🔒 piast-gate

Privacy-first LLM gateway — anonymize PII before it leaves your system.

Important

This is a MVP Version — Currently supporting only Google Gemini API and Polish-language prompts.

---

How It Works

piast-gate sits between your app and the model. It strips sensitive data before sending, then restores it after the response. The model never sees real PII.

Example

Input:

Mam na imię Jan Kowalski, mój email to jan@example.com, a PESEL: 85010112345

Sent to LLM:

Mam na imię <PERSON_1>, mój email to <EMAIL_1>, a PESEL: <PESEL_1>

Returned to client:

Mam na imię Jan Kowalski, mój email to jan@example.com, a PESEL: 85010112345

---

Quick Start

The recommended way to run this project is using uv.

1. Installation

Using uv (fast & recommended):

git clone https://github.com/your-org/piast-gate.git
cd piast-gate
uv sync

Using standard pip:

python -m venv .venv
.venv\Scripts\activate        # Windows alternate: source .venv/bin/activate
pip install -e ".[dev]"

2. Configure

cp .env.example .env

Edit the .env file and provide your GEMINI_API_KEY.

3. Run

Using uv:

uv run uvicorn main:app --workers 4

Using standard python:

uvicorn main:app --workers 4

4. Tests

uv run pytest     # using uv
pytest            # using venv

5. API Keys Configuration

Example configuration in .env:

LLM_PROVIDER=gemini
GEMINI_API_KEY=your_api_key_here
MODEL_NAME=gemini-2.5-flash
PL_NER_MODEL_NAME=pl_core_news_lg
RATE_LIMIT_PER_MINUTE=60
API_KEYS=["your-secret-key"]

Usage

curl -X POST http://localhost:8000/v1/api/chat \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer your-secret-key" \
  -d '{
    "model": "piast-gate",
    "messages": [
      {"role": "user", "content": "Mam na imię Jan Kowalski, email: jan@example.com"}
    ]
  }'

Response:

{
  "response": "Cześć Jan Kowalski! Jak mogę Ci pomóc?"
}

---

Performance

Benchmarks run with uvicorn main:app --workers 4.

By message length

Characters	avg	min	median	p(90)	p(95)
1290	391 ms	144 ms	236 ms	339 ms	367 ms
12900	1.46 s	839 ms	1.47 s	1.78 s	1.86 s
64500	7.32 s	5.45 s	7.33 s	8.70 s	9.11 s

By placeholder count

Placeholders	avg	min	median	p(90)	p(95)
3	162 ms	97 ms	164 ms	182 ms	186 ms
15	177 ms	102 ms	179 ms	198 ms	206 ms
60	260 ms	127 ms	257 ms	323 ms	333 ms
300	809 ms	562 ms	805 ms	916 ms	946 ms