video audio track extraction
Custom IAudioTranscriber injection Word-level timestamps
Environment-aware model size recommendation NVML-based VRAM probing (deferred to v0.2)

Model size selection

WhisperEnvironment.RecommendModelSize(QualityBias) picks a Whisper model size based on the local environment. CUDA/Vulkan availability is detected via driver-shipped nvcuda.dll / vulkan-1.dll; physical RAM via GlobalMemoryStatusEx; logical cores via Environment.ProcessorCount. VRAM is intentionally not probed in v0.1 — RAM ≥ 8 GB on a GPU host is treated as sufficient, and Whisper.net's runtime fallback (CUDA → Vulkan → CPU) handles real-device mismatches.

The balanced matrix (used directly by QualityBias.Balanced):

Environment Recommended model
GPU available, RAM ≥ 16 GB Large
GPU available, RAM ≥ 8 GB Medium
CPU only, RAM ≥ 16 GB, cores ≥ 8 Small
CPU only, RAM ≥ 8 GB Base
Otherwise Tiny

QualityBias.Accuracy (default) shifts the recommendation one tier up — appropriate for batch indexing where transcription latency is acceptable. QualityBias.Speed shifts one tier down — appropriate for interactive UI flows where the user is actively waiting.

Repository Structure

All library projects multi-target net8.0;net10.0.

src/
├── DocumentParsers/                     FieldCure.DocumentParsers 2.0 (net8.0 + net10.0)
│   ├── Ooxml/                           DocxParser, PptxParser, XlsxParser
│   ├── Hwpx/                            HwpxParser
│   ├── Html/                            HtmlParser
│   └── Pdf/                             PdfParser (text via PdfPig)
├── DocumentParsers.Imaging/             FieldCure.DocumentParsers.Imaging 1.0 (net8.0 + net10.0)
├── DocumentParsers.Ocr/                 FieldCure.DocumentParsers.Ocr 1.1 (net8.0 + net10.0)
├── DocumentParsers.Audio/               FieldCure.DocumentParsers.Audio 0.1 (net8.0 + net10.0)
├── DocumentParsers.Cli/                 Console tool for manual output inspection
├── DocumentParsers.Tests/               MSTest — core + PdfParser tests
├── DocumentParsers.Imaging.Tests/       MSTest — PdfImageRenderer tests
├── DocumentParsers.Ocr.Tests/           MSTest — OcrPdfParser + TesseractOcrEngine tests
└── DocumentParsers.Audio.Tests/         MSTest — Audio parser tests

Build & Test

dotnet build
dotnet test

See Also

Part of the AssistStudio ecosystem.

Release Notes

License

MIT — Copyright (c) 2026 FieldCure Co., Ltd.

About

Document text extraction library for DOCX, HWPX, XLSX, PPTX, and PDF. Supports OOXML math-to-LaTeX conversion, Hancom equation parsing, and IMediaDocumentParser for image extraction.

www.nuget.org/packages/FieldCure.DocumentParsers

Topics

csharpdocument-parserdocxdotnetequation-parserhwpxlatexnugetpdftext-extraction

Resources

Readme
MIT license

Contributing

Contributing
Activity
Custom properties

Stars

4 stars

Watchers

0 watching

Forks

0 forks
Report repository

Releases

Packages

Contributors

Languages