edgeparse

EdgeParse Skill

Enables AI agents to extract clean, structured content from any PDF — headings, tables, paragraphs, lists, bounding boxes — deterministically, without ML dependencies or GPU requirements.

Install: pip install edgeparse · Node.js: npm install edgeparse
Speed: ~0.023 s/doc (Apple M4 Max, 200-doc benchmark)

---

When to reach for this skill

Activate when the workflow involves:

• Reading or analyzing a PDF document on behalf of a user
• Building a RAG pipeline that ingests PDFs
• Feeding PDF content to an LLM for summarization, Q&A, or synthesis
• Extracting tables from financial reports, research papers, or invoices
• Processing a batch of documents for indexing or search
• An agent tool that must "open" a PDF and return its contents

---

Quick start

import edgeparse

# Convert any PDF to Markdown — best for LLM context windows
text = edgeparse.convert("report.pdf", format="markdown")

# Convert to JSON with bounding boxes and full structure
import json
doc = json.loads(edgeparse.convert("report.pdf", format="json"))

# Plain text (fast, minimal)
plain = edgeparse.convert("report.pdf", format="text")

The format parameter controls output:

Value	Best for
"markdown"	LLM context — headings, tables, lists in Markdown
"json"	Bounding boxes, citations, structured element metadata
"html"	Web rendering, semantic HTML5
"text"	Simple full-text search, minimal output

---

Core API

edgeparse.convert()

result: str = edgeparse.convert(
    input_path,             # str or Path — required
    format="markdown",      # output format (see table above)
    pages=None,             # e.g. "1-5" or "1,3,7-10" — specific pages only
    password=None,          # for password-protected PDFs
    reading_order="xycut",  # "xycut" (spatial sort, default) or "off"
    table_method="default", # "default" (ruling-line) or "cluster" (borderless)
    image_output="off",     # "off", "embedded" (base64), "external" (files)
)