---
id: wiki-2026-0508-pdf-format
title: PDF Format
category: 10_Wiki/Topics
status: verified
canonical_id: self
aliases: [Portable Document Format, ISO 32000, PDF/A]
duplicate_of: none
source_trust_level: A
confidence_score: 0.9
verification_status: applied
tags: [pdf, document, format, parsing, generation]
raw_sources: []
last_reinforced: 2026-05-10
github_commit: pending
tech_stack:
  language: python
  framework: pypdf
---

# PDF Format

## 매 한 줄
> **"매 cross-reference table 의 random-access 의 binary container"**. 매 Adobe (1993) 의 PostScript-derived 의 ISO 32000 의 standardize 의 page-fixed-layout 의 dominant interchange format. 매 2026 년 의 PDF/A-4 (archival) + PDF 2.0 의 modern variant 의 LLM-extraction 의 challenge 의 source (no semantic structure 의 guarantee).

## 매 핵심

### 매 file 구조
1. **Header** — `%PDF-2.0` (또는 1.x).
2. **Body** — sequence of indirect objects (`N G obj ... endobj`).
3. **Cross-reference table** (`xref`) — byte offset of each object.
4. **Trailer** — root + info + size + xref offset.

### 매 object types
- Boolean, Number, String (literal `()` or hex `<>`), Name (`/Name`), Array, Dictionary, Stream (filtered binary).
- Page tree (Catalog → Pages → Page) + Resources (Font, XObject, etc.).

### 매 응용
1. Text/table extraction (LLM 의 RAG ingest).
2. Form fill (AcroForm / XFA).
3. Digital signature (PAdES).
4. Print fidelity (PDF/X for press).
5. Archive (PDF/A — embed fonts, no encryption).

## 💻 패턴

### Text extraction (pypdf, 2026)
```python
from pypdf import PdfReader

reader = PdfReader("doc.pdf")
text = ""
for page in reader.pages:
    text += page.extract_text() + "\n"
# pypdf 5.x: layout-mode option for column-aware
text = "\n".join(p.extract_text(extraction_mode="layout") for p in reader.pages)
```

### Better extraction with pdfplumber (preserves layout)
```python
import pdfplumber

with pdfplumber.open("doc.pdf") as pdf:
    for page in pdf.pages:
        # Tables
        for table in page.extract_tables():
            print(table)
        # Words with bbox
        for word in page.extract_words():
            print(word['text'], word['x0'], word['top'])
```

### LLM-grade extraction with Unstructured (2026)
```python
from unstructured.partition.pdf import partition_pdf

elements = partition_pdf(
    filename="doc.pdf",
    strategy="hi_res",          # uses layout model
    infer_table_structure=True,
    extract_images_in_pdf=True,
)
# Each element: Title, NarrativeText, Table, Image
```

### Generate PDF (reportlab)
```python
from reportlab.lib.pagesizes import A4
from reportlab.pdfgen import canvas

c = canvas.Canvas("out.pdf", pagesize=A4)
c.setFont("Helvetica-Bold", 16)
c.drawString(72, 800, "Invoice #1234")
c.setFont("Helvetica", 10)
for i, line in enumerate(items):
    c.drawString(72, 760 - i*14, line)
c.showPage()
c.save()
```

### Modern HTML→PDF (Playwright, replaces wkhtmltopdf)
```python
from playwright.async_api import async_playwright

async def html_to_pdf(html, out):
    async with async_playwright() as p:
        browser = await p.chromium.launch()
        page = await browser.new_page()
        await page.set_content(html)
        await page.pdf(path=out, format="A4", print_background=True)
        await browser.close()
```

### Sign PDF (PAdES, pyhanko)
```python
from pyhanko.sign import signers, fields
from pyhanko.pdf_utils.incremental_writer import IncrementalPdfFileWriter

with open("input.pdf", "rb") as inf:
    w = IncrementalPdfFileWriter(inf)
    fields.append_signature_field(w, sig_field_spec=fields.SigFieldSpec("Sig1", box=(50, 50, 200, 100)))
    signer = signers.SimpleSigner.load("cert.pem", "key.pem")
    with open("signed.pdf", "wb") as out:
        signers.sign_pdf(w, signers.PdfSignatureMetadata(field_name="Sig1"), signer=signer, output=out)
```

### Repair / linearize (qpdf CLI)
```bash
qpdf --linearize input.pdf output.pdf
qpdf --object-streams=generate --compress-streams=y input.pdf small.pdf
qpdf --check input.pdf            # validate xref + structure
```

### Encrypted PDF
```python
from pypdf import PdfWriter

writer = PdfWriter(clone_from="doc.pdf")
writer.encrypt(user_password="user", owner_password="owner", algorithm="AES-256")
with open("encrypted.pdf", "wb") as f:
    writer.write(f)
```

## 매 결정 기준

| 상황 | Tool |
|---|---|
| Text extraction (simple) | pypdf 5.x |
| Layout / tables | pdfplumber |
| LLM RAG ingest | Unstructured + hi_res / Marker / Docling |
| Generation (reports) | reportlab / WeasyPrint |
| HTML → PDF (modern) | Playwright (Chrome headless) |
| Forms / signing | pyhanko + qpdf |
| Repair / optimize | qpdf, mutool |

**기본값**: 매 ingest → Unstructured (layout-aware), 매 generate → Playwright (HTML).

## 🔗 Graph
- 응용: [[Document AI]]
- Adjacent: [[OCR]]

## 🤖 LLM 활용
**언제**: 매 form-filled PDF 의 question. 매 extraction tool 의 selection. 매 schema mapping.
**언제 X**: 매 binary blob 의 direct edit 의 LLM 의 X. 매 spec-conformant tool 의 use.

## ❌ 안티패턴
- **Regex-based PDF parsing**: 매 binary + xref 의 fragile. 매 lib 의 사용.
- **Single extraction strategy**: 매 scanned PDF 의 OCR fallback. 매 hi_res strategy.
- **No PDF/A for archive**: 매 font 의 missing 의 future render fail.

## 🧪 검증 / 중복
- Verified (ISO 32000-2:2020, pypdf docs, Unstructured docs, qpdf manual).
- 신뢰도 A.

## 🕓 Changelog
| 날짜 | 변경 |
|---|---|
| 2026-05-08 | Phase 1 |
| 2026-05-10 | Manual cleanup — PDF structure + 2026 extraction/generation toolchain |