Flow Control

When using models, it's frequently necessary to change your prompt based on the output of the model. This can be done using conditional logic in your code.

In this section, we will learn how to capture the output of a completion and use it guide the flow of the prompt.

How it works

Defining your prompt as code allows you to prompt and generate text according to conditional statements and loops.

When you call gen, you can capture the output of the model by assigning it to a variable. You can then use this variable to guide the flow of your code.

Examples

In this example we prompt the model to answer an existential question. If the model's output includes the number 42, we prompt the model to provide a more serious answer. Otherwise, we ask it to provide an answer that references a novel commonly mentioned when discussing the meaning of life.

prompt(`What is the meaning of life? Answer in a single sentence.\n`, { role: 'user' });
const output = gen({limit: 500, role: 'assistant'});
 
if (output.includes('42')) { 
    prompt(`\n\nCorrect! But can you tell me a more serious answer that doesn't reference Hitchiker's Guide to the Galaxy?\n\n`, { role: 'user' });
} else { 
    prompt(`\n\nIncorrect! I'm looking for an answer that references a novel commonly mentioned when discussing this topic.\n\n`, { role: 'user' });
}
 
gen({limit: 500, role: 'assistant'});

This technique can also be used in a loop to get the model to iteratively refine its output.

prompt(`What is the meaning of life? Answer in a single sentence.\n`, { role: 'user' });
 
const NUM_TRIES = 5; 
let tries = 0;
while (tries < NUM_TRIES) { 
    gen({limit: 500, role: 'assistant'});
    tries++;
 
    if (tries < NUM_TRIES) { 
        prompt(`\nCan you be more brief in your answer?\n`, { role: 'user' });
    }
}