Get started prompt engineering with local LLMs

Ollama is an application for running large-language models locally on your computer. It gives you access to open-source LLMs that you can prompt directly with the command line or an endpoint.

To getting started with local LLMs:

1. Download and install Ollama: https://ollama.com/download
2. When prompted, install the ollama CLI
3. Download and run your first LLM: ollama run llama2
4. Send your first prompt: “What is the chief end of man?”

The response will be printed to the console.

Using the CLI is nice, but a better option is to create and send prompts with a scripting language. I’m going to use Python and OpenAI’s chat completion API, since that a popular combination. For an example with JavaScript, see this documentation.

1. Create a new python file: touch completions.py

2. Install the openai package: pip3 install openai

3. Set up your OpenAI client:

   # completions.py
   from openai import OpenAI
   
   client = OpenAI(
        base_url="http://localhost:11434", # ll43a looks like llama
        api_key="ollama" # Unused but required
   )
   

4. Create your first completion:

   # This function is not required, but it's nice to have
   def get_completion(prompt, model="llama2", temperature=0.0):
        messages = [{"role": "user", "content": prompt}]
        response = client.chat.completions.create(
            model=model,
            messages=messages,
            temperature=temperature,
        )
        return response.choices[0].message.content
   
   response = get_completion("What is the chief end of man?")
   print(response)
   

5. Run your script: python3 completions.py

That’s all it takes! For a good introduction to prompt writing, I recommend DeepLearning.AI’s ChatGPT Prompt Engineering for Developers course.