AI Integration & Automation for Rails applications

It looks like Rails, because it is

• Plain Ruby objects

  AI features are service objects and jobs your team can read, test and change. No framework lock-in.

• Permission-aware retrieval

  Retrieval runs through your existing scopes, so a model never sees a record the current user could not open.

• Async by default

  Slow model calls run in Sidekiq or Solid Queue with retries, timeouts and provider fallback.

• Tested like everything else

  Unit tests with recorded responses, plus evaluation suites that score real outputs on every change.

class Support::AnswerQuestion
  def call(question, user:)
    # 1. Retrieve: only records this user may see
    sources = Article.visible_to(user)
      .nearest_neighbors(:embedding, embed(question),
                         distance: "cosine")
      .first(5)

    # 2. Generate: grounded in those sources
    answer = LLM.chat(task: :support_answer)
      .ask(prompt_for(question, sources))

    # 3. Record: every call is traced and scored
    AiTrace.record!(user:, question:, sources:, answer:)
    answer
  end
end

Simplified example. LLM is a thin gateway in your application that wraps the provider SDKs; nearest_neighbors comes from the neighbor gem, which adds pgvector support to Active Record.