Skip to content

nvisycom/elide

Repository files navigation

Elide

Build

Composable, multimodal toolkit for detecting and redacting sensitive data.

Elide is a Rust toolkit for finding and removing PII and PHI from text, images, audio, and tabular data. It provides the building blocks (recognizers, deduplication, validation, redaction operators, and format codecs) that a consumer wires into their own document-processing flow. Elide is the toolkit layer only; the orchestrating runtime and gateway server live in separate projects.

Warning

Active development: API not stable. This project is under active development. Public APIs, configuration shapes, and on-disk formats may change without notice between releases. Pin a specific commit if you depend on this in production.

Features

  • Multimodal: detect and redact across text, images, audio, and tabular data through one entity model; OCR and speech-to-text lift images and audio into a text-recognizable surface so the same recognizers apply
  • Pattern detection: regex, dictionary, and checksum recognizers find structured PII and PHI across many common formats and jurisdictions
  • Model-driven recognition: NER with language detection, and LLM/VLM recognizers (text and image) behind a pluggable backend, alongside the pattern recognizers
  • Context-aware scoring: nearby keywords lift the confidence of ambiguous matches, so weak findings clear the threshold only when their surroundings support them
  • Deduplication: overlapping findings from multiple recognizers reconcile into a single set of entities, with conflict resolution and confidence calibration
  • Redaction operators: mask, replace, hash, or encrypt text; blur or black-box image regions; silence or beep audio; drop tabular rows and columns, and more. Reversible options (encrypt, pseudonymize) record what is needed to restore the original
  • Format codecs: read, edit, and write documents (text, JSON, HTML, DOCX, PDF, images, audio, CSV, and more) with faithful round-tripping that changes only the redacted parts
  • Provenance-first model: every entity carries its full audit trail of how it was found, scored, and hidden

Everything is feature-gated: take only the modalities, recognizers, and codecs you need.

Documentation

See docs/ for architecture, security, and API documentation.

Changelog

See CHANGELOG.md for release notes and version history.

License

Apache 2.0 License, see LICENSE.txt

Support

About

Composable redaction framework for building PII detection and removal pipelines over unstructured data. Built-in recognizers, custom rules, and ML models to analyze and anonymize multimodal documents.

Topics

Resources

License

Contributing

Stars

5 stars

Watchers

0 watching

Forks

Releases

No releases published

Contributors

Languages