🦉🫥 PDF Anonymizer CLI

A command-line interface for anonymizing PDF, Markdown, and plain text files using LLMs.

• High-Quality Anonymization: Leverages LLMs to identify and replace Personally Identifiable Information (PII) with high accuracy.
• Identity clues: With -p best-quality or --prompt-name detailed, the tool also hides phrases that point to one person without writing their name (for example "the CEO of Tesla"). The default best-speed profile does not.
• Checksum labels: After the fast number search, a real IBAN becomes IBAN_1 and a mistyped one becomes IBAN_LIKE_1. Both are hidden. Same idea for cards, VINs, and a few national IDs.
• Operators, reports, and lists: --operator TYPE=mask|hash|generalize|shift|fake, leftover scan (verify), linkage-risk score (report), --keep-list / --deny-list, --mapping-in, optional locked maps, HIPAA coverage aid. See the CLI History.
• Large File Support: Consistently anonymizes large files (tested up to 1GB).
• Multi-Provider & Cost-Effective: Free to use with local Ollama models. It also supports major providers like OpenAI, Anthropic, Google, Hugging Face, and OpenRouter.
• Reversible: Supports deanonymization to recover original data when needed.
• Multi-Format: Works with PDF, Markdown, and plain text files.

Installation

Install the CLI with your favorite package manager. To use a specific LLM provider, you must install the corresponding extra.

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

You can also install multiple extras at once:

pip install "pdf-anonymizer-cli[google,openrouter]"

This installs the pdf-anonymizer executable.

Environment Variables

The CLI will automatically load a .env file from the current directory or any parent directory. For consistency, it's recommended to place a single .env file at the root of the repository.

• GOOGLE_API_KEY: Required when using Google models.
• HUGGING_FACE_TOKEN: Required when using Hugging Face models. You can get a token from here.
• 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. Same as --mapping-passphrase (writes *.mapping.json.enc).
• ANONYMIZER_FAKE_SECRET: Optional. Seed for --operator TYPE=fake.

Example .env file:

GOOGLE_API_KEY="YOUR_API_KEY_HERE"
HUGGING_FACE_TOKEN="YOUR_HF_TOKEN_HERE"
OPENROUTER_API_KEY="YOUR_OPENROUTER_KEY"

Usage

Anonymize

The run command anonymizes one or more files.

pdf-anonymizer run FILE_PATH [FILE_PATH ...] \
  [-p | --config-profile {best-quality|best-speed|best-cost}] \
  [--characters-to-anonymize INTEGER] \
  [--prompt-name {simple|detailed}] \
  [--model-name TEXT] \
  [--anonymized-entities PATH] \
  [--countries US,GB] \
  [--verify / --no-verify] \
  [--verify-llm]

Arguments:

• FILE_PATH: Path to one or several PDF, Markdown, or text files for anonymization.

Options:

• -p, --config-profile {best-quality|best-speed|best-cost}: The configuration profile to use. Profiles bundle sensible defaults for model, prompt, chunk size, overlap, and retries (default: best-speed). Individual flags (--model-name, --prompt-name, --characters-to-anonymize) act as overrides on top of the chosen profile.
• --characters-to-anonymize INTEGER: Number of characters to process in each chunk (default: 100000; overrides profile).
• --prompt-name [simple|detailed]: The prompt template to use (default: detailed; overrides profile).
• --model-name TEXT: The language model to use (overrides profile).
• --anonymized-entities PATH: Path to a file with a list of entities to anonymize.
• --countries TEXT: ISO-2 country codes for national-ID regexes, comma-separated (e.g. US,GB). Universal patterns (email, IBAN, cards, …) always stay. Default: all countries.
• --verify / --no-verify: After masking, scan for leftovers (default: on). Writes data/stats/<stem>.residual_pii.json. Does not rewrite the file.
• --verify-llm: Also ask the language model to hunt for leftovers.
• --mapping-passphrase TEXT: Lock the mapping as *.mapping.json.enc. Also ANONYMIZER_MAPPING_KEY. Default: plaintext JSON.
• --operator TYPE=op: How to write a type (replace, mask, hash, generalize, shift, fake). Repeatable. Default is replace.
• --fake-secret TEXT: Seed for fake. Also ANONYMIZER_FAKE_SECRET.
• --risk / --no-risk: After masking, score identity-clue clumps (default: on). Writes data/stats/<stem>.risk.json.

pdf-anonymizer report FILE runs the same linkage-risk score on an already-masked file.

• --entity-profile hipaa-safe-harbor: coverage aid for the 18 Safe Harbor identifier classes (year-only dates, ZIP3, age 90+). Not a compliance certificate.
• --mapping-in PATH: reuse an existing mapping so the same person stays PERSON_1 across files. Files in one run share the map automatically.
• --keep-list PATH: phrases to leave visible (one per line).
• --deny-list PATH: phrases that must be hidden even if detection missed them.

pdf-anonymizer verify FILE runs the same leftover scan on an already-masked file.

Models: You can use any of the predefined models below, or specify a new model using the format "provider/model-name". For example: --model-name "google/gemini-flash-latest".

• Google: gemini-2.5-pro, gemini-2.5-flash (default), gemini-2.5-flash-lite.
• Ollama: gemma:7b, phi4-mini.
• Hugging Face: openai/gpt-oss-20b, mistralai/Mistral-7B-Instruct-v0.1, HuggingFaceH4/zephyr-7b-beta.
• OpenRouter: openai/gpt-4o, google/gemini-pro.
• OpenAI: gpt-4o, gpt-5.
• Anthropic: claude-4-sonet, claude-4.5-sonet.

Examples

Basic anonymization (uses the default best-speed profile):

pdf-anonymizer run document.pdf

High-quality run on an important document:

pdf-anonymizer run contract.pdf -p best-quality

Fast & cheap batch processing with a local model (override profile defaults):

pdf-anonymizer run notes/*.md -p best-cost --model-name "ollama/phi4-mini"

A new model (Google) and a simple prompt:

pdf-anonymizer run notes.md --model-name "google/gemini-flash-latest" --prompt-name simple

Using an OpenRouter model:

pdf-anonymizer run report.pdf --model-name "openai/gpt-4o"

Deanonymize

The deanonymize command reverts anonymization using a mapping file.

pdf-anonymizer deanonymize ANONYMIZED_FILE MAPPING_FILE [--mapping-passphrase TEXT]

Arguments:

• ANONYMIZED_FILE: Path to the anonymized text file.
• MAPPING_FILE: Path to the JSON mapping file (plaintext or *.mapping.json.enc).
• --mapping-passphrase: Required for an encrypted mapping. Also ANONYMIZER_MAPPING_KEY.

Example:

pdf-anonymizer deanonymize \
    data/anonymized/document.anonymized.md \
    data/mappings/document.mapping.json

This will create a deanonymized version of the file at data/deanonymized/document.deanonymized.md.

See Also

• Main Documentation — Full guides including the 101 course.
• Core Package README — Details on the underlying library.
• Recipes & Common Workflows — Practical CLI usage examples (profiles, batching, external workflows, etc.).
• CLI Reference (full) — Complete command reference on the docs site.
• Troubleshooting — Help with common CLI problems.
• Architecture — How the CLI and core work together.