usage.rst

Using the library

There are primarily two ways you can use prompttools in your LLM workflow:

1. Run experiments in notebooks.
2. Write unit tests and integrate them into your CI/CD workflow via Github Actions.

Notebooks

There are a few different ways to run an experiment in a notebook.

The simplest way is to define an experimentation harness and an evaluation function:

from prompttools.harness import PromptTemplateExperimentationHarness


def eval_fn(prompt: str, results: Dict, metadata: Dict) -> float:
    # Your logic here, or use a built-in one such as `prompttools.utils.similarity`.
    pass

prompt_templates = [
    "Answer the following question: {{input}}",
    "Respond the following query: {{input}}"
]

user_inputs = [
    {"input": "Who was the first president?"},
    {"input": "Who was the first president of India?"}
]

harness = PromptTemplateExperimentationHarness("text-davinci-003",
                                               prompt_templates,
                                               user_inputs)


harness.run()
harness.evaluate("metric_name", eval_fn)
harness.visualize()  # The results will be displayed as a table in your notebook

If you are interested to compare different models, the ModelComparison example may be of interest.

For an example of built-in evaluation function, please see this example of semantic similarity comparison for details.

You can also manually enter feedback to evaluate prompts, see HumanFeedback.ipynb.

Note: Above we used an ExperimentationHarness. Under the hood, that harness uses an Experiment to construct and make API calls to LLMs. The harness is responsible for managing higher level abstractions, like prompt templates or system prompts. To see how experiments work at a low level, see this example.

Unit Tests

Unit tests in prompttools are called prompttests. They use the @prompttest annotation to transform an evaluation function into an efficient unit test. The prompttest framework executes and evaluates experiments so you can test prompts over time. To turn an evaluation function into a re-usable test, you can add the annotation like this:

@prompttest.prompttest(
    experiment=OpenAICompletionExperiment,
    model_name="text-davinci-003",
    metric_name="similar_to_expected",
    prompt_template="Answer the following question: {{input}}",
    user_input=[{"input": "Who was the first president of the USA?"}],
)
def evaluation_fn(
    input_pair: Tuple[str, Dict[str, str]], results: Dict, metadata: Dict
) -> float:

The evaluation functions should accept one of the following as it's parameters:

• input_pair: Tuple[str, Dict[str, str]], results: Dict, metadata: Dict
• prompt: str, results: Dict, metadata: Dict
• messages: List[Dict[str,str], results: Dict, metadata: Dict

You can see an example test here and an example of that test being used as a Github Action here.