Support Complex Parameter Schemas

[jayelkaake]

Summary

Adds the ability to define more complex param schemas, while retaining the existing simple methods.

This new format is also forward-compatible once LLMs start supporting more JSON schema structures (and at the pace they are moving that might not be too long from now).

Bumped version to 1.3 since this is kind of a major feature.

Examples

# Existing structure still works and now is called 'shorthand' syntax:
param :latitude, desc: "Latitude (e.g., 52.5200)"

# But now you can treat the params as plain schema definitions. It no longer uses object params EXCEPT for `required` key:
param :unit, type: :string, enum: %w[f c], description: "Unit (e.g., celsius, fahrenheit)", required: false

# It now supports object params
param :location, type: :object, desc: "Country and city.", required: false, properties: {
  country: { type: :string, description: "Full name of the country." },
  city: { type: :string, description: "Full name of the city." }
}

# And also supports array params:
param :location, type: :array, desc: "Latitude and longitude", items: { type: :number, description: "Float value" }

How I tested

I tested all the supported LLM APIs directly.

Shout-outs

• @tpaulshippy for providing test code for this and starting the initial PR Support nested parameters for tools #112
• Everyone who contributed to the thread at Structured output response support #11

[tpaulshippy]

Well tested and documented. Compatible with my fork - appreciate it!

[kieranklaassen]

Nice one! I like it. The only issue is I really need support for Gemini myself, so I cannot use this one yet in production. I'm happy to close mine or merge any elements. I'll let @crmne know what he wants. My goal is to use RubyLLM for all production tasks, and I need structured outputs for Claude, Gemini, and OpenAI. Using the prompt method works very well.

[jayelkaake]

@kieranklaassen sounds great.

Yeah let's wait to hear from @crmne first then I can merge your code for Gemini support into my PR since that one is hardest to support if he decides to go that direction.

[tpaulshippy]

Are you guys commenting on the wrong PR or is it just me?

[jayelkaake]

@tpaulshippy not just you - comments went to wrong PR.

Those 2 comments should be on @kieranklaassen 's PR for structured output