PaperForge

PaperForge is a pure Elixir PDF generation engine. It builds PDF object graphs, page content streams, resources, cross-reference tables, trailers, text layout, vector graphics, metadata, and image XObjects directly in Elixir.

No browser, wkhtmltopdf, Chromium, ImageMagick, Ghostscript, or external rendering service is required.

PaperForge 1.1 provides a stable public API for pure-Elixir PDF authoring, bounded concurrent rendering, production telemetry, and reproducible performance tooling. Documented public modules follow Semantic Versioning throughout the 1.x series.

Start Here

Why PaperForge?

• Pure Elixir runtime. No browser, native executable, service, or installation-time compilation is required.
• Business-document layout. Build reports, invoices, statements, and contracts with automatic pagination, templates, tables, navigation, and reusable components.
• PDF-native output. Text, links, outlines, attachments, annotations, vectors, images, fonts, and metadata are emitted as PDF structures.
• Predictable production behavior. Rendering is deterministic, concurrent work is bounded, failures are isolated, and Telemetry events expose runtime cost.
• Two levels of control. Use PaperForge.Flow for structured documents or PaperForge.Page for exact coordinates and custom graphics.

Capability Overview

Choose An API

Installation

Add PaperForge to your dependencies:

def deps do
  [
    {:paper_forge, "~> 1.1"}
  ]
end

Then run:

mix deps.get

To use the GitHub release directly:

def deps do
  [
    {:paper_forge,
     github: "Manuel1471/paper_forge",
     tag: "v1.1.0"}
  ]
end

For local development against main:

def deps do
  [
    {:paper_forge,
     github: "Manuel1471/paper_forge",
     branch: "main"}
  ]
end

Quick Start

alias PaperForge.Flow

{document, report} =
  PaperForge.new(compress: true, pdf_version: "1.7")
  |> PaperForge.page_template(
    :default,
    size: :a4,
    margins: [top: 72, right: 54, bottom: 72, left: 54],
    header: "Quarterly Report",
    footer: "Page {page} of {total}"
  )
  |> PaperForge.layout(
    fn flow ->
      flow
      |> Flow.heading("Quarterly Report", level: 1)
      |> Flow.paragraph("""
      PaperForge measures, paginates, and renders document blocks automatically.
      """)
      |> Flow.list(
        ["Unified layout", "Automatic pagination", "Reusable templates"],
        type: :unordered
      )
      |> Flow.table(
        ["Metric", "Value"],
        [
          ["Revenue", "$120K"],
          ["Margin", "24%"]
        ],
        repeat_header: true
      )
    end,
    template: :default
  )

IO.inspect(report.pages, label: "Pages")
PaperForge.write!(document, "report.pdf")

Document Authoring

The authoring API builds structured documents on top of PaperForge.Flow. Register shared styles, reusable components, and inherited page templates once, then compose documents from measured blocks instead of calculating page coordinates manually.

alias PaperForge.Flow

document =
  PaperForge.new()
  |> PaperForge.style(:body, size: 10, line_height: 14)
  |> PaperForge.component(:customer, fn assigns ->
    Flow.new()
    |> Flow.rich_text([
      {assigns.name, [weight: :bold]},
      {"\n#{assigns.address}", [size: 9]}
    ])
  end)
  |> PaperForge.page_template(:base, size: :a4, margins: 54, footer: "Page {page} of {total}")
  |> PaperForge.page_template(:invoice, extends: :base, header: "Invoice")

{document, _report} =
  PaperForge.layout(document, fn flow ->
    flow
    |> Flow.table_of_contents()
    |> Flow.heading("Invoice")
    |> Flow.component(:customer, %{name: "Acme", address: "Monterrey, MX"})
    |> Flow.grid(2, ["Subtotal\n$1,200", "Due\n30 days"], cell_height: 60)
    |> Flow.columns(2, ["Terms and conditions...", "Payment instructions..."])
  end, template: :invoice)

Available authoring blocks include rich_text/3, table_of_contents/2, reference/3, component/4, grid/4, and columns/4. Tables accept explicit :column_widths, :header_fill_color, :header_color, and :stripe_fill_color options. See paper_forge_0_6_authoring.exs and linkedin_document_showcase.exs for complete documents.

Images support fit: :fill | :contain | :cover, horizontal and vertical alignment, and focal_point: {x, y}. Numbered images and tables create stable destinations for page-aware references.

The complete release example is paper_forge_0_6_complete.exs. It combines navigation, advanced tables, footnotes, endnotes, charts, SVG, QR, barcode, attachments, components, and custom report panels in one PDF.

Page-aware navigation is resolved with bounded multi-pass pagination:

flow
|> Flow.table_of_contents(title: "Contents")
|> Flow.heading("Financial results", destination: :financial_results)
|> Flow.reference(:financial_results, prefix: "Financial results begin on page ")

Custom blocks receive their measured block_x, block_y, block_width, and block_height in PageContext, so bespoke report panels can participate in normal flow without hard-coding page coordinates.

Typography And Report Visuals

Paragraph blocks can request hyphenation and minimum line counts around page breaks. Layout reports expose :measurements for each placed block.

flow
|> Flow.paragraph(long_copy, hyphenate: true, min_lines_at_top: 2, min_lines_at_bottom: 2)
|> Flow.chart([{"Q1", 418}, {"Q2", 432}, {"Q3", 451}], height: 140)
|> Flow.svg("<svg><rect x='0' y='0' width='80' height='30' fill='#0077b5'/></svg>", height: 40)
|> Flow.qr_code("https://example.com/pay/INV-2048", width: 96, height: 96)
|> Flow.barcode("20481234", width: 180, height: 64)

Advanced Tables And Notes

Table rows are measured from their wrapped cell content. :keep moves an oversized row to a fresh page, :split continues cell content across pages, and :error raises PaperForge.TableError when a row cannot fit.

Use Flow.cell/2 for composable cells with :colspan, :rowspan, :valign, per-cell colors, per-side borders, and nested flow blocks.

flow
|> Flow.table(
  ["Item", "Description"],
  rows,
  column_widths: [110, 340],
  repeat_header: true,
  row_split: :split,
  cell_line_height: 12
)
|> Flow.footnote("Values are unaudited and shown in USD.")
|> Flow.endnotes([])

Footnotes reserve space at the bottom of the current flow page, number themselves when the number is omitted, append a visible call marker to the preceding paragraph, rich-text block, heading, or final table cell, and continue on another page when necessary. Pass marker: false to author the call marker manually. Flow.endnotes/3 emits the collected notes as a document section.

Document Options

PaperForge.new/1 accepts:

• :compress - enables Flate compression for page content streams. Defaults to true.
• :pdf_version - sets the PDF header version. Supported values are "1.4", "1.5", "1.6", and "1.7". Defaults to "1.7".

PaperForge.new()
PaperForge.new(compress: false)
PaperForge.new(pdf_version: "1.4")
PaperForge.new(default_font: :helvetica)

Low-level Page API

Use PaperForge.Page when you need manual graphics, exact coordinates, or a lower-level drawing surface. New structured documents should usually start with PaperForge.Flow and PaperForge.layout/3.

Add a page with default options:

document =
  PaperForge.new()
  |> PaperForge.add_page(fn page ->
    Page.text(page, "Default A4 page", x: 72, y: 750)
  end)

Add a page with options:

document =
  PaperForge.new()
  |> PaperForge.add_page(
    [
      size: :letter,
      orientation: :landscape,
      origin: :top_left,
      margins: [top: 48, right: 54, bottom: 48, left: 54]
    ],
    fn page ->
      Page.text(page, "Landscape page", y: 48)
    end
  )

Supported page sizes:

:a3
:a4
:a5
:letter
:legal

Custom page sizes use {width, height} in PDF points:

Page.new(size: {500, 700})

All dimensions are expressed in PDF points.

1 point = 1/72 inch

Coordinates And Margins

PaperForge supports both PDF-native bottom-left coordinates and top-left coordinates.

Page.new(origin: :bottom_left)
Page.new(origin: :top_left)

You can also set the origin per operation:

Page.rectangle(page, x: 72, y: 72, width: 100, height: 40, origin: :top_left)

Margins can be uniform:

Page.new(margins: 72)

Or side-specific:

Page.new(margins: [top: 40, right: 50, bottom: 40, left: 50])

Content helpers:

Page.content_width(page)
Page.content_height(page)
Page.content_left(page)
Page.content_top(page)
Page.content_bottom(page)

Text

Draw a single line of text:

Page.text(
  page,
  "Centered title",
  x: Page.content_left(page),
  y: 72,
  width: Page.content_width(page),
  align: :center,
  font: :helvetica_bold,
  size: 24,
  color: Color.black()
)

Draw wrapped multiline text:

Page.text_box(
  page,
  """
  PaperForge wraps text into multiple lines using built-in font metrics.

  Explicit line breaks are preserved.
  """,
  x: Page.content_left(page),
  y: 120,
  width: Page.content_width(page),
  height: 160,
  font: :times_roman,
  size: 12,
  line_height: 17,
  align: :left
)

Supported alignment values:

:left
:center
:right

Fonts And Unicode Text

PaperForge supports two font paths: the 14 standard PDF Type 1 fonts and embedded TrueType fonts.

Standard Type 1 fonts are registered automatically when used:

:helvetica
:helvetica_bold
:helvetica_oblique
:helvetica_bold_oblique
:times_roman
:times_bold
:times_italic
:times_bold_italic
:courier
:courier_bold
:courier_oblique
:courier_bold_oblique
:symbol
:zapf_dingbats

Standard Type 1 fonts are convenient for simple Latin text, but they are not full Unicode fonts. For visible Unicode text, register a TrueType .ttf font before adding pages:

document =
  PaperForge.new()
  |> PaperForge.register_font(
    :inter,
    path: "assets/fonts/Inter-Regular.ttf"
  )

You can also register a font from an in-memory binary:

document =
  PaperForge.register_font(
    document,
    :inter,
    data: File.read!("assets/fonts/Inter-Regular.ttf")
  )

Then use the registered key in text operations:

Page.text(
  page,
  "El pingüino comió camarón — ¿listo? — Привет — Ω",
  x: 72,
  y: 720,
  font: :inter,
  size: 18
)

Embedded TrueType fonts are written as PDF Type 0 fonts with a CIDFontType2 descendant, Identity-H encoding, a /FontFile2 stream, widths from the TTF hmtx table, and a /ToUnicode CMap so text extraction and search can recover Unicode characters.

Supported embedded font input:

• TrueType outlines (.ttf)
• Unicode cmap format 4 or 12
• 8-bit and Unicode text strings supported by the font's glyph coverage

Current limitations:

• OpenType CFF (OTTO) and TrueType Collection (.ttc) files are not supported yet.
• Complex text shaping is not performed, so scripts such as Arabic, Devanagari, and advanced ligatures may not render as expected.
• PaperForge physically subsets embedded TrueType programs while preserving original glyph identifiers, and rebuilds affected glyf, loca, hmtx, and maxp tables with valid checksums.
• A missing glyph raises PaperForge.FontError unless a registered full-run fallback is configured through PaperForge.font_fallback/3.

Font Families

Register related TrueType files as a family:

document =
  PaperForge.new()
  |> PaperForge.register_font_family(
    :inter,
    regular: [path: "assets/fonts/Inter-Regular.ttf"],
    bold: [path: "assets/fonts/Inter-Bold.ttf"],
    italic: [path: "assets/fonts/Inter-Italic.ttf"],
    bold_italic: [path: "assets/fonts/Inter-BoldItalic.ttf"]
  )

Then select a variant with :weight and :style:

Page.text(
  page,
  "Important",
  x: 72,
  y: 720,
  font: :inter,
  weight: :bold,
  style: :italic
)

Set a document default font when most text should use the same font:

document =
  PaperForge.new()
  |> PaperForge.register_font(:inter, path: "assets/fonts/Inter-Regular.ttf")
  |> PaperForge.default_font(:inter)

Shapes

Lines

Page.line(
  page,
  x1: 72,
  y1: 700,
  x2: 300,
  y2: 700,
  width: 2,
  color: Color.rgb255(40, 70, 140)
)

Rectangles

Page.rectangle(
  page,
  x: 72,
  y: 560,
  width: 220,
  height: 100,
  fill: true,
  stroke: true,
  fill_color: Color.rgb255(235, 240, 250),
  stroke_color: Color.rgb255(40, 70, 140),
  line_width: 2
)

Circles

Page.circle(
  page,
  x: 400,
  y: 610,
  radius: 50,
  fill: true,
  stroke: true,
  fill_color: Color.rgb255(245, 180, 70),
  stroke_color: Color.rgb255(120, 70, 20),
  line_width: 2
)

PaperForge approximates circles using four cubic Bezier curves because PDF does not provide a native circle operator.

Colors

RGB values can be expressed from 0 to 1:

Color.rgb(1.0, 0.0, 0.0)

Or from 0 to 255:

Color.rgb255(255, 0, 0)

Grayscale helpers:

Color.gray(0.5)
Color.black()
Color.white()

Images

Page.image/3 accepts a supported image binary or a file path.

png = File.read!("logo.png")

page
|> Page.image(png, x: 72, y: 120, width: 200)
|> Page.image("photo.jpg", x: 72, y: 360, width: 200, height: 120)

When only one dimension is supplied, PaperForge preserves the source aspect ratio:

Page.image(page, "logo.png", x: 72, y: 120, width: 200)
Page.image(page, "logo.png", x: 72, y: 120, height: 80)

Supported JPEGs:

• grayscale
• RGB
• CMYK

Supported PNGs:

• non-interlaced 8-bit grayscale
• non-interlaced 8-bit RGB
• non-interlaced 8-bit grayscale with alpha
• non-interlaced 8-bit RGBA

PNG alpha is written as a PDF soft mask (/SMask). PNG grayscale/RGB images without alpha use the original compressed IDAT data directly with /FlateDecode and PNG predictor decode parameters. JPEG image data is embedded directly with /DCTDecode.

Images are deduplicated by SHA-256 hash, so drawing the same image several times does not embed duplicate image streams.

Unified Flow

PaperForge.flow/2 builds a document from layout blocks instead of manual page operations. The engine measures blocks, paginates them, calculates total pages, and then renders the final pages.

alias PaperForge.Flow

{document, report} =
  PaperForge.new()
  |> PaperForge.page_template(
    :report,
    size: :a4,
    margins: [top: 72, right: 54, bottom: 72, left: 54],
    header: "Quarterly report",
    footer: "Page {page} of {total}"
  )
  |> PaperForge.layout(
    fn flow ->
      flow
      |> Flow.heading("Quarterly report", level: 1)
      |> Flow.paragraph("Summary text that wraps and splits across pages.")
      |> Flow.list(["Revenue", "Expenses", "Cash"], type: :ordered)
      |> Flow.table(
        ["Metric", "Value"],
        [
          ["Revenue", "$120K"],
          ["Margin", "24%"]
        ],
        repeat_header: true
      )
      |> Flow.separator()
      |> Flow.page_break()
      |> Flow.section(:appendix, [title: "Appendix"], fn section ->
        section
        |> Flow.paragraph("Section content receives section metadata.")
      end)
    end,
    template: :report
  )

The report returned by PaperForge.layout/3 contains page count, block count, placements, warnings, and rendered page values. Placements include block ID, block type, page number, coordinates, dimensions, and section metadata:

{document, report} =
  PaperForge.layout(document, flow_function, template: :report)

report.pages
report.blocks
report.placements

Pagination options can be set on flow blocks:

flow
|> Flow.heading("Appendix", page_break_before: true, keep_with_next: true)
|> Flow.paragraph("This paragraph should stay visually connected.")
|> Flow.separator(page_break_after: true)

Sections group related content under a stable section ID. A section can add a title heading, start or end with page breaks, switch to a named page template, and pass section metadata into PageContext:

flow
|> Flow.section(:appendix, [title: "Appendix", template: :appendix], fn section ->
  section
  |> Flow.paragraph("Appendix content")
end)

Page templates can configure page geometry and reusable header/footer content:

document =
  PaperForge.new()
  |> PaperForge.page_template(
    :appendix,
    size: :letter,
    orientation: :landscape,
    margins: [top: 60, right: 48, bottom: 60, left: 48],
    header: fn page, context ->
      Page.text(page, "Appendix", x: context.content_left, y: 24)
    end,
    footer: "Page {page} of {total}"
  )

Custom blocks receive the current Page and PageContext:

Flow.custom(flow, fn page, context ->
  Page.text(
    page,
    "Page #{context.page_number} of #{context.total_pages}",
    x: context.content_left,
    y: context.content_top
  )
end, height: 24)

Debug reports summarize the generated document:

PaperForge.debug(document,
  show_margins: true,
  show_blocks: true,
  show_page_breaks: true
)

Existing Page, add_flow/4, and add_table/4 APIs remain supported for compatibility. New applications should prefer PaperForge.Flow and PaperForge.layout/3.

Legacy Flow And Page-level APIs

The APIs in this section remain supported for compatibility. New applications should prefer PaperForge.Flow and PaperForge.layout/3.

Flow text blocks across pages:

document =
  PaperForge.new()
  |> PaperForge.add_flow(
    [
      "First paragraph with enough text to wrap.",
      "Second paragraph. PaperForge creates new pages as needed."
    ],
    [size: :letter, margins: 72],
    font: :helvetica,
    size: 11,
    line_height: 15,
    gap: 8
  )

Get flow overflow information:

{document, report} =
  PaperForge.layout_flow(
    PaperForge.new(),
    ["A long paragraph", "Another long paragraph"],
    [size: :letter, margins: 72],
    header: "Quarterly report",
    footer: "Generated by PaperForge",
    keep_together: true
  )

report.pages_added
report.overflow?

Draw a basic table:

page =
  page
  |> Page.table(
    [
      ["Name", "Score"],
      ["Ana", 10],
      ["Luis", 9]
    ],
    x: Page.content_left(page),
    y: 96,
    width: Page.content_width(page),
    header: true
  )

Add a URI link annotation:

page =
  page
  |> Page.text("Project", x: 72, y: 720)
  |> Page.link(
    "https://github.com/Manuel1471/paper_forge",
    x: 72,
    y: 700,
    width: 180,
    height: 24
  )

Create internal navigation:

document =
  PaperForge.new()
  |> PaperForge.add_page(fn page ->
    page
    |> Page.destination(:intro, y: 720)
    |> Page.bookmark("Introduction", y: 720)
    |> Page.text("Introduction", x: 72, y: 720)
  end)
  |> PaperForge.add_page(fn page ->
    page
    |> Page.text("Back to intro", x: 72, y: 720)
    |> Page.link_to(:intro, x: 72, y: 700, width: 120, height: 24)
  end)

Add a paginated table with repeated headers:

document =
  PaperForge.add_table(
    document,
    rows,
    [size: :a4, margins: 72],
    repeat_header: true,
    row_split: :keep
  )

Metadata

document =
  PaperForge.new()
  |> PaperForge.metadata(
    title: "Reporte de Mexico",
    author: "Manuel Garcia",
    subject: "Informacion \u65E5\u672C\u8A9E",
    keywords: ["report", "elixir", "pdf"],
    creator: "PaperForge",
    producer: "PaperForge",
    creation_date: DateTime.utc_now(),
    modification_date: DateTime.utc_now()
  )

Metadata is written into the PDF Info dictionary and referenced from the document trailer. Latin-1-compatible strings are stored as PDF literal strings. Other Unicode strings are stored as UTF-16BE hexadecimal strings.

Binary Output

PaperForge can return the complete PDF as a binary:

pdf_binary =
  PaperForge.to_binary(document)

This can be used in Phoenix or Plug responses:

conn
|> put_resp_content_type("application/pdf")
|> put_resp_header(
  "content-disposition",
  ~s(attachment; filename="document.pdf")
)
|> send_resp(200, PaperForge.to_binary(document))

Write to disk:

PaperForge.write(document, "document.pdf")
PaperForge.write!(document, "document.pdf")

Architecture

PaperForge separates public drawing operations from low-level PDF objects.

PaperForge
|-- Document
|   |-- object allocation
|   |-- font registry
|   |-- image registry
|   `-- metadata reference
|-- Page
|   `-- high-level drawing operations
|-- Flow
|   `-- block-based document layout builder
|-- Layout
|   |-- Block
|   |-- Engine
|   `-- two-pass pagination and rendering
|-- PageCompiler
|   |-- coordinate transforms
|   |-- font registration
|   |-- image registration
|   `-- resource dictionaries
|-- Graphics
|   |-- Text
|   |-- TextBox
|   |-- Line
|   |-- Rectangle
|   |-- Circle
|   `-- Image
|-- Serializer
|   `-- Elixir values to PDF syntax
`-- Writer
    |-- PDF header
    |-- indirect objects
    |-- cross-reference table
    |-- trailer
    `-- EOF marker

The generated PDF uses traditional cross-reference tables. Tests verify that xref offsets point to the start of their corresponding indirect objects.

Examples

Examples write their generated files under tmp/. Start with the smallest example that matches the feature being evaluated:

Run an example from the project root:

mix run examples/linkedin_document_showcase.exs

The versioned paper_forge_0_4_showcase.exs and paper_forge_0_5_showcase.exs files remain available as historical migration references.

Development

Clone the repository:

git clone git@github.com:Manuel1471/paper_forge.git
cd paper_forge

Run the test suite:

mix test

Compile with warnings treated as errors:

mix compile --warnings-as-errors

Format the source code:

mix format

Run all checks:

mix do format, compile --warnings-as-errors, test

Run the TrueType and Unicode benchmark script:

mix run benchmarks/truetype.exs

Scope Boundaries

• The unified layout engine is the preferred rendering path for new applications, while page-level APIs remain available for compatibility.
• PaperForge 1.x targets standard PDF authoring and Unicode text supported by embedded TrueType fonts. Complex-script shaping and bidirectional layout are outside the current compatibility contract.
• Widow and orphan control uses configurable minimum line counts. Optional hyphenation deterministically breaks overlong words rather than consulting language dictionaries.
• PaperForge emits standard PDF 1.4 through 1.7 files. PDF/A output profiles are outside the current compatibility contract.

Performance Envelope

PaperForge 1.1 adds bounded process-local caches for repeated text metrics and Flate-compressed streams. Font metrics use stable font identities, and the PDF writer accumulates serialized objects linearly before producing the final binary. Documents without page-aware contents or references paginate once; navigation-aware documents retain bounded multi-pass convergence. Caches require no server process, remain isolated between render processes, and cannot grow without a fixed limit.

Latency benchmarks

Latency profiles answer: How long does one document take to generate?

They render increasingly large table-based reports with 25, 500, and 5,000 rows. These are the appropriate figures for estimating the latency and memory cost of an individual business document.

Run reproducible small, medium, and large profiles in the production environment:

MIX_ENV=prod SAMPLES=10 WARMUPS=2 mix run benchmarks/render_profiles.exs

Use PROFILE=small, PROFILE=medium, or PROFILE=large to isolate one scale. For release measurements, increase SAMPLES to 30. Each sample runs in a fresh process after explicit garbage collection.

The benchmark reports median, p95, minimum, and maximum values for layout, cold serialization, warm serialization, total time, peak process memory, reductions, garbage collections, and reclaimed words. It also records the Elixir version, OTP release, scheduler count, Mix environment, PDF size, page count, and cache hits. benchmarks/document_scale.exs remains available for comparison with the original 5,000-row baseline. Results are reference measurements rather than fixed guarantees across machines.

The following measurements compare the unchanged 1.0.0 commit with 1.1.0 using the same runner, 2 warmups, 10 samples, Elixir 1.20.2, OTP 29, 10 schedulers, and MIX_ENV=prod:

For the 5,000-row profile, median layout fell from 756.72 ms to 717.66 ms, reductions from 67.1 million to 56.8 million, and garbage collections from 196 to 177. Peak process memory remained effectively flat near 81.0 MB. The generated document remained byte-for-byte identical at 179 pages and 616,018 bytes.

Concurrent Rendering

PaperForge can render independent documents concurrently with bounded demand, per-document isolation, timeouts, cancellation, and runtime metrics. Phoenix applications should own the task supervisor:

children = [
  {Task.Supervisor, name: MyApp.PDFSupervisor}
]

Consume a lazy, backpressured stream without loading every result in advance:

MyApp.PDFSupervisor
|> PaperForge.Concurrent.stream(invoices, &InvoicePDF.render/1,
  max_concurrency: System.schedulers_online(),
  timeout: 30_000,
  ordered: false,
  job_id: fn invoice, _index -> invoice.id end
)
|> Stream.each(fn
  %{status: :ok, value: pdf, id: id} ->
    Storage.put_pdf(id, pdf)

  %{status: status, error: error, id: id} ->
    Logger.error("PDF #{id} failed with #{status}: #{inspect(error)}")
end)
|> Stream.run()

PaperForge.Concurrent.run/3 collects a finite batch and creates a temporary supervisor when the caller does not provide one. start_job/4 returns a task handle that can be passed to cancel/2. Every result reports duration, reductions, and process-local garbage collections.

The concurrency limit bounds active jobs and provides backpressure through the lazy input stream. Errors and timeouts affect only their originating document. Font, image, text-metric, and compression state remains immutable or process-local, avoiding shared mutable cache locks between renders.

Common production options:

Successful and failed jobs return a PaperForge.Concurrent.Result with :status, :id, :index, :attempts, :duration_us, :reductions, :garbage_collections, :peak_memory_bytes, and either :value or :error. Possible statuses are :ok, :error, :timeout, and :resource_limit.

Scalability benchmarks

Scalability profiles answer: How does the runtime behave as concurrent demand increases?

The bundled concurrency benchmark deliberately generates minimal, single-page PDFs. It measures scheduling overhead, backpressure, task isolation, and the effect of worker limits. Its throughput must not be interpreted as the expected throughput for invoices, image-heavy reports, embedded fonts, or large tables.

Run the 1,000-document minimal-workload benchmark:

MIX_ENV=prod JOBS=1000 CONCURRENCY=1,10,20 \
  mix run benchmarks/concurrent_renders.exs

Minimal single-page workload only:

These numbers describe concurrency infrastructure overhead. They are not document-generation capacity claims for real invoices or reports.

Representative table-report workloads on the same Elixir 1.20.2, OTP 29, 10-scheduler development machine:

Medium workload: 100 documents, 500 rows and 18 pages each

Large workload: 20 documents, 5,000 rows and 179 pages each

Reproduce these workloads:

MIX_ENV=prod WORKLOAD=medium JOBS=100 CONCURRENCY=1,5,10 \
  mix run benchmarks/concurrent_renders.exs

MIX_ENV=prod WORKLOAD=large JOBS=20 CONCURRENCY=1,5,10 \
  mix run benchmarks/concurrent_renders.exs

Higher concurrency improves batch throughput but increases per-document latency and total VM memory. For the measured large workload, moving from one to ten workers improves throughput by about 4.2x while increasing peak BEAM memory from roughly 308 MB to 1.86 GB. Choose the worker limit according to the application's latency target, memory budget, queue depth, and schedulers.

Production workloads can additionally set max_memory_bytes, max_reductions, max_attempts, retry_delay, and retry_on. File-oriented jobs should prefer PaperForge.write/2, which now serializes objects incrementally to a temporary file and atomically renames the finished PDF.

See PRODUCTION.md for staged rendering, optional Oban integration, distributed-node strategies, failure recovery, very large documents, and deployment sizing.

Telemetry

PaperForge emits stable Telemetry events for standalone rendering and concurrent production workloads:

Measurements include duration, bytes, memory, reductions, and gc. Batch completion also includes jobs. Metadata includes pages, output, status, id, index, attempt, status counts, ordering, and the configured concurrency limit where applicable.

duration follows Telemetry convention and uses native time units. Convert it for display with:

System.convert_time_unit(duration, :native, :millisecond)

Example metrics definitions:

import Telemetry.Metrics

[
  distribution("paperforge.render.stop.duration",
    event_name: [:paperforge, :render, :stop],
    measurement: :duration,
    unit: {:native, :millisecond},
    tags: [:status, :output]
  ),
  counter("paperforge.render.exception.count",
    event_name: [:paperforge, :render, :exception]
  ),
  distribution("paperforge.batch.job.duration",
    event_name: [:paperforge, :batch, :job],
    measurement: :duration,
    unit: {:native, :millisecond},
    tags: [:status, :attempt]
  ),
  sum("paperforge.batch.complete.jobs",
    event_name: [:paperforge, :batch, :complete],
    measurement: :jobs
  )
]

PaperForge emits standard Telemetry data and does not require a particular reporter. Applications can connect these definitions to Prometheus, OpenTelemetry, StatsD, or their existing observability stack.

For production capacity planning, replace the renderer in benchmarks/concurrent_renders.exs with a representative application document and keep the same JOBS and CONCURRENCY matrix. Report both categories:

• Document latency: median and p95 for representative small, medium, and large documents.
• Concurrent scalability: elapsed batch time, throughput, per-job median and p95, failures, and timeouts at each worker limit.

See API.md for the public compatibility policy and MIGRATING.md for the 0.6-to-1.0 upgrade guide.

Production Hardening

See MIGRATING.md for the stable 1.0 compatibility contract.

Future Enhancements

These ideas are not required by, or promised as part of, the stable 1.x API. They describe possible directions for later minor and major releases.

International Typography

• Connect complex-script shaping, bidirectional paragraph layout, and glyph-level font fallback after harfbuzz_ex exposes numeric glyph IDs. PaperForge will consume that public API without adding a private NIF or an installation-time native build.
• Add language dictionaries for discretionary hyphenation. The current hyphenate: true option only breaks overlong words deterministically.
• Establish PDF/A output profiles. VeraPDF remains an optional external validator and is not required to install or run PaperForge.

Declarative Authoring

• Introduce a versioned .paperforge template format backed by a safe, declarative document model.
• Support validated variables, loops, conditions, reusable components, and resource references without evaluating arbitrary Elixir code.
• Use the same format as the interchange contract for a future visual document designer.

Distributed Generation

• Optional adapters for durable job systems beyond the documented Oban integration pattern.
• Cross-node admission control and cluster-wide resource budgets.
• Resumable generation for documents that exceed a single job's memory or execution window.
• Compiled declarative templates after a safe .paperforge format exists.

HTML And CSS

• HTML parsing
• CSS cascade and computed styles
• HTML/CSS to PaperForge Layout IR
• Paged-media layout
• Flexbox and advanced document styling

Project Status

PaperForge 1.1 is suitable for production PDF authoring within the documented scope. Public APIs listed in API.md follow Semantic Versioning throughout the 1.x series. Runtime installation remains pure Elixir and does not require native compilation or external rendering services.

Contributing

Contributions, bug reports, architecture discussions, and PDF examples are welcome.

Before opening a pull request:

mix format
mix compile --warnings-as-errors
mix test

License

PaperForge is available under the terms specified in the LICENSE.