Code Prompter

Prompt engineering can be a tedious task inaccurate tasks. However, with a formal language (such as python) the prompts are more strutured.

The output of an LLM optimized to predict code is also highly structured.

For example, for the following prompt we would expect the LLM to output a valid array

  assert extract_tags("https://www.argmaxml.com") == [

This simple OpenAI and AI21 wrapper structures the prompt as a python code, then it filters out only the outputs that are valid syntactically.

Example usages

Classification

Using AI21 j2-grande-instruct

    from ai21_prompter import ClassificationQuery
    cq = ClassificationQuery("ai21_apikey")
    
    print(cq.tag("Tinder", most_common=3))
    # output: [("Dating", 4), ("Social", 2), ("hookups", 1)]

    print(cq.classify("Gucci sweaters now on sale as stock drops", ["fashion", "sport", "finance"]))
    # output: [{"sport": True, "finance": False, "fashion":True}]
    

Inverting a function

Using OpenAI codex

    from codex_prompter import ExtrapolationQuery
    eq = ExtrapolationQuery("openai_secret")
    
    print(eq.extrapolate_function_value("abbreviate", {"User Id": "userid", "Document id": "docid",}, "time of day"))
    # output: ["ToD", "timeofday"]

    print(eq.reverse_extrapolate_function("abbreviate", {"User Id": "userid", "Document id": "docid",}, "dow"))
    # output: ["Day of week", "definition of work"]