Ruby AI eXtensions

Raix (pronounced "ray" because the x is silent) is a library that gives you everything you need to add discrete large-language model (LLM) AI components to your Ruby applications. Raix consists of proven code that has been extracted from Olympia, the world's leading virtual AI team platform, and probably one of the biggest and most successful AI chat projects written completely in Ruby.

Understanding the how to use discrete AI components in otherwise normal code is key to productively leveraging Raix, and the subject of a book written by Raix's author Obie Fernandez, titled Patterns of Application Development Using AI. You can easily support the ongoing development of this project by buying the book at Leanpub.

At the moment, Raix natively supports use of either OpenAI or OpenRouter as its underlying AI provider. Eventually you will be able to specify your AI provider via an adapter, kind of like ActiveRecord maps to databases. Note that you can also use Raix to add AI capabilities to non-Rails applications as long as you include ActiveSupport as a dependency. Extracting the base code to its own standalone library without Rails dependencies is on the roadmap, but not a high priority.

Raix consists of three modules that can be mixed in to Ruby classes to give them AI powers. The first (and mandatory) module is ChatCompletion, which provides transcript and chat_completion methods.

class MeaningOfLife
  include Raix::ChatCompletion
end

>> ai = MeaningOfLife.new
>> ai.transcript << { user: "What is the meaning of life?" }
>> ai.chat_completion

=> "The question of the meaning of life is one of the most profound and enduring inquiries in philosophy, religion, and science.
    Different perspectives offer various answers..."

The transcript accepts both abbreviated and standard OpenAI message hash formats. The abbreviated format, suitable for system, assistant, and user messages is simply a mapping of role => content, as show in the example above.

transcript << { user: "What is the meaning of life?" }

As mentioned, Raix also understands standard OpenAI messages hashes. The previous example could be written as:

transcript << { role: "user", content: "What is the meaning of life?" }

One of the advantages of OpenRouter and the reason that it is used by default by this library is that it handles mapping message formats from the OpenAI standard to whatever other model you're wanting to use (Anthropic, Cohere, etc.)

Raix supports Predicted Outputs with the prediction parameter for OpenAI.

>> ai.chat_completion(openai: "gpt-4o", params: { prediction: })

Raix supports Anthropic-style prompt caching when using Anthropic's Claud family of models. You can specify a cache_at parameter when doing a chat completion. If the character count for the content of a particular message is longer than the cache_at parameter, it will be sent to Anthropic as a multipart message with a cache control "breakpoint" set to "ephemeral".

Note that there is a limit of four breakpoints, and the cache will expire within five minutes. Therefore, it is recommended to reserve the cache breakpoints for large bodies of text, such as character cards, CSV data, RAG data, book chapters, etc. Raix does not enforce a limit on the number of breakpoints, which means that you might get an error if you try to cache too many messages.

>> my_class.chat_completion(params: { cache_at: 1000 })
=> {
  "messages": [
    {
      "role": "system",
      "content": [
        {
          "type": "text",
          "text": "HUGE TEXT BODY LONGER THAN 1000 CHARACTERS",
          "cache_control": {
            "type": "ephemeral"
          }
        }
      ]
    },

The second (optional) module that you can add to your Ruby classes after ChatCompletion is FunctionDispatch. It lets you declare and implement functions to be called at the AI's discretion as part of a chat completion "loop" in a declarative, Rails-like "DSL" fashion.

Most end-user facing AI components that include functions should be invoked using chat_completion(loop: true), so that the results of each function call are added to the transcript and chat completion is triggered again. The looping will continue until the AI generates a plain text response.

class WhatIsTheWeather
  include Raix::ChatCompletion
  include Raix::FunctionDispatch

  function :check_weather, "Check the weather for a location", location: { type: "string" } do |arguments|
    "The weather in #{arguments[:location]} is hot and sunny"
  end
end

RSpec.describe WhatIsTheWeather do
  subject { described_class.new }

  it "can call a function and loop to provide text response" do
    subject.transcript << { user: "What is the weather in Zipolite, Oaxaca?" }
    response = subject.chat_completion(openai: "gpt-4o", loop: true)
    expect(response).to include("hot and sunny")
  end
end

Some AI models (like GPT-4) can make multiple tool calls in a single response. When this happens, Raix will automatically handle all the function calls sequentially and return an array of their results. Here's an example:

class MultipleToolExample
  include Raix::ChatCompletion
  include Raix::FunctionDispatch

  function :first_tool do |arguments|
    "Result from first tool"
  end

  function :second_tool do |arguments|
    "Result from second tool"
  end
end

example = MultipleToolExample.new
example.transcript << { user: "Please use both tools" }
results = example.chat_completion(openai: "gpt-4o")
# => ["Result from first tool", "Result from second tool"]

To loop AI components that don't interact with end users, at least one function block should invoke stop_looping! whenever you're ready to stop processing.

class OrderProcessor
  include Raix::ChatCompletion
  include Raix::FunctionDispatch

  SYSTEM_DIRECTIVE = "You are an order processor, tasked with order validation, inventory check,
                      payment processing, and shipping."

  attr_accessor :order

  def initialize(order)
    self.order = order
    transcript << { system: SYSTEM_DIRECTIVE }
    transcript << { user: order.to_json }
  end

  def perform
    # will continue looping until `stop_looping!` is called
    chat_completion(loop: true)
  end


  # implementation of functions that can be called by the AI
  # entirely at its discretion, depending on the needs of the order.
  # The return value of each `perform` method will be added to the
  # transcript of the conversation as a function result.

  function :validate_order do
    OrderValidationWorker.perform(@order)
  end

  function :check_inventory do
    InventoryCheckWorker.perform(@order)
  end

  function :process_payment do
    PaymentProcessingWorker.perform(@order)
  end

  function :schedule_shipping do
    ShippingSchedulerWorker.perform(@order)
  end

  function :send_confirmation do
    OrderConfirmationWorker.perform(@order)
  end

  function :finished_processing do
    order.update!(transcript:, processed_at: Time.current)
    stop_looping!
  end
end

The third (also optional) module that you can add mix in along with ChatCompletion is PromptDeclarations. It provides the ability to declare a "Prompt Chain" (series of prompts to be called in a sequence), and also features a declarative, Rails-like "DSL" of its own. Prompts can be defined inline or delegate to callable prompt objects, which themselves implement ChatCompletion.

The following example is a rough excerpt of the main "Conversation Loop" in Olympia, which pre-processes user messages to check for the presence of URLs and scan memory before submitting as a prompt to GPT-4. Note that prompt declarations are executed in the order that they are declared. The FetchUrlCheck callable prompt class is included for instructional purposes. Note that it is passed the an instance of the object that is calling it in its initializer as its context. The passing of context means that you can assemble composite prompt structures of arbitrary depth.

class PromptSubscriber
  include Raix::ChatCompletion
  include Raix::PromptDeclarations

  attr_accessor :conversation, :bot_message, :user_message

  # many other declarations ommitted...

  prompt call: FetchUrlCheck

  prompt call: MemoryScan

  prompt text: -> { user_message.content }, stream: -> { ReplyStream.new(self) }, until: -> { bot_message.complete? }

  def initialize(conversation)
    self.conversation = conversation
  end

  def message_created(user_message)
    self.user_message = user_message
    self.bot_message = conversation.bot_message!(responding_to: user_message)

    chat_completion(loop: true, openai: "gpt-4o")
  end

  ...

end

class FetchUrlCheck
  include ChatCompletion
  include FunctionDispatch

  REGEX = %r{\b(?:http(s)?://)?(?:www\.)?[a-zA-Z0-9-]+(\.[a-zA-Z]{2,})+(/[^\s]*)?\b}

  attr_accessor :context, :conversation

  delegate :user_message, to: :context
  delegate :content, to: :user_message

  def initialize(context)
    self.context = context
    self.conversation = context.conversation
    self.model = "anthropic/claude-3-haiku"
  end

  def call
    return unless content&.match?(REGEX)

    transcript << { system: "Call the `fetch` function if the user mentions a website, otherwise say nil" }
    transcript << { user: content }

    chat_completion # TODO: consider looping to fetch more than one URL per user message
  end

  function :fetch, "Gets the plain text contents of a web page", url: { type: "string" } do |arguments|
    Tools::FetchUrl.fetch(arguments[:url]).tap do |result|
      parent = conversation.function_call!("fetch_url", arguments, parent: user_message)
      conversation.function_result!("fetch_url", result, parent:)
    end
  end

Notably, Olympia does not use the FunctionDispatch module in its primary conversation loop because it does not have a fixed set of tools that are included in every single prompt. Functions are made available dynamically based on a number of factors including the user's plan tier and capabilities of the assistant with whom the user is conversing.

Streaming of the AI's response to the end user is handled by the ReplyStream class, passed to the final prompt declaration as its stream parameter. Patterns of Application Development Using AI devotes a whole chapter to describing how to write your own ReplyStream class.

The Raix::Predicate module provides a simple way to handle yes/no/maybe questions using AI chat completion. It allows you to define blocks that handle different types of responses with their explanations. It is one of the concrete patterns described in the "Discrete Components" chapter of Patterns of Application Development Using AI.

Include the Raix::Predicate module in your class and define handlers using block syntax:

class Question
  include Raix::Predicate

  yes? do |explanation|
    puts "Affirmative: #{explanation}"
  end

  no? do |explanation|
    puts "Negative: #{explanation}"
  end

  maybe? do |explanation|
    puts "Uncertain: #{explanation}"
  end
end

question = Question.new
question.ask("Is Ruby a programming language?")
# => Affirmative: Yes, Ruby is a dynamic, object-oriented programming language...

• Define handlers for yes, no, and/or maybe responses using the declarative class level block syntax.
• At least one handler (yes, no, or maybe) must be defined.
• Handlers receive the full AI response including explanation as an argument.
• Responses always start with "Yes, ", "No, ", or "Maybe, " followed by an explanation.
• Make sure to ask a question that can be answered with yes, no, or maybe (otherwise the results are indeterminate).

You can define only the handlers you need:

class SimpleQuestion
  include Raix::Predicate

  # Only handle positive responses
  yes? do |explanation|
    puts "✅ #{explanation}"
  end
end

question = SimpleQuestion.new
question.ask("Is 2 + 2 = 4?")
# => ✅ Yes, 2 + 2 equals 4, this is a fundamental mathematical fact.

The module will raise a RuntimeError if you attempt to ask a question without defining any response handlers:

class InvalidQuestion
  include Raix::Predicate
end

question = InvalidQuestion.new
question.ask("Any question")
# => RuntimeError: Please define a yes and/or no block

The ResponseFormat class provides a way to declare a JSON schema for the response format of an AI chat completion. It's particularly useful when you need structured responses from AI models, ensuring the output conforms to your application's requirements.

• Converts Ruby hashes and arrays into JSON schema format
• Supports nested structures and arrays
• Enforces strict validation with additionalProperties: false
• Automatically marks all top-level properties as required
• Handles both simple type definitions and complex nested schemas

# Simple schema with basic types
format = Raix::ResponseFormat.new("PersonInfo", {
  name: { type: "string" },
  age: { type: "integer" }
})

# Use in chat completion
my_ai.chat_completion(response_format: format)

# Nested structure with arrays
format = Raix::ResponseFormat.new("CompanyInfo", {
  company: {
    name: { type: "string" },
    employees: [
      {
        name: { type: "string" },
        role: { type: "string" },
        skills: ["string"]
      }
    ],
    locations: ["string"]
  }
})

The ResponseFormat class generates a schema that follows this structure:

{
  "type": "json_schema",
  "json_schema": {
    "name": "SchemaName",
    "schema": {
      "type": "object",
      "properties": {
        "property1": { "type": "string" },
        "property2": { "type": "integer" }
      },
      "required": ["property1", "property2"],
      "additionalProperties": false
    },
    "strict": true
  }
}

When used with chat completion, the AI model will format its response according to your schema:

class StructuredResponse
  include Raix::ChatCompletion

  def analyze_person(name)
    format = Raix::ResponseFormat.new("PersonAnalysis", {
      full_name: { type: "string" },
      age_estimate: { type: "integer" },
      personality_traits: ["string"]
    })

    transcript << { user: "Analyze the person named #{name}" }
    chat_completion(response_format: format)
  end
end

response = StructuredResponse.new.analyze_person("Alice")
# Returns a hash matching the defined schema

Install the gem and add to the application's Gemfile by executing:

$ bundle add raix

If bundler is not being used to manage dependencies, install the gem by executing:

$ gem install raix

If you are using the default OpenRouter API, Raix expects Raix.configuration.openrouter_client to initialized with the OpenRouter API client instance.

You can add an initializer to your application's config/initializers directory:

  # config/initializers/raix.rb
  Raix.configure do |config|
    config.openrouter_client = OpenRouter::Client.new
  end

You will also need to configure the OpenRouter API access token as per the instructions here: https://github.com/OlympiaAI/open_router?tab=readme-ov-file#quickstart

After checking out the repo, run bin/setup to install dependencies. Then, run rake spec to run the tests. You can also run bin/console for an interactive prompt that will allow you to experiment.

Specs require OR_ACCESS_TOKEN and OAI_ACCESS_TOKEN environment variables, for access to OpenRouter and OpenAI, respectively. You can add those keys to a local unversionsed .env file and they will be picked up by the dotenv gem.

To install this gem onto your local machine, run bundle exec rake install. To release a new version, update the version number in version.rb, and then run bundle exec rake release, which will create a git tag for the version, push git commits and the created tag, and push the .gem file to rubygems.org.

Bug reports and pull requests are welcome on GitHub at https://github.com/[OlympiaAI]/raix. This project is intended to be a safe, welcoming space for collaboration, and contributors are expected to adhere to the code of conduct.

The gem is available as open source under the terms of the MIT License.

Everyone interacting in the Raix project's codebases, issue trackers, chat rooms and mailing lists is expected to follow the code of conduct.