🦉🫥 PDF Anonymizer Core

This package provides the core functionality for the PDF/Text anonymizer, including text extraction, LLM-driven anonymization, and deanonymization logic. It is used by pdf-anonymizer-cli.

Installation

Install the base package with your favorite package manager:

pip install pdf-anonymizer-core

To use a specific LLM provider, you must install the corresponding extra. This helps to keep the installation lightweight by only downloading the libraries you need.

• Google: pip install "pdf-anonymizer-core[google]"
• Ollama: pip install "pdf-anonymizer-core[ollama]"
• Hugging Face: pip install "pdf-anonymizer-core[huggingface]"
• OpenRouter: pip install "pdf-anonymizer-core[openrouter]"
• OpenAI: pip install "pdf-anonymizer-core[openai]"
• Anthropic: pip install "pdf-anonymizer-core[anthropic]"

You can also install multiple extras at once:

pip install "pdf-anonymizer-core[google,ollama]"

Full documentation, including the Python SDK guide, CLI reference, recipes, and architecture, lives at leo-gan.github.io/anonymizer/ (especially the Recipes & Common Workflows section).

Environment Variables

The core library itself does not load .env files. Environment variables must be loaded by the application that uses this library (e.g., pdf-anonymizer-cli) or set in your shell.

• GOOGLE_API_KEY: Required when using Google models.
• HUGGING_FACE_TOKEN: Required when using Hugging Face models.
• OPENROUTER_API_KEY: Required when using OpenRouter models.
• OPENAI_API_KEY: Required when using OpenAI models.
• ANTHROPIC_API_KEY: Required when using Anthropic models.
• OLLAMA_HOST: Optional, defaults to http://localhost:11434 when using Ollama models.
• ANONYMIZER_MAPPING_KEY: Optional passphrase for encrypted mapping files.
• ANONYMIZER_FAKE_SECRET: Optional seed for the fake operator.

API Usage

anonymize_file()

Anonymizes a single file and returns the anonymized text and a mapping of original entities to their placeholders.

from pdf_anonymizer_core.core import anonymize_file
from pdf_anonymizer_core.prompts import detailed

# Example of programmatic usage
# detailed.prompt_template also asks the model to hide identity clues:
# phrases that point to one person without writing their name
# (for example "the CEO of Tesla"). Use simple.prompt_template to skip that.
text, mapping = anonymize_file(
    file_path="/path/to/file.pdf",
    prompt_template=detailed.prompt_template,
    model_name="gemini-2.5-pro"  # Can also be a new model like "google/gemini-flash-latest"
)

if text and mapping:
    print("Anonymized Text:", text)
    print("Mapping:", mapping)

deanonymize_file()

Reverts anonymization using a mapping file. The function writes the restored document and a stats file to conventional locations and returns their paths.

from pdf_anonymizer_core.utils import deanonymize_file

# Assumes you have an anonymized file and a mapping file
deanonymized_file_path, stats_file_path = deanonymize_file(
    "path/to/anonymized.md",
    "path/to/mapping.json",
    # mapping_passphrase="secret",  # required for *.mapping.json.enc
    # expected_source_sha256="...",  # optional AAD check
)

print("Deanonymized file:", deanonymized_file_path)
print("Stats file:", stats_file_path)

Configuration

You can import default configurations, profiles, and available models from the conf module.

from pdf_anonymizer_core.conf import (
    DEFAULT_MODEL_NAME,
    ModelName,
    PromptEnum,
    ConfigProfile,
    get_config_for_profile,
    DEFAULT_REGEX_PATTERNS,
)

print(f"Default model: {DEFAULT_MODEL_NAME}")
print(f"Available Google models: {[m.value for m in ModelName if m.provider == 'google']}")
print("Regex first-stage covers:", sorted(DEFAULT_REGEX_PATTERNS.keys())[:10], "...")

# Recommended way to obtain bundled settings (used by the CLI's --config-profile / -p)
cfg = get_config_for_profile(ConfigProfile.BEST_SPEED)
print(cfg.model_name, cfg.chunk_size)

The CLI exposes --config-profile (short: -p) with the values best-quality, best-speed (default), and best-cost.
These profiles control model, prompt, chunk size, overlap, retries, etc. You can override individual values when calling get_config_for_profile(...) or when using the CLI.

The first-stage regex (hybrid NER) is now powered by the RE2 engine (google-re2 package).
DEFAULT_REGEX_PATTERNS contains 70+ patterns partitioned by country (ISO-2 suffixes) covering emails, phones, URLs, credit cards, crypto, IBAN/BIC, VIN, MAC, IPv4/6, dates plus national IDs, tax IDs, driver licences, VAT/business numbers, passports, medical licenses etc. for 30+ countries (mandatory: US, CA, GB, ES, IT, FR, IN, CN + many others). After a structural match, validators.py runs a cheap checksum (Luhn, IBAN mod-97, VIN check digit, and a few national IDs). Failures are kept and labeled TYPE_LIKE (for example IBAN_LIKE_1) so a mistyped number is still hidden. Types with no check are unchanged. Listing IBAN in a type filter also includes IBAN_LIKE. See conf.py, regex_ner.py, and validators.py.

Pass keep_list= / deny_list= phrase lists into anonymize_file (CLI: --keep-list / --deny-list). Keep wins if a phrase is on both.

Pass seed_mapping= (original → written) into anonymize_file so the same person stays PERSON_1 across documents. The CLI --mapping-in flag loads that file (including encrypted maps).

EntityProfile.HIPAA_SAFE_HARBOR is a coverage aid for Safe Harbor identifier classes (year-only dates, ZIP3, age 90+). It is not a compliance certificate.

Per-type operators (replace, mask, hash, generalize, shift, fake) go in operators={"PERSON": "fake"} on anonymize_file. Pass fake_secret= so the same person always gets the same fake.

Replacement is span-based: mentions are located in the full document, the longer interval wins on overlap, and slices are written from the end.

To score a gold fixture (mention vs entity recall, direct vs quasi) run uv run python scripts/eval_tab.py. Tests and scripts only.

To score identity-clue clumps in masked text (report only):

from pdf_anonymizer_core.risk import assess_linkage_risk, write_risk_report

report = assess_linkage_risk(anonymized_text)
write_risk_report(report, "data/anonymized/note.anonymized.md")

To scan a masked string for leftovers (report only, no rewrite):

from pdf_anonymizer_core.verify import verify_anonymized_text, write_residual_report

report = verify_anonymized_text(anonymized_text)
write_residual_report(report, "data/anonymized/note.anonymized.md")

To keep only some countries' national-ID regexes (plus every universal pattern):

from pdf_anonymizer_core.conf import filter_regex_patterns, get_config_for_profile, ConfigProfile

only_us_gb = filter_regex_patterns(["US", "GB"])
cfg = get_config_for_profile(ConfigProfile.BEST_SPEED, countries=["US", "GB"])

See Also

• Main Documentation — Full project guides and 101 course.
• CLI Package README — Command line interface usage.
• Recipes & Common Workflows — Practical SDK examples (including profiles, caching, and advanced usage).
• API Reference (auto) — Detailed function signatures.
• Troubleshooting — Common issues when using the core library.
• Architecture — How the anonymization pipeline works internally.