redact

A local command-line tool that redacts personally identifiable information (PII) in common document formats. It replaces names, email addresses, ID codes, and usernames with consistent placeholders, and writes a separate mapping file that can restore the originals at any time.

All detection patterns live in a plain TOML config file — adding a new pattern requires no Python knowledge.

---

What it does

Given an input document, the tool produces two files:

File	Purpose
report_redacted.docx	Redacted copy — safe to share
report_mapping.json	Pseudonym → original map — keep this secure

The mapping file is the only way to reverse the process. Store it separately from the redacted document.

---

Detected PII (default config)

Type	Example input	Replaced with
Email address	alice@university.ac.uk	person001@anon.invalid
Person name	Alice Johnson	Person001
7- or 8-digit ID	1234567 / 12345678	0000001 / 00000001 (same width)
8-char alphanumeric code / username	ab12cd34	user0001xx

• The same original value always gets the same pseudonym within a file (consistent replacement).
• Names use spaCy NER when available, falling back to a title-case word heuristic.
• All patterns are defined in redact_config.toml and can be extended without touching the Python code.

---

Supported formats

Format	Extension	Redacted output
Microsoft Word	.docx	.docx
OpenDocument Text	.odt	.odt
Microsoft Excel	.xlsx	.xlsx
PDF	.pdf	.txt (plain text; PDF formatting cannot be preserved)

---

Installation

1. Clone the repository

git clone https://github.com/sharpic/redact.git
cd redact

2. Install Python dependencies

pip install python-docx odfpy openpyxl pdfplumber

3. Install spaCy for better name detection (recommended)

pip install spacy
python -m spacy download en_core_web_sm

Without spaCy the tool falls back to a title-case word heuristic for names, which is less accurate but requires no additional dependencies.

Python version: 3.11 or later required (uses the built-in tomllib parser).

---

Usage

Redact a document

python3 redact.py <input_file>

Output files are created alongside the input:

report.docx            ← original (unchanged)
report_redacted.docx   ← redacted copy
report_mapping.json    ← reversal key

Redact with custom output paths

python3 redact.py data.xlsx -o data_safe.xlsx -m data_keys.json

Redact a PDF

python3 redact.py notes.pdf
# → notes_redacted.txt  (plain text — PDF formatting is not preserved)

Restore originals

The tool auto-locates the mapping file from the filename:

python3 redact.py report_redacted.docx --restore
# → report_redacted_restored.docx

Or point to the mapping file explicitly:

python3 redact.py report_redacted.docx -r -m /secure/report_mapping.json

Use a custom config file

python3 redact.py report.docx --config my_patterns.toml

Disable spaCy (faster, regex-only name detection)

python3 redact.py report.docx --no-spacy

Full option reference

positional arguments:
  input               Input file (.docx / .odt / .xlsx / .pdf)

options:
  -o, --output FILE   Output path  (default: <input>_redacted.<ext>)
  -m, --mapping FILE  Mapping JSON (default: <input>_mapping.json)
  -c, --config FILE   Config TOML  (default: redact_config.toml)
  --no-spacy          Use regex heuristic instead of spaCy for names
  -r, --restore       Reverse mode: restore originals from mapping file

---

Adding new patterns

Open redact_config.toml and append a [[patterns]] block. No Python changes are needed.

Example: UK National Insurance numbers

[[patterns]]
name        = "ni_number"
regex       = '\b[A-CEGHJ-PR-TW-Z]{2}\d{6}[A-D]\b'
regex_flags = ["IGNORECASE"]
template    = "NI{n:04d}"

Example: UK phone numbers

[[patterns]]
name     = "phone_uk"
regex    = '\b(?:0|\+44)\s*\d[\d\s]{8,12}\d\b'
template = "PHONE{n:03d}"

Example: UK postcodes

[[patterns]]
name        = "postcode"
regex       = '\b[A-Z]{1,2}\d[A-Z\d]?\s*\d[A-Z]{2}\b'
regex_flags = ["IGNORECASE"]
template    = "PC{n:03d}"

Pattern fields reference

Field	Type	Description
name	string	Identifier shown in stats and the mapping file
regex	string	Python regex — use single quotes to keep \b etc. literal
spacy_ner	bool	Use spaCy NER instead of (or in addition to) the regex
spacy_labels	list	NER entity labels to match (default: ["PERSON"])
template	string	Pseudonym format — see below
exclusions	list	Words that disqualify a match (case-sensitive)
min_word_length	int	Minimum character length per word in a multi-word match
regex_flags	list	re module flag names, e.g. ["IGNORECASE"]

Template format

Template value	Result
"Person{n:03d}"	Person001, Person002, …
"person{n:03d}@anon.invalid"	person001@anon.invalid, …
"preserve_length"	Counter zero-padded to the same width as the original
"{orig_len}-char-{n}"	Embeds original length, e.g. 7-char-1

TOML tip: always use single quotes ('...') for regex strings. Double-quoted strings in TOML process escape sequences, turning \b into a backspace character rather than a word boundary.

---

How it works

1. The config is loaded and each [[patterns]] block is compiled into a PatternDef.
2. Each page / paragraph / cell of the document is scanned against all patterns in config order.
3. When a span is claimed by a pattern, no later pattern can overlap it — so email beats name for John.Smith@example.com.
4. Each unique original value is assigned exactly one pseudonym (stored in PseudonymRegistry). The same value always maps to the same pseudonym within a file.
5. The mapping is written to _mapping.json as { pseudonym: original }.
6. Restoration reads the mapping and applies plain string replacements in longest-first order.

---

Limitations

• DOCX / ODT: paragraph-level formatting (styles, spacing, alignment) is preserved. Per-run character formatting (bold/italic on individual words) may be collapsed within replaced spans.
• PDF: output is plain text. The redacted file loses all layout and formatting.
• Name detection without spaCy: the title-case heuristic catches consecutive capitalized words, which can produce false positives in documents with many heading-style phrases. Install spaCy for significantly better accuracy.
• Cross-cell PII: a value split across multiple cells or paragraphs will not be detected. PII must appear within a single text unit.
• No guarantee of completeness: redaction is a best-effort process. Review the output and the stats before sharing a file.

---

Running the tests

python3 -m pytest redact_tests.py -v

116 tests covering templates, span overlap, exclusions, registry consistency, round-trip redaction/restoration for all file formats, and spaCy NER.

---

Project structure

redact.py              Main script
redact_config.toml     Pattern definitions (edit this to extend)
redact_tests.py        Test suite
README.md              This file

---

Disclaimer

This tool is provided as a convenience utility. It is not a certified anonymization or de-identification solution and makes no guarantees of completeness. Always review the redacted output before sharing, especially in contexts subject to GDPR, HIPAA, or other data-protection regulations.

The direction and requirements were provided by the user; all code was generated by Claude (Anthropic). Developers should review and validate the script against their own data and requirements before relying on it in production.