Presidio

Presidio is an open-source data protection SDK for detecting, redacting, masking, and anonymizing personally identifiable information across text, images, and structured data. It combines NLP-based named entity recognition (spaCy, Stanza, Hugging Face transformers), pattern matching, checksums, and custom recognizers in a pluggable Analyzer, with a separate Anonymizer that replaces, masks, hashes, or encrypts the detected spans using configurable operators. The image redactor handles OCR text and DICOM medical images, and the structured module covers tabular columns. It is the standard way to put a PII scrubbing layer in front of LLM or RAG pipelines without vendor lock-in.

Microsoft developed Presidio in the open starting in 2018, and in 2026 the project transitioned to community-led governance under the independent Data Privacy Stack organization, with Microsoft's support and the MIT license unchanged; Docker images moved to ghcr.io/data-privacy-stack. With roughly 11,350 GitHub stars and releases within days of this writing, it is the most adopted open-source PII anonymization framework available. Microsoft's own documentation is candid that Presidio does not guarantee finding all sensitive information, so it belongs as one layer in a defense-in-depth strategy with human review for compliance-critical flows, not as a standalone guarantee.

1. Install via pip: pip install presidio-analyzer presidio-anonymizer (plus presidio-image-redactor / presidio-structured as needed), or use the Docker images.
2. Run the Analyzer on your text to get detected PII spans with entity types and scores.
3. Pass the spans to the Anonymizer with your chosen operators (replace, mask, hash, encrypt).
4. Add custom PatternRecognizers for internal ID formats, and wire the pipeline before LLM calls and after responses.