Youarebot Classifier

FastAPI service that receives chat messages, stores them in PostgreSQL, and returns is_bot_probability for the whole dialog.

The public contract is intentionally small:

• GET /health returns service status.
• POST /predict accepts one message and returns a prediction for the dialog.

Stack

• Python 3.12
• uv
• FastAPI + Uvicorn
• Pydantic v2
• PostgreSQL
• Transformers zero-shot classification pipeline

The default model is typeform/distilbert-base-uncased-mnli, a lightweight DistilBERT MNLI model used through the Transformers zero-shot classification pipeline.

Local Run

Install dependencies:

uv sync

Start PostgreSQL locally, for example through Compose:

docker compose up -d postgres

Run the API:

DB_HOST=localhost uv run uvicorn src.main:app --reload

The service listens on http://127.0.0.1:8000.

Check health:

curl http://127.0.0.1:8000/health

Docker Compose

Run the full stack:

docker compose up --build

Compose exposes the API as http://localhost:443 on the host and keeps the container listening on port 8000.

API

GET /health

{
  "status": "ok"
}

POST /predict

Request:

{
  "text": "Hi, how are you?",
  "dialog_id": "0f2d4682-d939-4af6-beb9-880a5da202a2",
  "id": "13ab658b-8463-4410-a747-33ea2dfb7b68",
  "participant_index": 0
}

Response:

{
  "id": "814c9637-5f4a-4633-84d2-f86c7fa62a48",
  "message_id": "13ab658b-8463-4410-a747-33ea2dfb7b68",
  "dialog_id": "0f2d4682-d939-4af6-beb9-880a5da202a2",
  "participant_index": 0,
  "is_bot_probability": 0.42
}

is_bot_probability is validated by Pydantic and must stay in [0, 1].

Configuration

Environment variables:

DB_USER=student
DB_PASSWORD=student_pass
DB_HOST=postgres
DB_PORT=5432
DB_NAME=chat_db
MODEL_NAME=typeform/distilbert-base-uncased-mnli

The model is loaded lazily on the first /predict call.