Crate tiktag 

Text anonymization library backed by a built-in multilingual NER model.

tiktag loads one bundled profile and exposes a small library surface: construct Tiktag once, then call Tiktag::anonymize for each input. The built-in pipeline uses the Xenova distilbert-base-multilingual-cased-ner-hrl model for person, org, and location entities, then applies additive regex recognizers such as email.

Example

use std::path::Path;

use tiktag::Tiktag;

let mut tiktag = Tiktag::new(Path::new("models/profiles.toml"))?;
let out = tiktag.anonymize("Maria Garcia from OpenAI visited Berlin.")?;
println!("{}", out.anonymization.anonymized_text);

Notes

• Tiktag::new is the expensive step: it loads the profile, tokenizer, and ONNX runtime session.
• Tiktag::anonymize takes &mut self and reuses that loaded runtime.
• Placeholder numbering is stable within one call only.
• Model-based anonymization can miss entities; treat tiktag as an assistive control, not a sole safety boundary.

Structs

AnonymizationResult

Complete anonymization payload for one input document.

Profiles

Parsed built-in profile configuration before resolution.

Replacement

One accepted replacement span in the original text.

ResolvedProfile

The built-in model config after validation and path resolution.

Tiktag

Reusable anonymizer instance backed by one loaded profile and ONNX session.

TiktagOutput

Result of one Tiktag::anonymize call.

Enums

PlaceholderFamily

Placeholder family written into anonymized output.

TiktagError

Public error type for the tiktag library surface.

Constants

BUILTIN_PROFILE_NAME

Logical name of the built-in bundled profile.

REQUIRED_MODEL_FILES

Relative file paths required in a valid model bundle directory.

Functions

missing_model_files

Returns required bundle files that are missing from model_dir.

validate_model_bundle

Validates that model_dir contains the files required by the built-in bundle layout.

Type Aliases

TiktagResult

Library-wide Result alias that defaults to TiktagError.