Philter

Philter is an application that redacts protected health information (PHI), personally identifiable information (PII), non-public personal information (NPPI), and other sensitive information from text. Philter processes plain text and PDF documents. Philter runs in your private cloud so your sensitive data never has to traverse the public internet. Use Philter's API to process text from virtually any system or process. Philter is open source. Learn more on GitHub.

Phileas (the open source library underneath Philter) came first, and Philter followed as the API layer on top. The toolkit has been continuously maintained since 2019 and predates the current wave of LLM-wrapper privacy tools. The detection engine is a hybrid of purpose-built NLP models and pattern matching, not a prompt sent to a third-party LLM. You can inspect the full commit history, every model, and every line of policy logic on GitHub.

No. Philter never transmits your text or documents to any third-party service. Philter can run in a firewalled (or even air-gapped) environment. For example, in AWS you can deploy Philter to a private subnet and use security groups and network ACLs to prevent any outbound traffic from the instance and its subnet, and we recommend doing so to strengthen your overall security posture.

Philter runs in your own environment and Philterd never receives your PHI, so there is no BAA to sign with us. This guide explains why.

Yes. Philter is open source under the Apache License, Version 2.0. Check out the code, file issues, and contribute on GitHub.

Philter detects many entity types and we add new ones regularly. Among them: Ages · Bitcoin Addresses · US Cities · US Counties · Credit Card Numbers · Custom Dictionaries · Custom Identifiers (e.g. medical record numbers, transaction numbers) · Dates · US Driver's License Numbers · Email Addresses · Hospital Names · IBAN Codes · IP Addresses (IPv4 and IPv6) · MAC Addresses · Passport Numbers · Persons' Names (fuzzy matching, first/last/whole) · Phone/Fax Numbers · Physician Names · SSNs and TINs · Shipping Tracking Numbers · US States · URLs · VINs · US Zip Codes

You create policies that tell Philter which types of PII and PHI to find and how to handle them. A policy lists the entity types (phone numbers, names, etc.), when to act on them, and how to redact them. You can have as many policies as you need and select which one to apply per request. Policies are documented in Philter's User's Guide.

Philter uses a variety of methods, including specially trained machine learning models. A Philter lens is a trained model for a specific kind of text. Using a lens that matches your data gives you more accurate results. Philter includes the English Names (Small) lens, which its default policy uses to detect names inside Philter, with no separate service. Larger English lenses are in the lens catalog; run one in PhEye and point your policy at it.

Philter can be deployed as a container or directly into your cloud from the AWS, Google Cloud, or Microsoft Azure marketplaces in just a few minutes. Deploying Philter in 5 minutes on the AWS Marketplace walks through the marketplace route end to end, from subscribing to the first redacted response. For containers, see the quick start.

Three options: 1. Call the HTTP API directly. Philter accepts text and returns the redacted text. Generate a client in any language from the Philter OpenAPI specification. 2. Use the Philter CLI for convenient command-line access. 3. Use the open source Java SDK.

Precision and recall depend greatly on your data. Every dataset is different, so quoting a single F1 score across users would be meaningless. If a vendor cites accuracy without seeing your documents, be very cautious. We'd rather take some representative text from you and spend a few days gathering precision and recall metrics specific to your data, then send those back to you alongside the redacted output.

Philter uses state-of-the-art NLP technology, which is fundamentally non-deterministic. Identification accuracy depends on how similar your text is to the training corpus, how the text is formatted, and how long it is. For that reason, it's important to assess Philter's performance on your own data before relying on it in production. Every detected entity has a confidence score between 0 and 100. The confidence condition in a filter strategy lets you tune detection. For example, confidence > 75 ignores entities the model isn't sure about and only redacts high-confidence matches.

Direct comparisons are tricky because Philter is designed differently. Philter goes beyond identification. It includes disambiguation, ignore lists, value replacement, and consistent pseudonymization out of the box. The other services may technically support some of these, but only if you build them yourself on top. Philter is also not a managed SaaS API. You deploy it into your environment and call its API over your own network, so your text never has to leave your perimeter. We think that's substantially more secure than handing sensitive text to a third-party API. Finally, Philter is flexible: you can use your own models and you have full control over the filtering pipeline. For a side-by-side breakdown against Amazon Comprehend, Google Cloud DLP, Microsoft Presidio, Private AI, and others, see our comparison page.

Philter is open source software under the Apache License, Version 2.0. The full Philter license agreement is published on this site, and the licensing details are also available on the project repository.