PII Management Prerequisites

import os

from pathlib import Path
from gliner import GLiNER
from transformers import AutoTokenizer

def download_model(model_path: Path, model_name: str) -> None:
    """Download a Gliner model to the specified directory."""
    # Check if the directory already exists
    if not os.path.exists(model_path):
        # Create the directory
        os.makedirs(model_path)

    # The default tokenizer for Gliner is DeBERTa v2, which results in
    # protobuf issues as we are using protobuf 5.29.5.
    # Use BERT tokenizer instead of DeBERTa v2.
    print(f"Downloading GLiNER model: {model_name}")
    print("Using BERT tokenizer to avoid protobuf issues...")

    try:
        # Load a BERT tokenizer that's compatible
        tokenizer = AutoTokenizer.from_pretrained("bert-base-uncased")

        # Load GLiNER model with custom tokenizer
        model = GLiNER.from_pretrained(model_name, tokenizer=tokenizer)
        model.save_pretrained(model_path)
        print(f"Successfully downloaded and saved model to {model_path}")

    except Exception as e:
        print(f"Error with BERT tokenizer: {e}")
        print("Trying with default tokenizer...")
        # Fallback to default tokenizer
        model = GLiNER.from_pretrained(model_name)
        model.save_pretrained(model_path)
        print(f"Successfully downloaded and saved model to {model_path}")

if __name__ == "__main__":
    local_model_path = Path("./gliner_model").resolve() # You can modify this path to your desired location
    download_model(
        model_path=local_model_path,
        model_name="urchade/gliner_multi_pii-v1"
    )

FROM rasa/rasa-pro:3.13.0 AS builder
USER root

# Install dependencies
RUN python3 -m venv /opt/venv &&  \
    . /opt/venv/bin/activate &&  \
    pip install --no-cache-dir -U "pip==24.*" &&  \
    pip install --no-cache-dir "rasa-pro[pii]===3.13.0" &&  \
    # pin transformers to avoid connection errors with their \
    # OpenTelemetry metrics collector occurring from >=4.53.0
    pip install --no-cache-dir "transformers==4.52.4"

# Download HF model
COPY ./gliner_model_download_script.py gliner_model_download_script.py
# Force slow tokenizers and use pure-Python protobuf only for the download step
# This is needed to avoid Protobuf 5 compatibility issues with the fast tokenizers
RUN TRANSFORMERS_NO_FAST_TOKENIZER=1 PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION=python python3 -m gliner_model_download_script

FROM builder
WORKDIR /app
COPY --from=builder /app/gliner_model /app/gliner_model
RUN --mount=type=bind,target=/app/gliner_model

# Update permissions
RUN chown -R 1001:0 /app/gliner_model && \
    chmod -R g=u /app/gliner_model && \
    chmod o+wr /app/gliner_model
USER 1001
# Check if the model files are downloaded correctly
RUN ls -l /app/gliner_model

x-license: &license
  RASA_LICENSE: ${RASA_LICENSE}
  OPENAI_API_KEY: ${OPENAI_API_KEY}
x-pii-config: &pii-env-vars
  USER_CHAT_INACTIVITY_IN_MINUTES: 15
  GLINER_MODEL_PATH: "/app/gliner_model"

services:
  pii_assistant:
    image: pii-assistant:latest
    container_name: pii_assistant
    volumes:
      - "${BOT_PATH}:/app/bot"
    working_dir: /app/bot
    entrypoint: ""
    command: rasa run -p 5005 --enable-api
    ports:
      - "5005:5005"
    networks:
      - default
    environment:
      <<: [ *license, *pii-env-vars ]
    healthcheck:
      test: curl localhost:5005 || exit 1
      interval: 10s
      retries: 10
      start_period: 15s
      timeout: 10s
    user: "rasa"