0.4.0

acroforge

Turn flat PDFs into real, fillable AcroForms. Permissive (Apache-2.0), deterministic, zero-copyleft.

Left: a flat PDF - just printed lines and an empty box. Right: the same PDF after acroforge - real, fillable form fields, filled and rendered correctly. No Adobe, no cloud, no AGPL.

---

What it does

acroforge takes any PDF - vector or scanned - and injects real AcroForm fields at positions you specify. The result is a standards-compliant fillable PDF that renders correctly in Chrome's pdfium and Firefox's pdf.js.

Four operations:

All accept and return plain bytes, making them easy to compose in any pipeline.

---

Tested on real-world forms

The deterministic core (build / fill / flatten / read_fields) is validated against 125 real public PDF forms - IRS and other government forms (VA, OPM, GSA), CMS / Medicare and hospital healthcare forms, federal- and state-court legal forms, and SBA / USPTO / vendor business forms. It reads every one of them, and fills, flattens, and round-trips every fillable one (over 11,000 real fields) without a single crash. Every field type renders correctly in Chrome's pdfium and Firefox's pdf.js, golden-image tested in CI.

The detect() layer below is separate and clearly labeled best-effort.

---

Install

pip install acroforge

Or from source:

git clone https://github.com/san64777/acroforge
cd acroforge
pip install -e .   # or: uv pip install -e .

---

Python usage

import io
from reportlab.pdfgen import canvas  # any PDF source works
import acroforge as af
from acroforge import FieldSpec, FieldType

# --- Step 0: obtain a flat PDF (bytes) any way you like ---
buf = io.BytesIO()
c = canvas.Canvas(buf, pagesize=(612, 792))
c.drawString(72, 720, "Name:")
c.drawString(72, 680, "Agree to terms:")
c.save()
flat_pdf: bytes = buf.getvalue()

# --- Step 1: describe the fields you want ---
fields = [
    FieldSpec(
        type=FieldType.TEXT,
        page=0,
        rect=(200, 700, 450, 730),  # (x0, y0, x1, y1) in PDF points
        name="full_name",
    ),
    FieldSpec(
        type=FieldType.CHECKBOX,
        page=0,
        rect=(200, 660, 220, 680),
        name="agree",
        export_value="Yes",
    ),
]

# --- Step 2: inject the fields ---
fillable: bytes = af.build(flat_pdf, fields)

# --- Step 3: fill values ---
filled: bytes = af.fill(fillable, {"full_name": "Jane Doe", "agree": True})

# --- Step 4: flatten (optional - locks the form) ---
final: bytes = af.flatten(filled)

# Write to disk
with open("output.pdf", "wb") as f:
    f.write(final)

---

CLI usage

# 1. Inject fields described in a JSON manifest
acroforge build in.pdf manifest.json fillable.pdf

# 2. Fill fields from a JSON object {name: value}
acroforge fill fillable.pdf data.json filled.pdf

# 3. Flatten (bake and lock)
acroforge flatten filled.pdf final.pdf

Example manifest.json:

[
  {
    "type": "text",
    "page": 0,
    "rect": [200, 700, 450, 730],
    "name": "full_name"
  },
  {
    "type": "checkbox",
    "page": 0,
    "rect": [200, 660, 220, 680],
    "name": "agree",
    "export_value": "Yes"
  },
  {
    "type": "radio",
    "page": 0,
    "rect": [200, 620, 220, 640],
    "name": "plan",
    "options": ["basic", "pro", "enterprise"],
    "export_value": "pro"
  }
]

Example data.json:

{"full_name": "Jane Doe", "agree": true, "plan": "pro"}

---

Field types

FieldSpec reference

class FieldSpec(BaseModel):
    type: FieldType
    page: int                                    # 0-indexed
    rect: tuple[float, float, float, float]      # (x0, y0, x1, y1) in PDF points
    name: str                                    # AcroForm field name
    options: list[str] | list[tuple[str, str]] | None = None  # choice options (str or (export, label))
    maxlen: int | None = None                    # TEXT cap / COMB cell count
    export_value: str | None = None              # radio/checkbox on-value
    list_box: bool = False                       # CHOICE: False=dropdown, True=list box
    multi_select: bool = False                   # CHOICE list box: allow multiple selections
    editable: bool = False                       # CHOICE combo: accept free-typed text
    confidence: float = 1.0                      # 1.0 = explicit; <1.0 = best-effort guess

Dropdowns and list boxes (FieldType.CHOICE)

# dropdown (combo box)
FieldSpec(type=FieldType.CHOICE, page=0, rect=(200, 620, 360, 640),
          name="state", options=["CA", "NY", "TX"])

# (export, label) pairs: store "CA", display "California"
FieldSpec(type=FieldType.CHOICE, page=0, rect=(200, 580, 360, 600),
          name="st", options=[("CA", "California"), ("NY", "New York")])

# scrolling list box, multi-select
FieldSpec(type=FieldType.CHOICE, page=0, rect=(200, 500, 360, 570),
          name="langs", options=["en", "fr", "de"], list_box=True, multi_select=True)

All four variants - dropdown, single-select list box, editable dropdown, and multi-select list box - are cross-viewer verified: the selected value renders in both pdfium and pdf.js. read_fields recovers a choice field's structure (its options and the list_box / multi_select / editable flags); it does not recover the current selection, since a FieldSpec describes the field, not its filled value.

---

Detection (best-effort)

In addition to the deterministic engine, acroforge ships an optional, best-effort detector that guesses where fields belong on a flat vector PDF by reading its vector geometry and nearby text labels. It handles both common form archetypes:

• Underline forms - write-on rules become text fields.
• Table/grid forms - bordered table cells become text fields (label-aware: the field is placed in the writable area below the label, multi-column cells are split, and section-header rows are skipped).
• Checkboxes - both vector squares and font glyphs (☐ / ☑ / ☒).

import acroforge as af

pdf = open("form.pdf", "rb").read()

# Inspect candidate fields (a FormManifest); every field has confidence < 1.0
manifest = af.detect(pdf)
for f in manifest.fields:
    print(f.type, f.name, f.rect, f.confidence)

# Or go straight to a fillable PDF (detect() then build())
fillable: bytes = af.make_fillable(pdf)

CLI:

# Print the detected manifest as JSON (review it!)
acroforge detect form.pdf

# Detect and write a fillable PDF in one step
acroforge make-fillable form.pdf fillable.pdf

Read this before relying on it:

• Heuristic. Detection guesses from vector shapes and text proximity. It will miss fields and invent spurious ones.
• Vector-only. It reads the PDF's vector content stream. Scanned (image-only) PDFs are refused with ScannedPDFError - there is no OCR.
• Confidence-scored. Every detected FieldSpec carries confidence < 1.0 to flag it as a guess. Explicitly authored specs use confidence = 1.0.
• Meant to be reviewed. Treat the output of detect() / make-fillable as a draft manifest to inspect and correct, not a finished form.
• No accuracy claims. We make no promise about detection precision or recall on any form. Quality varies wildly by document.
• No AI. There are no models, no inference, no network calls - just deterministic geometry heuristics over the PDF's own vectors.

Reading existing fields

read_fields(pdf) ingests the AcroForm fields already present in a fillable PDF as FieldSpecs (real registered fields, so confidence = 1.0). It is the inverse of build, so the two round-trip:

import acroforge as af

specs = af.read_fields(open("fillable.pdf", "rb").read())   # -> list[FieldSpec]
for s in specs:
    print(s.type.value, s.name, s.rect)

# copy one form's field layout onto another PDF
af.build(other_pdf, af.read_fields(template_pdf))

(One FieldSpec per widget, with coordinates, type, name, and checkbox/radio on-states recovered. Dropdowns are reported as text. Pushbuttons are skipped.)

Removing fields

remove(pdf, names) deletes specific fields by the name read_fields reports, so the two compose. Handy when make_fillable over-detects, or to strip a field before sending a form:

specs = af.read_fields(pdf)
junk = [s.name for s in specs if s.type == af.FieldType.SIGNATURE]
clean = af.remove(pdf, junk)        # raises ValueError if any name is missing

Naming a radio group removes the whole group; removing the last field leaves an empty, re-usable /AcroForm.

Serializing a manifest

detect() returns a FormManifest and read_fields() returns list[FieldSpec] - both pydantic models, so store / send-to-a-UI / round-trip them with pydantic's built-ins (no extra API to learn):

data = manifest.model_dump_json()                  # -> JSON string
manifest = FormManifest.model_validate_json(data)  # -> back to a FormManifest
af.build(pdf, manifest.fields)                      # build from the (edited) specs

(export, label) option pairs round-trip as [export, label] arrays and back to tuples; generate a TypeScript type from FormManifest.model_json_schema().

---

Scope and honest limits

The reliable part is the deterministic build / fill / flatten engine. You supply field positions via FieldSpecs - acroforge injects, fills, and flattens them reliably at exactly the coordinates you give it, on any PDF (vector or scanned).

detect() / make_fillable() are the best-effort layer described above: use them to bootstrap a manifest, then review and hand off the corrected specs to the engine.

XFA / dynamic forms: some PDFs (many government forms) carry a dynamic XFA layer over the standard AcroForm. acroforge operates on the AcroForm layer - which is what most viewers render - and drops the XFA layer on output. Flattened output is unambiguous everywhere; for interactive output, an XFA-first viewer (some Adobe configurations) may prefer the dropped layer, so flatten the result if you need cross-Adobe fidelity.

There is no AI in this package, and no copyrighted form templates are bundled - bring your own PDFs.

---

Engine and dependencies

Runtime dependencies are strictly permissive:

Optional extras:

• [fallback] - adds pikepdf (MPL-2.0) as a fallback PDF writer; not required for the default engine path.
• [harness] - adds pypdfium2 + Pillow for cross-viewer visual regression tests.

No GPL, AGPL, LGPL, or SSPL in the runtime tree. CI enforces this on every push via pip-licenses --fail-on='GPL;AGPL;LGPL;SSPL'.

---

License

Apache-2.0. See LICENSE.

No copyrighted form templates are included or bundled. Bring your own PDFs.