extract-pdf

name	extract-pdf
description	Extract clean text and markdown from any PDF. Fixes broken Unicode mappings that make Chrome's copy-paste produce gibberish. Returns plain text, basic markdown, or structured markdown with TOC, headings, page markers, and token counts. Works locally via WASM — your PDFs never leave your machine.

Preamble (run first)

_PLUGIN_DIR="${CLAUDE_PLUGIN_ROOT:-$(cd "$(dirname "$0")/.." 2>/dev/null && pwd || echo "$HOME/.config/pdf-to-text")}"
_UPD=$("$_PLUGIN_DIR/bin/update-check" 2>/dev/null || true)
[ -n "$_UPD" ] && echo "$_UPD" || true

If UPGRADE_AVAILABLE <old> <new> is output: tell the user a new version is available and ask if they want to upgrade. If yes, run $_PLUGIN_DIR/hooks/install-engine.sh. If JUST_UPGRADED <old> <new>: tell the user "PDF to Text engine updated to v{new}!" and continue.

PDF to Text

A local PDF-to-markdown extraction engine with a 7-level fallback cascade that recovers text from PDFs whose embedded Unicode mappings are broken or missing. It outputs three markdown formats: plain, basic (one # Page N per page), and structured (YAML frontmatter + TOC + font-detected headings + fenced code blocks).

Use this skill when you need clean text from a PDF, not when you need the visual layout. For tables, figures, and images, look elsewhere.

When to invoke

Use glyph-api when:

• The user asks you to read, summarize, or extract content from a PDF
• The user hits "gibberish on paste" from a PDF in their browser or editor
• The user wants to feed a PDF into an LLM (RAG ingest, summarization, question-answering, extraction pipeline)
• The user has a URL that ends in .pdf and wants the content
• Chrome's native PDF viewer or cat/pdftotext produced garbage on a particular PDF
• You need structured output (TOC, heading hierarchy, page markers) for navigation or agentic querying

Do NOT use glyph-api for:

• Tables — current extraction collapses tabular data into flat prose. Use a real table-aware tool (camelot, tabula-py, pdfplumber).
• Figures / diagrams / images — glyph-api extracts their text labels but not the visual structure. Images are ignored.
• Scanned PDFs with no text layer — glyph-api is not an OCR engine. Use tesseract or a cloud OCR service first, then feed the resulting text-layered PDF to glyph-api.
• Non-Latin scripts (CJK, Arabic, Hebrew, Devanagari) — MVP is Latin-only; complex script support is on the roadmap.

How to invoke

Four invocation paths, ordered by agent-friendliness.

Path 0 — MCP tools (preferred for Claude Code agents)

This plugin registers MCP tools automatically. Use them directly:

• extract_pdf — pass url or path, get plain/basic/structured markdown + stats
• render_markdown — fetch and parse any .md URL, get sections + token count
• list_recent — see recently extracted PDFs from the local cache

Use the extract_pdf tool with path: "/path/to/document.pdf" and format: "structured"

Path 1 — Chrome extension agentic API (for browser-automation agents)

If you're driving a browser (Claude in Chrome, Playwright, Puppeteer, CDP):

// Navigate to any .pdf URL. The extension's DNR rule intercepts and
// redirects to the viewer. The viewer runs the WASM extractor and
// publishes the result to a page-world global.
await navigate("https://example.com/document.pdf");

// Wait for extraction to finish.
await waitFor(() => window.__glyph?.status === "ready");

// Read the three markdown formats (all pre-computed).
const plain      = window.__glyph.markdown.plain;
const basic      = window.__glyph.markdown.basic;
const structured = window.__glyph.markdown.structured;

The viewer also fires a glyph:status CustomEvent and sets document.body.dataset.glyphStatus to loading / ready / error, so agents that prefer selector-based waits can use:

await waitForSelector('body[data-glyph-status="ready"]');

Iframe-embedding pages receive a postMessage with {type: "glyph:status", status, markdown} once ready.

Requirements: the Glyph Chrome extension must be installed (unpacked or from the Chrome Web Store). The extension ID is stable for a given store listing; print it from any viewer tab via chrome.runtime.id in the DevTools console.

Path 2 — Native Rust CLI (preferred for shell/pipeline agents)

cd /Users/jordan/Code/glyph-api
./target/release/glyph-api path/to/document.pdf

Outputs plain text to stdout. One form-feed (\f) between pages. No frontmatter, no headings, no markdown syntax — just the extracted text.

To build the binary from source (~5s):

cd /Users/jordan/Code/glyph-api
cargo build --release --bin glyph-api

Path 3 — WASM module from Node/TypeScript

import initWasm, { extract_chars_with_positions } from "./pkg/glyph_api.js";
import { readFileSync } from "node:fs";

await initWasm({ module_or_path: readFileSync("./pkg/glyph_api_bg.wasm") });

const pdfBytes = readFileSync("document.pdf");
const json = extract_chars_with_positions(pdfBytes);
const parsed = JSON.parse(json);
// parsed.pages[i].text  — per-page extracted text
// parsed.pages[i].chars — per-char position metadata (for spatial queries)
// parsed.total_chars, total_resolved, total_unresolved

To also get the three markdown formats, import the markdown module:

import {
  toPlain,
  toBasicMarkdown,
  toStructuredMarkdown,
} from "./extension/dist/markdown.js";

const structured = toStructuredMarkdown(parsed, {
  srcUrl: "https://example.com/document.pdf",
});

Output format reference

Plain

<page 1 text>

<page 2 text>

...

No markdown. Blank line between pages. Use for embeddings and keyword search.

Basic

# Page 1

<escaped page 1 text>

---

# Page 2

<escaped page 2 text>

Page headers, minimal escaping. Use when you want page-boundary awareness without heading detection.

Structured (RAG-ready)

---
source: https://example.com/document.pdf
pages: 9
chars: 21146
tokens: ~5.3k
headings: 13
extracted_at: 2026-04-11T17:00:00.000Z
extractor: glyph-api
---

## Table of Contents

- Bitcoin: A Peer-to-Peer Electronic Cash System _(p.1)_
- 1. Introduction _(p.1)_
- 2. Transactions _(p.2)_
...

<!-- page 1 -->

## Bitcoin: A Peer-to-Peer Electronic Cash System

Satoshi Nakamoto
[e-mail]
www.bitcoin.org

## Abstract

A purely peer-to-peer version of electronic cash...

## 1. Introduction

Commerce on the Internet has come to rely...

Notable features:

• YAML frontmatter with source URL, page/char/token counts, ISO timestamp
• Auto-generated TOC from detected headings
• <!-- page N --> markers — invisible when rendered, queryable for per-page chunking in RAG pipelines
• Heading detection via three channels: font-size clustering, bold-font promotion, and a numbered-section pattern (^\d+\.\s+[A-Z])
• Known section words (Abstract, References, Acknowledgements, Bibliography, Appendix, Conclusion) promoted to H2 even when set in body-size bold
• Fenced code blocks for C (#include), Python (def), math formulas (∑ ⋅ ≤), and diagram labels — keeps GitHub's markdown renderer from choking on them

Gotchas / current limitations

• Synthetic stress-test PDFs look repetitive — dense_20p_70l_r2.pdf contains the same sentence repeated 1400 times by design. That's the input, not an extraction bug.
• The arxiv fi ligature bug — papers using ligatures (fi, fl, ffi, ffl) may emit raw glyph codes like 002gures instead of figures. Future work: extend the L3/L4 cascade to recognize ligature glyph names.
• Form field glue — IRS-style forms that position each digit with a separate Tj operator produce Form10 4 0 2025 instead of Form 1040 (2025). Future work: detect tight-spacing runs and apply word grouping.
• Letter-spaced titles — 110TH CONGRESS set with positive letter-tracking extracts as 110 TH C ONGRESS. Future work: detect heading-level letter-spacing and collapse.
• Math formulas are not LaTeX — they're fenced as \``textfor legibility but won't render via MathJax. Future work:$$...$$` wrapping for paragraphs where symbol density is high enough.
• Tables are prose — glyph-api doesn't emit GFM tables. Multi-column data collapses into rows that may or may not align.

Verification

If you're unsure whether glyph-api improved on Chrome's native extraction for a given PDF, look at the frontmatter's headings: count and the structured output's TOC — both should be non-empty for any real document. For the famous "Attention Is All You Need" paper (arxiv 1706.03762), the structured output has:

• 31 sections in the TOC
• The H2 title + all section numbers 1–7 + sub-sub-sections like 3.2.1
• Cross-validates clean against docmap --type code --lang c on the appendix

If the output has headings: 0 and no TOC, the PDF likely has unusual content structure — try feeding it back and reporting the file shape.

Cross-validate with docmap

If you have docmap (v0.4.0+) installed, you can verify the output's structural integrity:

docmap extracted.structured.md           # section tree
docmap extracted.structured.md --type code --lang c     # find C blocks
docmap extracted.structured.md --type math              # find math blocks
docmap extracted.structured.md --json | jq '.documents[0].sections'

This gives independent confirmation that the extracted markdown is semantically navigable — agents can jump to sections, count constructs, and reason about document structure.