kakushi 0.1.0

Japanese PII masking for LLM pipelines

kakushi 隠し

Japanese PII masking for LLM pipelines. Strip personal names, company names, phone numbers, emails, amounts, and my-numbers from Japanese business text before it leaves your server. Restore them in the model reply.

Live demo on HuggingFace Spaces · 日本語の説明

---

Why

You want to send a Japanese email, meeting note, or invoice to OpenAI / Anthropic / Gemini for summarisation, rewriting, or QA. The text contains customer names (田中太郎), company names (株式会社山田商事), deal amounts (¥1,500,000), and contact details. You don't want those bytes leaving your server.

kakushi gives you a three-line workaround.

from kakushi import mask, unmask

masked, mapping = mask(
    "田中太郎様、株式会社山田商事の件で¥1,500,000の見積書を [e-mail] にお送りします",
    dictionary={"株式会社山田商事": "取引先A"},   # optional — pin sensitive names
)
# masked  → "<PERSON_1>様、取引先Aの件で<AMOUNT_1>の見積書を <EMAIL_1> にお送りします"
# mapping → in-memory only. Never persisted.

reply = call_openai(masked)                   # your LLM call here
final = unmask(reply, mapping)                # placeholders → originals

Install

pip install kakushi          # regex layer only — fast, no model download
pip install kakushi[ner]     # + GiNZA Japanese NER (person / org / location)

What gets caught

Three layers, lowest priority first.

1. GiNZA NER — PERSON, ORG, LOCATION from the Japanese spaCy pipeline.
2. Regex — phone, postal code, email, my-number, yen amount, IPv4, driver's license, credit card, URL, contract / order numbers, and Japanese company names with 株式会社 / 合同会社 / (株) markers.
3. Company dictionary (optional) — {"株式会社山田商事": "取引先A"}. Highest priority. The value can be None, in which case kakushi auto-assigns <COMPANY_DICT_N>.

Overlapping spans are resolved with priority dict > regex > NER, with the longer span winning at the same priority.

Recall on synthetic corpus (100 cases, NER on)

Entity	Recall
AMOUNT	100%
EMAIL	100%
PERSON	100%
PHONE	100%
ORG	92%
LOCATION	60.9%
Overall	95.7%

Reproduce with python scripts/eval.py --n 100.

Known limitations

• Hiragana-leading company names (株式会社さくら通信, 合同会社みどり) are missed by the regex because including hiragana in the body would destroy the particle boundary. Register them via the company dictionary.
• Compound place names like 京都府京都市 are sometimes split by GiNZA.
• My-number checksums are not validated yet. Any 12-digit dddd-dddd-dddd pattern is matched.
• The mapping returned from mask() is in-memory only. Do not serialize it, log it, or write it to disk. It contains the originals you just removed.

Try it without writing code

The HuggingFace Space lets you paste any Japanese text, optionally provide a company dictionary, and see exactly what would be sent to the LLM. Nothing is logged.

Project layout

kakushi/
├── src/kakushi/
│   ├── normalizer.py     # NFKC + context-aware hyphen folding
│   ├── patterns.py       # 12 regex patterns (PII + Japanese company names)
│   ├── ner.py            # GiNZA wrapper (lazy load, optional)
│   ├── resolver.py       # span overlap resolution
│   ├── masker.py         # mask() / unmask() / company dictionary
│   ├── synth.py          # synthetic business-text generator
│   └── types.py          # EntityType, Match
├── tests/                # 70 cases, all green
├── scripts/eval.py       # recall measurement on synthetic corpus
├── examples/openai_demo.py
└── space/                # HuggingFace Space app (Gradio)

License

MIT.