Use containers for generative AI development

Table of contents

Prerequisites

Complete Containerize a generative AI application.

Overview

In this section, you'll learn how to set up a development environment to access all the services that your generative AI (GenAI) application needs. This includes:

Adding a local database
Adding a local or remote LLM service

Note
You can see more samples of containerized GenAI applications in the GenAI Stack demo applications.

You can use containers to set up local services, like a database. In this section, you'll update the compose.yaml file to define a database service. In addition, you'll specify an environment variables file to load the database connection information rather than manually entering the information every time.

To run the database service:

In the cloned repository's directory, rename env.example file to .env. This file contains the environment variables that the containers will use.
In the cloned repository's directory, open the compose.yaml file in an IDE or text editor.

In the compose.yaml file, add the following:

Add instructions to run a Neo4j database
Specify the environment file under the server service in order to pass in the environment variables for the connection

The following is the updated compose.yaml file. All comments have been removed.

services:
  server:
    build:
      context: .
    ports:
      - 8000:8000
    env_file:
      - .env
    depends_on:
      database:
        condition: service_healthy
  database:
    image: neo4j:5.11
    ports:
      - "7474:7474"
      - "7687:7687"
    environment:
      - NEO4J_AUTH=${NEO4J_USERNAME}/${NEO4J_PASSWORD}
    healthcheck:
      test: ["CMD-SHELL", "wget --no-verbose --tries=1 --spider localhost:7474 || exit 1"]
      interval: 5s
      timeout: 3s
      retries: 5

Note
To learn more about Neo4j, see the Neo4j Official Docker Image.

Run the application. Inside the docker-genai-sample directory, run the following command in a terminal.
$ docker compose up --build
Access the application. Open a browser and view the application at http://localhost:8000. You should see a simple Streamlit application. Note that asking questions to a PDF will cause the application to fail because the LLM service specified in the .env file isn't running yet.
Stop the application. In the terminal, press ctrl+c to stop the application.

Add a local or remote LLM service

The sample application supports both Ollama and OpenAI. This guide provides instructions for the following scenarios:

Run Ollama in a container
Run Ollama outside of a container
Use OpenAI

While all platforms can use any of the previous scenarios, the performance and GPU support may vary. You can use the following guidelines to help you choose the appropriate option:

Run Ollama in a container if you're on Linux, and using a native installation of the Docker Engine, or Windows 10/11, and using Docker Desktop, you have a CUDA-supported GPU, and your system has at least 8 GB of RAM.
Run Ollama outside of a container if you're on an Apple silicon Mac.
Use OpenAI if the previous two scenarios don't apply to you.

Choose one of the following options for your LLM service.

When running Ollama in a container, you should have a CUDA-supported GPU. While you can run Ollama in a container without a supported GPU, the performance may not be acceptable. Only Linux and Windows 11 support GPU access to containers.

To run Ollama in a container and provide GPU access:

Install the prerequisites.
- For Docker Engine on Linux, install the NVIDIA Container Toolkit.
- For Docker Desktop on Windows 10/11, install the latest NVIDIA driver and make sure you are using the WSL2 backend

Add the Ollama service and a volume in your compose.yaml. The following is the updated compose.yaml:

services:
  server:
    build:
      context: .
    ports:
      - 8000:8000
    env_file:
      - .env
    depends_on:
      database:
        condition: service_healthy
  database:
    image: neo4j:5.11
    ports:
      - "7474:7474"
      - "7687:7687"
    environment:
      - NEO4J_AUTH=${NEO4J_USERNAME}/${NEO4J_PASSWORD}
    healthcheck:
      test:
        [
          "CMD-SHELL",
          "wget --no-verbose --tries=1 --spider localhost:7474 || exit 1",
        ]
      interval: 5s
      timeout: 3s
      retries: 5
  ollama:
    image: ollama/ollama:latest
    ports:
      - "11434:11434"
    volumes:
      - ollama_volume:/root/.ollama
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: all
              capabilities: [gpu]
volumes:
  ollama_volume:

Note
For more details about the Compose instructions, see Turn on GPU access with Docker Compose.

Add the ollama-pull service to your compose.yaml file. This service uses the docker/genai:ollama-pull image, based on the GenAI Stack's pull_model.Dockerfile. The service will automatically pull the model for your Ollama container. The following is the updated section of the compose.yaml file:

services:
  server:
    build:
      context: .
    ports:
      - 8000:8000
    env_file:
      - .env
    depends_on:
      database:
        condition: service_healthy
      ollama-pull:
        condition: service_completed_successfully
  ollama-pull:
    image: docker/genai:ollama-pull
    env_file:
      - .env
  # ...

To run Ollama outside of a container:

Install and run Ollama on your host machine.
Update the OLLAMA_BASE_URL value in your .env file to http://host.docker.internal:11434.
Pull the model to Ollama using the following command.
$ ollama pull llama2

Important
Using OpenAI requires an OpenAI account. OpenAI is a third-party hosted service and charges may apply.

Update the LLM value in your .env file to gpt-3.5.
Uncomment and update the OPENAI_API_KEY value in your .env file to your OpenAI API key.

Run your GenAI application

At this point, you have the following services in your Compose file:

Server service for your main GenAI application
Database service to store vectors in a Neo4j database
(optional) Ollama service to run the LLM
(optional) Ollama-pull service to automatically pull the model for the Ollama service

To run all the services, run the following command in your docker-genai-sample directory:

$ docker compose up --build

If your Compose file has the ollama-pull service, it may take several minutes for the ollama-pull service to pull the model. The ollama-pull service will continuously update the console with its status. After pulling the model, the ollama-pull service container will stop and you can access the application.

Once the application is running, open a browser and access the application at http://localhost:8000.

Upload a PDF file, for example the Docker CLI Cheat Sheet, and ask a question about the PDF.

Depending on your system and the LLM service that you chose, it may take several minutes to answer. If you are using Ollama and the performance isn't acceptable, try using OpenAI.

Summary

In this section, you learned how to set up a development environment to provide access all the services that your GenAI application needs.

Related information:

Next steps

See samples of more GenAI applications in the GenAI Stack demo applications.