rdh/active_intelligence README

A Rails engine that provides Rails-y AI integration

This engine is in (very) early development and breaking changes are expected. Use at your own risk or contribute at your pleasure.

Add this line to your application's Gemfile:

gem 'active_intelligence', git: 'git@github.com:rdh/active_intelligence.git', branch: 'main'

And then execute:

$ bundle

To install migrations:

rails active_intelligence:install:migrations

Configure your LLM in config/ai/llm.yml, something like:

openai: &openai
  adapter: openai
  access_token: <%= ENV.fetch('OPENAI_ACCESS_TOKEN') %>
  organization_id: <%= ENV.fetch('OPENAI_ORGANIZATION_ID')  %>
  request_timeout: 120

development:
  <<: *openai
  model: gpt-4-turbo
  temperature: 0.0

adapter = ActiveIntelligence::LLM.adapter
puts adapter.generate("Tell me a joke")

• Prompts live in app/prompts. They are ERB files that use a model as binding.
• The default prompt per-model is named after the model, e.g. app/prompts/users.erb
• Named prompts per-model live in a subdirectory named adter the model, e.g. app/prompts/users/invite.erb

Add include ActiveIntelligence::Promptable to your model, which adds the #to_prompt and #from_llm methods.

default_response = user.from_llm 
invite_response = user.from_llm(:invite)

erDiagram
    Chat {
        integer id
    }
    ChatMessage {
        integer id
        integer chat_id
        string role
        string content
    }
    Chat ||--o{ ChatMessage : has_many

# app/prompts/active_intelligence/chat.erb

* Your name is Poe.  You are a fan of Edgar Allan Poe.
* You are the AI proprietor of the Raven Hotel.
* You exhibit the utmost sincerity and hospitality.

include ActiveIntelligence

chat = Chat.create!
chat.messages.create!(role: 'user', content: "Hi!  Who are you?")
puts chat.reply.content

rake active_intelligence:chat
rake active_intelligence:chat[id] 

erDiagram
    Embedding {
        integer id
        integer embeddable_id
        integer embeddable_type
        vector embedding
    }
    Embeddable {
        integer id
    }
    Embeddable ||--o{ Embedding : has_many

Configure your LLM in config/ai/embeddings.yml, something like:

openai: &openai
  adapter: openai
  access_token: <%= ENV.fetch('OPENAI_ACCESS_TOKEN') %>
  organization_id: <%= ENV.fetch('OPENAI_ORGANIZATION_ID')  %>
  request_timeout: 30
  model: text-embedding-3-small

development:
  <<: *openai

class Greeting < ApplicationRecord
  include ActiveIntelligence::Embeddable
  
  def self.seed(text)
    create.add_embedding(text)
  end
end

Greeting.seed('Hello darkness, my old friend')
Greeting.seed('Aloha!')

greetings Greeting.semantic_search('Hello')  

This relies on pg_vector and the neighbor gem.

The included logic supports simple use cases.
For more complex cases, you may want to add an embedding vector directly to your model, rather than use the Embeddable concern.

Configure your LLM in config/ai/llm.yml, something like:

aws: &aws
  adapter: aws
  access_key_id: <%= ENV.fetch('AWS_ACCESS_KEY_ID') %>
  secret_access_key: <%= ENV.fetch('AWS_SECRET_ACCESS_KEY') %>
  region: <%= ENV.fetch('AWS_REGION') %>
  bucket: <%= ENV.fetch('AWS_TRANSCRIBE_BUCKET') %>
  folder: <%= ENV.fetch('AWS_TRANSCRIBE_FOLDER') %>
  language_code: en-US
  
openai: &openai
  adapter: openai
  access_token: <%= ENV.fetch('OPENAI_ACCESS_TOKEN') %>
  organization_id: <%= ENV.fetch('OPENAI_ORGANIZATION_ID')  %>
  request_timeout: 300
  model: whisper-1
  
development:
  <<: *openai

adapter = ActiveIntelligence::ASR.adapter
puts adapter.transcribe('spec/data/audio/ebn.wav')

Configure your LLM in config/ai/tts.yml, something like:

eleven_labs: &eleven_labs
  adapter: eleven_labs
  api_key: <%= ENV.fetch('ELEVEN_LABS_API_KEY') %>

charlie: &charlie
  <<: *eleven_labs
  voice_id: IKne3meq5aSn9XLyUdCD

development:
  <<: *charlie

adapter = ActiveIntelligence::TTS.adapter
adapter.generate_file('Hello darkness, my old friend', 'tmp/hello.mp3')

The engine currently has three significant modules: ASR, LLM, and TTS. Each module has a common Config and Adapter pattern.

The config is a constructor for the adapter.
By default, it uses the Rails.env as the key, but you can specify one:

adapter = ActiveIntelligence::ASR.adapter # uses Rails.env
adapter = ActiveIntelligence::ASR.adapter(:foobar) # uses the named configuration

Values in a configuration will "flow through" to services called by the adapter, so you can set defaults in the configuration. Options provided directly to calls will take precedence over the configuration.

Contribution directions go here.

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