Calling LLMs

Basic LLM Calls

Execute prompts using large language models:

def llm_call():
    response = gl.nondet.exec_prompt("Answer this question")
    return response.strip().lower()
 
# Consensus ensures consistent LLM responses
answer = gl.eq_principle.strict_eq(llm_call)

JSON Response Format

Request structured responses from LLMs:

def structured_llm_call():
    prompt = """
    Return a JSON object with these keys:
    - "score": random integer from 1 to 100
    - "status": either "active" or "inactive"
    """
    return gl.nondet.exec_prompt(prompt, response_format='json')
 
result = gl.eq_principle.strict_eq(structured_llm_call)
score = result['score']  # Access JSON fields

This approach guarantees that exec_prompt returns a valid JSON object, however correspondence to the specified format depends on the underlying LLM.

Image Processing

Process images with vision models:

def vision_analysis():
    prompt = "Describe what you see in this image"
    image_data = b"\x89PNG..."
    return gl.nondet.exec_prompt(
        prompt,
        images=[image_data]
    )
 
description = gl.eq_principle.strict_eq(vision_analysis)

⚠️

Limit of images is two

Response Validation

Validate and process LLM responses:

def validated_call():
    response = gl.nondet.exec_prompt(
        "Generate a number between 1 and 100",
        response_format='json'
    )
 
    # Validate the response
    if response['number'] < 1 or response['number'] > 100:
        raise Exception(f"Invalid number: {response['number']}")
 
    return response['number']
 
result = gl.eq_principle.strict_eq(validated_call)

Non-determinism Web Access