Add retry mechanism for openai API calls

[harshkumarchourasia]

Add retrying mechanism for openai api calls.

Sample use cases

1. complete = models.text_completion.openai("text-davinci-003", **{"wait_exponential_multiplier":1000, "wait_exponential_max":10000, "stop_max_attempt_number":5})
2. complete = models.text_completion.openai("text-davinci-003", **{"wait_random_min":1000, "wait_random_max":10000, "stop_max_attempt_number":5})

Closes #115

[brandonwillard]

Don't forget to set up pre-commit locally. Also, retrying needs to be added to the dependencies in pyproject.toml.

[arunpatro]

I would advise using https://pypi.org/project/backoff/ instead of retying. Backoff seems to have more stars, and I have seen multiple people using it like the official tree-of-thoughts too

See: https://github.com/ysymyth/tree-of-thought-llm/blob/9f0f986dc30604b83f8d07203cf09b745ada36aa/models.py#L18

[brandonwillard]

I would advise using https://pypi.org/project/backoff/ instead of retying. Backoff seems to have more stars, and I have seen multiple people using it like the official tree-of-thoughts too

See: https://github.com/ysymyth/tree-of-thought-llm/blob/9f0f986dc30604b83f8d07203cf09b745ada36aa/models.py#L18

Is there a technical/functional/support reason one of these libraries should be used over the other?

[arunpatro]

Nothing too technical as of my current knowledge. Just some thoughts based on seeing various implementations of retrying. I found another simple and intuitive library - https://github.com/invl/retry (600 stars). One reason to prefer this one or backoff (2000 stars) is that retrying has only 29 stars.

I feel using backoff or retry can create a more intuitive api extension, I can also try creating a PR to compare.

[brandonwillard]

Nothing too technical as of my current knowledge. Just some thoughts based on seeing various implementations of retrying. I found another simple and intuitive library - https://github.com/invl/retry (600 stars). One reason to prefer this one or backoff (2000 stars) is that retrying has only 29 stars.

I feel using backoff or retry can create a more intuitive api extension, I can also try creating a PR to compare.

Assuming stars as a proxy for something like quality or maturity of the libraries, and that the changes are trivial, they might as well be made here.

[rlouf]

I've always thought tenacity (a fork of retrying) was the de facto standard for this, and it seems to be very similar to backoff. If there is a technical reason to prefer backoff, let's use that. Otherwise we can stick with tenacity.

[rlouf]

I rebased the changes, modified the code to use tenacity and fixed a few bugs I noticed while making the changes. This is ready for review.

I used the parameters found in this file in the openai-python repo.