pdf-compiler 0.2.0

Stitch large PDFs from a YAML spec: title pages, ToC, markdown, image galleries

pdf-compiler

Stitch large PDFs from a YAML spec. Title pages, clickable tables of contents, markdown sections with auto-indexed headings, embedded PDFs with page-range selection, and packed image galleries — all from one declarative file.

uv run pdfc compile spec.yaml -o out.pdf

---

Install

# clone, then:
uv sync

The console script pdfc is installed into the project venv. Run it with uv run pdfc ... or activate the venv with source .venv/bin/activate and call pdfc directly.

Requires Python ≥ 3.14. WeasyPrint pulls in cairo/pango — on macOS brew install cairo pango gdk-pixbuf libffi is the usual prerequisite; on Debian/Ubuntu, apt install libcairo2 libpango-1.0-0 libpangoft2-1.0-0.

---

Quickstart

A minimal spec:

# minimal.yaml
output: out.pdf
metadata:
  title: My Document

sections:
  - type: title
    title: My Document
    subtitle: A demonstration
    date: 2026-05-21

  - type: toc

  - type: markdown
    path: intro.md

Compile it:

uv run pdfc compile minimal.yaml

The output PDF has a title page, a clickable ToC pointing at every markdown heading, and the rendered markdown body.

---

CLI

pdfc compile  SPEC [--out OUT] [-j N] [--no-cache]
pdfc validate SPEC
pdfc watch    SPEC [--out OUT]
pdfc cache    clear
pdfc --version

• compile runs the full pipeline and writes the output PDF. -j N sets the worker count for parallel section compilation (default: cpu_count() - 1). --no-cache forces a fresh build.
• validate parses the spec and checks every referenced input (markdown files exist, PDFs open, page ranges are in bounds, image files decode) without producing any output. Exits non-zero on problems — useful in CI.
• watch runs an initial compile, then re-runs on every change under the spec's directory. Errors don't kill the watcher.
• cache clear wipes every cached compiled-section PDF from the user cache directory ($XDG_CACHE_HOME/pdf-compiler or ~/.cache/pdf-compiler).

---

The YAML spec

The top-level keys are output, metadata, defaults, and sections. Only sections is required.

output: build/report.pdf

metadata:
  title:   "Annual Report 2025"
  author:  "Ivan Weissburg"
  subject: "Fiscal year summary"
  keywords: [report, 2025, demo]

defaults:
  index_headers: true        # markdown headings become ToC entries
  page_size: letter          # letter | legal | a4 | a5 | tabloid
  margin: 0.75in
  regularize_pages: false    # scale embedded PDFs to fit page_size
  page_numbering:
    enabled: false           # stamp page numbers on each page
    front_matter: roman      # roman | arabic | none
    body: arabic
    position: bottom-center  # bottom-{center,left,right}, top-…

vars:                        # see "Variables" below
  petitioner: "Jane Smith"
  filing_no:  "I-751"

sections:
  - …                        # see below

Unknown keys are rejected with a YAML line number — typos surface immediately instead of being silently ignored.

Section types

Each section has a type field that selects its schema. Sections appear in the output in the order they're listed.

title — cover page

- type: title
  title:    "Annual Report 2025"      # optional; falls back to metadata.title
  subtitle: "Fiscal Year Summary"     # optional
  author:   "Author Name"             # optional; falls back to metadata.author
  date:     2026-05-21                # optional; see below
  front_matter: true                  # use roman numerals for this page
  in_toc:   false                     # default: don't list in ToC

title, author, and date all fall back to the top-level metadata block when omitted on the section, so a minimal cover can be as short as:

metadata:
  title: "Annual Report 2025"
  author: "Author Name"
sections:
  - type: title

date resolves in this order: section wins → else metadata → else today's date. Setting date: ~ (YAML null) at either level explicitly disables the date.

toc — table of contents

- type: toc
  title: "Table of Contents"   # optional, default shown
  depth: 3                     # max heading level to include (1–6)
  front_matter: true

The ToC is rendered with dotted leaders and resolved page numbers. Every entry is a clickable internal link. Place it anywhere in the section list; you can include multiple ToCs (e.g., one per part).

header — divider page

- type: header
  title:    "Part II — Financials"
  subtitle: "Detailed breakdowns"    # optional
  body: |                            # optional markdown shown below
    Introductory paragraph in **markdown**.
  in_toc: true
  subtoc: false                      # add a mini-ToC for this part
  subtoc_depth: 3

Set subtoc: true to follow the divider with a mini table-of-contents listing every entry from this header up to the next header section (or the end of the document). Useful for multi-part documents where each part deserves its own overview page.

markdown — chapter rendered from a .md file

- type: markdown
  path: chapters/intro.md
  title: "Introduction"      # optional; else the first H1 in the file
  index_headers: true        # optional; else inherits defaults.index_headers

When index_headers is on, every heading in the markdown becomes a nested ToC entry. The heading hierarchy maps to ToC depth.

Markdown rendering uses CommonMark with GFM-style pipe tables, strikethrough (~~gone~~), and URL autolinking enabled.

pdf — embed an existing PDF

- type: pdf
  path: vendor/q1-report.pdf
  pages: "1-10,15,20-"       # 1-based, inclusive; "20-" = to end
  title: "Q1 Vendor Report"  # optional; else the file stem
  rotate: 0                  # 0 | 90 | 180 | 270
  preserve_bookmarks: true   # merge included PDF's outline under this entry
  regularize_pages: null     # null=inherit defaults.regularize_pages; true/false to override
  in_toc: true

Set regularize_pages: true (or enable it on defaults) when the embedded PDFs come from a mix of sources — letter scans, A4 PDFs, and oversized originals. Each source page is scaled & centered onto a target-sized blank page so the final document has uniform on-screen dimensions. Pages that already match the target are passed through untouched.

Page-range syntax:

token	meaning
5	only page 5
2-4	pages 2 through 4 (inclusive)
5-	from page 5 to the end
-3	from page 1 to page 3
omitted	all pages

images — packed image gallery

- type: images
  title: "Site Photographs"
  per_page: 4                # images per page (grid layout)
  layout: grid               # grid | autopack
  captions: below            # below | above | overlay | none
  variable_heights: false    # proportional row heights, preserve order
  optimize_packing: false    # sort by aspect ratio + proportional heights
  images:
    - { path: site/a.jpg, caption: "Entrance, looking north" }
    - { path: site/b.jpg, caption: "South wall, post-repair" }
    - { path: site/c.jpg, caption: "Portrait shot", rotate: 90 }

Layout modes:

• grid places exactly per_page images per page on a √N grid (e.g. per_page: 4 → 2×2; per_page: 6 → 2×3). By default each row gets an equal share of the page height.
• autopack uses a justified-rows algorithm: images fill each row to the page width, rows stack until the page is full. Variable images per page, non-overlapping by construction.

Packing options (grid layout):

• variable_heights: true — row heights are proportional to the natural dimensions of the images in that row rather than equal fractions. A page with one portrait and one landscape image fills edge-to-edge instead of leaving up to 40% whitespace. Image order is preserved.
• optimize_packing: true — implies variable_heights and also sorts images widest-first before assigning pages, grouping similar aspect ratios together for the most uniform pages. Image order is not preserved.

Per-image rotation:

EXIF orientation is applied automatically. For manual corrections, set rotate (degrees clockwise) on any image:

images:
  - { path: photo.jpg, caption: "Normal" }
  - { path: sideways.jpg, caption: "Rotated CW", rotate: 90 }
  - { path: upside_down.jpg, rotate: 180 }

---

Variables

Any user-facing string — titles, subtitles, captions, image captions, markdown content, header bodies, and PDF metadata fields — can reference {{ name }} placeholders. Names resolve from a merged dict of user-defined vars: and a set of builtins:

name	value
today	today's date in ISO format (2026-05-21)
year	four-digit year (2026)
month	zero-padded month (05)
day	zero-padded day (21)
month_name	full English month name (May)

User entries in vars: override builtins with the same name.

vars:
  petitioner: "Jane Smith"
  case_no:    "MSC-2026-0421"

sections:
  - type: title
    title:    "Petition by {{petitioner}}"
    subtitle: "Filed {{today}} — Case {{case_no}}"

  - type: markdown
    path: cover_letter.md   # may also use {{petitioner}}, {{today}}, …

Unknown names render as the literal source ({{nothere}}) — existing documents that happen to contain double-brace text are not broken by the feature. Values are stringified with str(), so YAML ints, floats, and bools work as expected. Changing a variable invalidates the section cache for any section that interpolated it.

---

How it works

The pipeline is functional and runs in four phases:

parse → validate → resolve paths →
  compile non-deferred sections (parallel) →
  reserve pages for ToC + subtoc headers →
  render deferred sections against resolved offsets →
  assemble + metadata + outline + page-number stamps → out.pdf

• Sections speak in named destinations (sec-0003-intro-h2-foo), never in page numbers. Each section's compile() returns a temp PDF plus a list of destination names. Assembly remaps them to global page references via a /Catalog/Names/Dests name tree. As a free consequence, every <a href="#anchor"> link WeasyPrint emits in a markdown body or in the ToC becomes a working PDF link — no link rewriting needed.

• Two-pass deferred rendering, no iteration. Step 1 compiles every non-deferred section so we know its page count. Step 2 reserves N blank pages at each deferred slot (main ToC, plus any header with subtoc: true) based on entry count, then renders each deferred section with the resolved page labels. If anything overflows, the pipeline widens the plan once and re-renders. Named destinations mean page numbers in any ToC always resolve correctly regardless of where it lands.

• Content-addressed cache. Each section's output is keyed by blake3(spec_section + defaults + input_file_bytes + package_version). Re-running on the same inputs short-circuits to the cached PDF. Modify one markdown file and only that section recompiles.

• Parallel compilation uses multiprocessing (WeasyPrint isn't thread-safe). The pool is bypassed when -j 1 or when there's only one section.

• Front matter vs body numbering. Sections marked front_matter: true produce roman-numeral page labels in the ToC; body sections get arabic numerals starting at 1.

• Global page numbers. Set defaults.page_numbering.enabled: true to stamp the resolved label onto every page during the final assembly step. The stamp uses the same roman/arabic split as the ToC, so a Part II divider page reads "1" while a title page reads "i". A single shared Helvetica resource keeps the per-page overhead to one small content-stream object.

• Page-size regularization. With regularize_pages: true the embedder wraps each source page onto a fresh target-sized page via pikepdf's overlay primitive, scaling & centering to fit while preserving aspect ratio. Pages already at the target size are kept in place (no overhead) — only oversized or undersized inputs pay the wrap cost.

---

Development

uv sync                       # install deps + dev tools
uv run pytest                 # 158 tests; runs in ~4s
uv run pytest --cov           # with coverage
uv run ruff check src tests   # lint

Project layout:

src/pdf_compiler/
├── cli.py                  # typer CLI surface (lazy imports)
├── spec.py                 # pydantic models, discriminated union
├── loader.py               # ruamel.yaml → pydantic with line-number errors
├── pipeline.py             # public compile_spec / validate_spec / watch_spec
├── pipeline_impl.py        # orchestration (parallel compile + ToC + assemble)
├── context.py              # BuildContext (paths, cache, tmpdir, workers)
├── cache.py                # blake3 content-addressed section cache
├── assemble.py             # pikepdf concat + named-destinations + page-number stamps
├── md_ast.py               # markdown-it AST → headings + anchor injection
├── numbering.py            # roman / arabic page-number formatting
├── page_range.py           # "1-10,15,20-" parser
├── lengths.py              # CSS length parser + page-size table
├── interpolate.py          # {{name}} variable substitution + builtins
├── validate.py             # standalone input validation
├── watcher.py              # watchdog-based --watch
├── util.py                 # slugify
├── sections/
│   ├── base.py             # Section protocol, CompiledSection, TocEntry
│   ├── _common.py          # SectionMeta, helpers
│   ├── title.py            #   ↓
│   ├── header.py           #   each section type's impl
│   ├── markdown_doc.py     #   ↓
│   ├── pdf_ref.py          #   ↓
│   ├── images.py           #   ↓
│   └── toc.py              # two-pass ToC renderer (also: subtoc headers)
├── render/
│   ├── html.py             # jinja2 + WeasyPrint
│   └── templates/*.{html,css}
└── layout/
    └── pack.py             # grid + justified-rows image packers

tests/
├── unit/                   # ~135 unit tests, table-driven + hypothesis
├── integration/            # full-pipeline assertions via pdfplumber
├── conftest.py             # shared fixtures (make_pdf, png_bytes)
└── fixtures/

examples/                   # demonstration specs
examples_content/           # markdown + images referenced by examples

The CLI module only imports heavy dependencies inside function bodies; pdfc --help and pdfc validate start in well under 200 ms.

---

Examples

examples/report.yaml exercises every section type (title, ToC, two markdown chapters, a header divider with a subtoc, and a 5-image gallery), with stamped page numbers and roman/arabic front-matter numbering enabled:

uv run pdfc compile examples/report.yaml

examples/minimal.yaml is the smallest possible spec.

Real-world: an evidence packet

Bundling dozens of scanned-and-non-scanned PDFs into a single navigable document (immigration evidence, legal exhibits, audit binders) is what motivated the three "uniform-output" features — global page numbers, subtoc parts, and page-size regularization.

output: evidence.pdf

metadata:
  title: "I-751 Joint Petition Evidence"
  author: "Petitioner Name"

defaults:
  page_size: letter
  regularize_pages: true          # scanned A4/legal pages → letter
  page_numbering:
    enabled: true                 # stamp global numbers on every page
    body: arabic
    position: bottom-center

sections:
  - type: title
    subtitle: "Supporting Documentation"

  - type: toc
    depth: 2

  - type: header
    title: "Identity & Status"
    subtitle: "Petitioner and beneficiary identity documents"
    subtoc: true
  - type: pdf
    path: "[Main Form] I-751.pdf"
  - type: pdf
    path: "Green Card.pdf"
  - type: pdf
    path: "Iain Passport.pdf"

  - type: header
    title: "Financial Co-mingling"
    subtitle: "Joint accounts, tax returns, shared expenses"
    subtoc: true
  - type: pdf
    path: "2025 Tax Return.pdf"
  - type: pdf
    path: "Joint Checking Opening.pdf"
  - type: pdf
    path: "Apple Card Statements - First Pages.pdf"

  - type: header
    title: "Joint Residence"
    subtoc: true
  - type: pdf
    path: "6629 Fathom Way Goleta Lease Weissburg Moreno Jun 2024 signed.pdf"
  - type: pdf
    path: "PGE Bills - First Pages.pdf"

  - type: images
    title: "Photographs"
    layout: autopack
    captions: below
    images:
      - { path: photos/wedding.jpg,    caption: "Wedding, June 2024" }
      - { path: photos/anniversary.jpg, caption: "Anniversary, June 2025" }

Every embedded PDF — whether the original was 8.5×11", scanned at A4, or a phone-photographed image — comes out at uniform letter size. Every page bears a sequential arabic page number that matches the top-level ToC and the per-section subtoc.