Validate and retry LLM outputs with RubyLLM::Contract

[justi]

Hi everyone! I've built RubyLLM::Contract, a gem that wraps RubyLLM::Chat with input/output contracts, business-rule validation, retry with model escalation, pre-flight cost ceilings, and a regression-eval framework.

The problem it solves: LLMs can return JSON that looks correct - valid shape, right types - while being silently wrong in ways schema validation doesn't catch. Contract lets you declare business rules next to the prompt, escalate to a stronger model when the cheap one fails them, and gate CI on eval regressions.

class SummarizeArticle < RubyLLM::Contract::Step::Base
  prompt "Summarize this article for a UI card...\n\n{input}"

  output_schema do
    string :tldr
    array  :takeaways, of: :string, min_items: 3, max_items: 5
  end

  validate("TL;DR fits the card") { |o, _| o[:tldr].length <= 200 }
  retry_policy models: %w[gpt-4.1-nano gpt-4.1-mini gpt-4.1]
end

It composes with RubyLLM::Tribunal - Contract gates at runtime (before output reaches your code, with retries on failure), Tribunal grades at test time (in your spec suite). Ecosystem PR: #808. Feedback very welcome!