Translate max_tokens_to_sample for OpenAI

[scottcmoore]

While attempting to use the semantic chunker with OpenAI, I ran into this issue:

[2] pry(main)> url = "https://www.scottcmoore.com"
data = Langchain::Loader.load(url).value
llm = LlmFactory.create!("OpenAI")
Langchain::Chunker::Semantic.new(
  data,
  llm: llm
DEPRECATED: `Langchain::LLM::OpenAI#complete` is deprecated, and will be removed in the next major version. Use `Langchain::LLM::OpenAI#chat` instead.
ArgumentError: unknown keyword: :max_tokens_to_sample

OpenAI expects this parameter as 'max_tokens', while other providers appear to expect 'max_tokens_to_sample'. This PR updates the OpenAI chat method to accept either version of the parameter.
Let me know what you think of this approach, thanks!

[codenamev]

This looks good, although should start thinking about unifying the LLM api for this reason.

[andreibondarev]

@scottcmoore Which LLM are you using if you don't mind me asking? If instead of this we dropped the max_tokens_to_sample: 50000 from this line, would it still work for you?

The way I'm thinking about this is that OpenAI sets the "gold-standard" in terms of API interfaces and we adapt all other LLM provider interfaces to match it.

In practice this looks like the following... Every Langchain::LLM::Base subclass implements the following interface:

def chat(
  messages: [],
  model: defaults[:chat_completion_model_name],
  max_tokens: nil
  ...
)

But the underlying calls, depending on the LLM class, may need to remap to their respective maxTokens or max_tokens_to_sample, etc.

I won't claim that this is implemented everywhere with 100% accuracy but this is where I'd like it to converge towards.

[scottcmoore]

@andreibondarev I was using gpt-3.5-turbo when testing this.
If the max_tokens_to_sample: 50000 were dropped I think it would work fine for our use case.

Just to make sure, am I understanding correctly that you'd like to keep the OpenAI provider interface as it is, and update the others to accept max tokens as a param to their chat methods, and do any necessary translation of parameter names in those providers?