Custom Python

The python provider allows you to use a Python script as an API provider for evaluating prompts. This is useful when you have custom logic or models implemented in Python that you want to integrate with your test suite.

Configuration

To configure the Python provider, you need to specify the path to your Python script and any additional options you want to pass to the script. Here's an example configuration in YAML format:

providers:
  - id: 'python:my_script.py'
    label: 'Test script 1'   # Optional display label for this provider
    config:
      additionalOption: 123

Python script

Your Python script should accept a prompt, options, and context as arguments. It should return a JSON-encoded ProviderResponse.

• The ProviderResponse must include an output field containing the result of the API call.
• Optionally, it can include an error field if something goes wrong, and a tokenUsage field to report the number of tokens used.

Here's an example of a Python script that could be used with the Python provider, which includes handling for the prompt, options, and context:

# my_script.py
import json

def call_api(prompt, options, context):
    # The 'options' parameter contains additional configuration for the API call.
    config = options.get('config', None)
    additional_option = config.get('additionalOption', None)

    # The 'context' parameter provides info about which vars were used to create the final prompt.
    user_variable = context['vars'].get('userVariable', None)

    # The prompt is the final prompt string after the variables have been processed.
    # Custom logic to process the prompt goes here.
    # For instance, you might call an external API or run some computations.
    output = call_llm(prompt)


    # The result should be a dictionary with at least an 'output' field.
    result = {
        "output": output,
    }

    if some_error_condition:
        result['error'] = "An error occurred during processing"

    if token_usage_calculated:
        # If you want to report token usage, you can set the 'tokenUsage' field.
        result['tokenUsage'] = {"total": token_count, "prompt": prompt_token_count, "completion": completion_token_count}

    return result

Types

The types passed into the Python script function and the ProviderResponse return type are defined as follows:

class ProviderOptions:
    id: Optional[str]
    config: Optional[Dict[str, Any]]

class CallApiContextParams:
    vars: Dict[str, str]

class TokenUsage:
    total: int
    prompt: int
    completion: int

class ProviderResponse:
    output: Optional[str]
    error: Optional[str]
    tokenUsage: Optional[TokenUsage]
    cost: Optional[float]
    cached: Optional[bool]
    logProbs: Optional[List[float]]

Setting the Python executable

In some scenarios, you may need to specify a custom Python executable. This is particularly useful when working with virtual environments or when the default Python path does not point to the desired Python interpreter.

Here's an example of how you can override the Python executable using the pythonExecutable option:

providers:
  - id: 'python:my_script.py'
    config:
      pythonExecutable: /path/to/python3.11

Troubleshooting

Viewing python output

If you use print statements in your python script, set LOG_LEVEL=debug to view script invocations and output:

LOG_LEVEL=debug npx promptfoo@latest eval

Setting the Python binary path

If you are using a specific Python binary (e.g. from a virtualenv or poetry), set the PROMPTFOO_PYTHON environment variable to be the binary location.

Also note that promptfoo will respect the PYTHONPATH. You can use this to tell the python interpreter where your custom modules live.

For example:

PROMPTFOO_PYTHON=venv/bin/python3.9 PYTHONPATH=/usr/lib/foo npx promptfoo@latest eval