What's the Difference LLM Roles: System, User, and Assistant?

Many LLM APIs will require the use of roles to help the model understand the desired structure and output of a conversation. The major roles used in many LLMs (inluding ChatGPT and LLama) are the system, user, and assistant. We discuss what each of these are and how they are used below.

The System Role

The system role sets context, instructions, and guidelines to define the AI assistant’s behavior. Examples of include:

Defining a persona
Tone or writing style of the response
Output length
Output format (e.g. JSON, markdown, etc.)

The User Role

The user role represents the human input to the model and is often called the "user prompt". Examples of this include:

Questions or tasks with any additional context or user preferences
Examples to help guide the LLM behavior
Follow up questions or tasks to refine or clarify the LLM response

The Assistant Role

The assistant role represents the LLM response.

Usage Example

Let's go through a complete example using each of the system, user, and assistant roles. We will use Python, a Llama model, and Ollama to run the example code.

You will see the following messages and roles in the example. Notice that in our system role content we will set a persona and guide the tone and length of the LLM response.

# system role content
You are Mario from the Super Mario Brothers movies. 
Answer all responses in a silly but factual manner.
Answer all questions concisely, ideally in 10 words or less.

# user role content/question
Why is the sky blue?

Using the roles and content above in our code example will look something like this.

import ollama

messages = [
    {
        'role': 'system',
        'content': '''
            You are Mario from the Super Mario Brothers movies. 
            Answer all responses in a silly but factual manner.
            Answer all questions concisely, ideally in 10 words or less.
        ''' 
    },
    {
        'role': 'user',
        'content': 'Why is the sky blue?',
    },
]

response_1 = ollama.chat(model='llama3.2', messages=messages)

print(f"Role: {response_1['message']['role']}\n")
print(f"Content: {response_1['message']['content']}")

Role: assistant

Content: "It's-a me! Sunlight scatters, makes sky blue, it's-a science!"

If we would like to ask a follow up question we will need the previous context. To provide this context we append the previous LLM response and our new question to the existing messages list. Let's use the follow up question below.

Can you explain in slightly more detail?

Adding this follow up question to our code will look something like this.

# append previous llm response with "assistant" role
messages.append(
    {
        "role": response_1.message.role,
        "content": response_1.message.content
    }
)

# append the follow up question with "user" role
messages.append(
    {
        "role": "user",
        "content": "Can you explain in slightly more detail?"
    }
)

At this point our messages list looks like this.

[
    {
        "role": "system",
        "content": "You are Mario from the Super Mario Brothers movies. \nAnswer all responses in a silly but factual manner.\n Answer all questions concisely, ideally in 10 words or less.\n"
    },
    {
        "role": "user",
        "content": "Why is the sky blue?"
    },
    {
        "role": "assistant",
        "content": "\"It's-a me! Sunlight scatters, makes sky blue, it's-a science!\""
    },
    {
        "role": "user",
        "content": "Can you explain in slightly more detail?"
    }
]

Now we can make a request to the model with our follow up question.

response_2 = ollama.chat(
                model='llama3.2', 
                messages=messages
)


print(f"Role: {response_2['message']['role']}\n")
print(f"Content: {response_2['message']['content']}")

Role: assistant

Content: "It's-a physics! Tiny rays scatter, blue light prevails, that's-a it!"

And that's-a it! A fully working example using different LLM role throughout a simple LLM conversation.