IBM Research created Docling and donated it to the LF AI & Data Foundation on 29 April 2025. It is MIT licensed. PDFluent is a general-purpose PDF SDK (parsing, rendering, forms, signatures, PDF/A, redaction) that also ships a preview structured-extraction layer aimed at the same RAG and LLM use case. For Docling's own scope and supported formats, check its documentation.
use pdfluent::{ExtractOptions, PdfDocument};
fn main() -> pdfluent::Result<()> {
let doc = PdfDocument::open("report.pdf")?;
// Structured extraction (Preview)
let markdown = doc.extract_markdown(&ExtractOptions::default())?;
println!("{}", markdown);
Ok(())
}from docling.document_converter import DocumentConverter
source = "report.pdf"
converter = DocumentConverter()
result = converter.convert(source)
print(result.document.export_to_markdown())| Feature | PDFluent | Docling |
|---|---|---|
| License model | AGPL v3 or commercial | MIT |
| Core engine language | Rust | Python package on PyPI |
| Official Rust bindings | Not assessed here | |
| Official Node.js bindings | Not assessed here | |
| Official .NET bindings | Not assessed here | |
| Official Java bindings | Not assessed here | |
| WebAssembly (runs in the browser) | Not assessed here | |
| PDF creation & editing | Editing, for example image placement and replacement, redaction and form filling | Not assessed here |
| Merge / split PDFs | Yes, in the PDFluent editor | Not assessed here |
| Digital signatures | PAdES B-B, B-T and B-LT | Not assessed here |
| AcroForms / XFA forms | Fills XFA forms (read fields, set values, save) and flattens them to a regular PDF; AcroForm calculate and format actions run | Not assessed here |
| PDF/A validation & conversion | Validation for PDF/A-1, A-2 and A-3; conversion is a Preview feature | Not assessed here |
| Redaction | Yes: removes the redacted content from the file instead of covering it | Not assessed here |
| Markdown / LLM-ready text extraction | Preview: extract_markdown() / extract_blocks() (1.1.0) | Measured in the same run, on the same pages [A62, 2026-10-17] |
| Table structure recognition | Table model on the balanced tier (Preview, opt-in) | Included in the 19 measures below, same run [A62, 2026-10-17] |
| Extraction quality, 19 measures | Balanced tier with the table model (Preview): average 0.721 against 0.672 for Docling 2.126.0 over 19 measures* [A62, 2026-10-17]. *Balanced tier with the table model (Preview), on our own held-out half of four public datasets, 17 October 2026. The table model was trained on the source datasets of two of them (PubTables-1M, FinTabNet), and most of those held-out pages are in their training split. On the 13 measures without that overlap the ranking is the same (0.650, 0.642, 0.610). Details on the benchmarks page. | Measured in the same run, on the same pages [A62, 2026-10-17] |
| Speed and memory, one c3-standard-8 (8 vCPU) | Fast tier: 19 ms per page and 43 MB peak memory on all 8 cores. Balanced tier with the table model: 885 ms per page on all cores, slower than Docling on one thread [C22, 2026-10-17] | 1,526 ms per page and 1,629 MB peak memory on all cores; 4,187 ms per page on one thread [C22, 2026-10-17] |
| Non-PDF input formats (DOCX, PPTX, XLSX, HTML) | DOCX, PPTX, XLSX and HTML to Markdown (Preview) [A31-A34, 2026-10-04] | Read DOCX, PPTX, XLSX and HTML in our Office comparison [A31-A34, 2026-10-04]; check Docling's documentation for the full list of inputs |
| Vision-language-model pipeline option | No vision-language-model option in the extraction tiers described here | Not assessed here; check Docling's documentation |
| Local, offline execution | Yes: runs on your machine; optional features and model installs use the network | Not assessed here |
| Published, self-service pricing | Yes, prices published (as of 7 September 2026) | MIT licence; check Docling's documentation for any paid offering |
Choose PDFluent if you need a single engine that both works on PDFs (edit, sign, fill forms, redact, validate PDF/A) and extracts structured, LLM-ready text from them, with bindings beyond Python, and the fast tier's speed and memory on the axes we've measured.
Choose Docling if you want an MIT-licensed tool, if you need input formats or a vision-language-model pipeline that PDFluent does not provide (check Docling's documentation for both), or if you run on a GPU, which PDFluent has no route for.
Docling and PDFluent overlap on one axis: turning a PDF into clean text for an AI pipeline. On that axis we measured both tools, plus pymupdf4llm 1.28.2, in one run on the same pages. PDFluent's balanced tier with the table model averaged 0.721 across 19 measures, against 0.672 for Docling 2.126.0 and 0.693 for pymupdf4llm 1.28.2* [A62, 2026-10-17]. *Balanced tier with the table model (Preview), on our own held-out half of four public datasets, 17 October 2026. The table model was trained on the source datasets of two of them (PubTables-1M, FinTabNet), and most of those held-out pages are in their training split. On the 13 measures without that overlap the ranking is the same (0.650, 0.642, 0.610). Details on the benchmarks page. Speed is a separate question with a tier attached. On one 8-vCPU c3-standard-8, PDFluent's fast tier took 19 ms per page against Docling's 1,526 ms on all cores [C22, 2026-10-17]. PDFluent also edits, signs and redacts PDFs and fills forms; for what Docling covers beyond extraction, read its documentation.
Free under the AGPLv3, or a commercial licence if you would rather not publish your own source.