Semantic PDF extraction

Understand the page
before simplifying it.

PDFs do not contain the clean document tree that HTML does. @boxpdf/reader uses the page layout to put content back into a useful reading order, with paragraphs, tables, nested sections, cards, addresses, labels, and values.

Reading order Paragraphs and lines Table structure Source coordinates retained Accessible HTML LLM-ready context

Use semantic mode Try the visual reader

Live semantic comparison

The designed page on the left. Its document flow on the right.

Choose a real PDF. The browser renders its original pages on the left while boxpdf turns the same document into reflowable headings, paragraphs, lists, cards, address groups, totals, and tables on the right.

Visual PDFbrowser renderer · scrollable

Semantic HTMLreflowable document structure

—open

—semantic HTML

—pages structured

—peak reader cache

Loading the semantic comparison…

The important distinction

Keep the page layout before simplifying it.

Text stored next to each other in a PDF may belong to different columns. Text stored far apart may belong to the same sentence. boxpdf looks at where content appears on the page before deciding how it should read.

visual

Preserve what the author placed

Coordinates, baselines, font metrics, vectors, images, clipping, transforms, and paint order remain available. The output can be rendered and compared to the PDF.

Best for viewers and previews
Best for visual LLM input
Keeps the layout needed to build semantic HTML

semantic

Produce HTML that reads naturally

boxpdf orders the text, groups related content, builds real tables, removes repeated headers and footers, and keeps sections and continued tables together across pages.

Best for accessibility and search
Best for extraction and indexing
Best for compact language-model context

How it works

Read the PDF. Rebuild the page. Produce useful HTML.

1PDF objectsfonts · text · paths · images

→

2Visual modelgeometry · paint order · source refs

→

3Semantic flowsections · cards · paragraphs · tables

The semantic output keeps links to the original text and page coordinates. You can highlight the source, show the matching page, or let an LLM check the result against the visual document.

Use the right representation for the next job.

Accessible reading order

Turn positioned content into headings, paragraphs, lists, and tables that assistive technology can navigate.

Grounded LLM extraction

Send compact semantic content for reasoning and retain visual HTML for layout-sensitive verification and citations.

Search and indexing

Index coherent blocks instead of isolated glyph runs, while preserving page and bounding-box provenance for every result.

Tables and forms

Recover real rows, continued tables, totals, addresses, cards, labels, and values that plain-text extraction cannot keep connected.

Same extraction, different intent

Choose semantic output when meaning matters most.

import { blobSource, openPdf } from "@boxpdf/reader";
import { writeHtmlDocument } from "@boxpdf/html-writer";

const pdf = await openPdf(await blobSource(file));

await writeHtmlDocument(pdf.pages(), write, {
  profile: "semantic",
  semanticLookaheadPages: 4
});

Semantic output streams with bounded lookahead, so cross-page structure improves without turning a large document into a large memory spike.