Allow setting default params in config?

[vkryukov]

Right now, there is an asymmetry between params and config in Instructor.chat_completion, which makes it harder to switch between different providers.

For example, if I need to send a message to Claude, ChatGPT, and Gemini, respectively, I would need to use slightly different code:

# Claude
Instructor.chat_completion(
    max_token: 4096, # Unique for Claude
    model: "claude-3-5-sonnet-latest", 
    messages: [...]
)

# Gemini
Instructor.chat_completion(
    mode: json_schema, # Unique for Gemini
    model: "gemini-exp-1114", 
    messages: [...]
)

# ChatGPT
Instructor.chat_completion(
    model: "gpt-4o",
    messages: [...]
)

Ideally, I would just set these parameters (including even the default model) in my config.exs:

# config/config.exs
config :instructor,
  # adapter: Instructor.Adapters.Anthropic,
  adapter: Instructor.Adapters.Gemini,
  anthropic: [
    api_key: System.fetch_env!("ANTHROPIC_API_KEY"),
    max_tokens: 4096,
    model: "claude-3-5-sonnet-latest"
  ],
  gemini: [
    api_key: System.fetch_env!("GOOGLE_API_KEY"),
    mode: :json_schema,
    model: "gemini-exp-1114"
  ],
  openai: [
    api_key: System.fetch_env!("OPENAI_API_KEY"),
    model: "gpt-4o-mini"
  ]

then my application code could simply become Instructor.chat_completion(messages: [...] and I can switch to a different provider by only changing :adapter.

If @thmsmlr you are aligned with this approach, I can send a pull request.

[thmsmlr]

I'm torn on this.

On the one hand, I totally see your point. It would be nice to set a default params as part of the config, where you could default the mode, the model, really any set of parameters.

However, in practice, when you're switching between different models, you often have to tweak your prompts as well. And so while I think superficially it's useful, in practice it won't yield the results you actually want.

My recommendation is to just create a wrapper function for the time being. When more concrete use cases arise, maybe we can revisit.