tiktag AGENTS.md

AGENTS.md

Authoritative project contract for tiktag. Keep this file to contract, invariants, and real caveats only. README.md stays user-facing and short.

Working rules

• state assumptions when they matter; if ambiguity changes implementation, ask instead of guessing
• prefer smallest change that solves request; no speculative features, abstractions, or configurability
• touch requested scope only; do not refactor or clean unrelated code
• verify behavior-changing work with targeted checks before declaring done

Project

tiktag = text anonymizer.

• one Rust crate: library + thin CLI
• built-in model: Xenova/distilbert-base-multilingual-cased-ner-hrl (quantized ONNX)
• built-in model scope: PERSON, ORG, LOCATION
• regex recognizers are additive supplements after model inference
• current built-in regex recognizer: email

Library contract

use tiktag::{Tiktag, TiktagError, TiktagOutput};

let mut tiktag = Tiktag::new(&profiles_path)?;   // loads tokenizer + ONNX session once
let out: TiktagOutput = tiktag.anonymize(text)?; // reuses runtime per call
let text = &out.anonymization.anonymized_text;

• construct once, call many: new is expensive; anonymize reuses runtime
• anonymize takes &mut self; no internal locking
• multi-thread hosts use Mutex<Tiktag> or per-thread instance
• profiles_path is explicit
• relative model_dir resolves against profile file parent only
• TiktagOutput = anonymization + sequence_len + window_count
• errors are TiktagError; keep typed variants for profile, bundle, and inference boundaries
• placeholder numbering is stable per call only; no cross-document identity

CLI contract

• tiktag "<text>" or tiktag --stdin prints anonymized text to stdout with trailing newline
• tiktag --json emits machine-readable output without reversible metadata
• tiktag --debug-json emits reversible metadata for local/debug use only
• tiktag download fetches bundled model assets
• diagnostics and timing logs go to stderr via log
• JSON modes also emit stats.timings on stdout payload
• CLI resolves models/profiles.toml from app-data first, then legacy fallback: <exe_dir>/models/profiles.toml, then cwd
• no model/profile selection flags
• prefer --stdin for large inputs