ruby_llm 0.1.0.pre30 → 0.1.0.pre33

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: 33b602ae544211e79207cf51bbfad5f1ed9c78cd1d9710d4f57916c95377bb60
-  data.tar.gz: 3fac0d6aed252218d7f7bd8e460f7b281c0306ca066e0194193e11a44ac8400d
+  metadata.gz: e36a864dcd3852d03d467684aafcc811273f0f44dbd1925484440808c6c8895b
+  data.tar.gz: 5ed1b436a1b91f7fac06ab76d43fd0c7dfee925cc1af280a10bb40e448dabf0b
 SHA512:
-  metadata.gz: fa21a4d89b3704f384c257a7bcca71e42b0442b2f660ddd2c93d108f77bcf547b5d1c05101d85269abc5aab8fbe37e3bec3da8a4fd09d77f8184ed923fea0986
-  data.tar.gz: 5295d0c1ec5660e4ce9e7388836bbc37d82eae0ccce77a4f601f59ed91933cb2dc31b84f79f5aab3271d5b8b318aa190ca0224926a28c69a0e409ed9035db4c8
+  metadata.gz: 6dd5a24c4a374a5fd291c7f47363dd282abcc850361464e67859f82ad88f51f49f0d4fda89299cf3871df838896acb111170a3615963c215a2aedb6b4154d0bd
+  data.tar.gz: 07f0931ca91abe8c42b7ca52910f4ac31f34ec4e049a321a7a4d59808cd190813aab1534312bb279d609a94f2ee6fbb2164faa6330a82a80ff520697741a7088

data/.github/workflows/{gem-push.yml → cicd.yml}

@@ -1,18 +1,46 @@
-name: Ruby Gem
+name: CI
 
 on:
   push:
     branches: [ "main" ]
   pull_request:
     branches: [ "main" ]
+  workflow_call:
 
 jobs:
   test:
-    uses: ./.github/workflows/test.yml
+    runs-on: ubuntu-latest
+    strategy:
+      matrix:
+        ruby-version: ['3.1']
+
+    steps:
+    - uses: actions/checkout@v4
+
+    - name: Set up Ruby
+      uses: ruby/setup-ruby@v1
+      with:
+        ruby-version: ${{ matrix.ruby-version }}
+        bundler-cache: true
+
+    - name: Install dependencies
+      run: bundle install
+
+    - name: Check code format
+      run: bundle exec rubocop
+
+    - name: Run tests
+      env:
+        OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }}
+        ANTHROPIC_API_KEY: ${{ secrets.ANTHROPIC_API_KEY }}
+        GEMINI_API_KEY: ${{ secrets.GEMINI_API_KEY }}
+        DEEPSEEK_API_KEY: ${{ secrets.DEEPSEEK_API_KEY }}
+      run: bundle exec rspec
 
-  build:
-    needs: test # This ensures tests must pass before building/publishing
+  publish:
     name: Build + Publish
+    # needs: test
+    if: github.event_name == 'push' && github.ref == 'refs/heads/main'
     runs-on: ubuntu-latest
     permissions:
       contents: read

data/.rspec_status

@@ -0,0 +1,38 @@
+example_id                                         | status | run_time        |
+-------------------------------------------------- | ------ | --------------- |
+./spec/integration/chat_spec.rb[1:1:1:1]           | passed | 0.5826 seconds  |
+./spec/integration/chat_spec.rb[1:1:1:2]           | passed | 3.43 seconds    |
+./spec/integration/chat_spec.rb[1:1:2:1]           | passed | 0.49383 seconds |
+./spec/integration/chat_spec.rb[1:1:2:2]           | passed | 1.36 seconds    |
+./spec/integration/chat_spec.rb[1:1:3:1]           | passed | 0.74049 seconds |
+./spec/integration/chat_spec.rb[1:1:3:2]           | passed | 4.11 seconds    |
+./spec/integration/content_spec.rb[1:1:1]          | passed | 3.87 seconds    |
+./spec/integration/content_spec.rb[1:1:2]          | passed | 1.23 seconds    |
+./spec/integration/content_spec.rb[1:1:3]          | passed | 2.07 seconds    |
+./spec/integration/content_spec.rb[1:2:1]          | passed | 2.05 seconds    |
+./spec/integration/content_spec.rb[1:2:2]          | passed | 2.88 seconds    |
+./spec/integration/embeddings_spec.rb[1:1:1:1]     | passed | 0.30185 seconds |
+./spec/integration/embeddings_spec.rb[1:1:1:2]     | passed | 0.30812 seconds |
+./spec/integration/embeddings_spec.rb[1:1:2:1]     | passed | 13.05 seconds   |
+./spec/integration/embeddings_spec.rb[1:1:2:2]     | passed | 0.78135 seconds |
+./spec/integration/error_handling_spec.rb[1:1]     | passed | 0.21297 seconds |
+./spec/integration/image_generation_spec.rb[1:1:1] | passed | 12.44 seconds   |
+./spec/integration/image_generation_spec.rb[1:1:2] | passed | 17.66 seconds   |
+./spec/integration/image_generation_spec.rb[1:1:3] | passed | 0.00324 seconds |
+./spec/integration/image_generation_spec.rb[1:1:4] | failed | 0.15682 seconds |
+./spec/integration/image_generation_spec.rb[1:1:5] | passed | 18.69 seconds   |
+./spec/integration/image_generation_spec.rb[1:1:6] | passed | 0.00032 seconds |
+./spec/integration/rails_spec.rb[1:1]              | passed | 4.05 seconds    |
+./spec/integration/rails_spec.rb[1:2]              | passed | 1.82 seconds    |
+./spec/integration/streaming_spec.rb[1:1:1:1]      | passed | 0.58445 seconds |
+./spec/integration/streaming_spec.rb[1:1:1:2]      | passed | 6.04 seconds    |
+./spec/integration/streaming_spec.rb[1:1:2:1]      | passed | 0.47171 seconds |
+./spec/integration/streaming_spec.rb[1:1:2:2]      | passed | 2.39 seconds    |
+./spec/integration/streaming_spec.rb[1:1:3:1]      | passed | 0.72016 seconds |
+./spec/integration/streaming_spec.rb[1:1:3:2]      | passed | 3.59 seconds    |
+./spec/integration/tools_spec.rb[1:1:1:1]          | passed | 3.1 seconds     |
+./spec/integration/tools_spec.rb[1:1:1:2]          | passed | 7.04 seconds    |
+./spec/integration/tools_spec.rb[1:1:2:1]          | passed | 1.42 seconds    |
+./spec/integration/tools_spec.rb[1:1:2:2]          | passed | 2.24 seconds    |
+./spec/integration/tools_spec.rb[1:1:3:1]          | passed | 2.16 seconds    |
+./spec/integration/tools_spec.rb[1:1:3:2]          | passed | 5.26 seconds    |

data/README.md

@@ -1,20 +1,33 @@
 # RubyLLM
 
-A delightful Ruby way to work with AI language models. Provides a unified interface to OpenAI, Anthropic, Google, and DeepSeek models with automatic token counting, proper streaming support, and a focus on developer happiness. No wrapping your head around multiple APIs - just clean Ruby code that works.
+A delightful Ruby way to work with AI. Chat in text, analyze and generate images, understand audio, and use tools through a unified interface to OpenAI, Anthropic, Google, and DeepSeek. Built for developer happiness with automatic token counting, proper streaming, and Rails integration. No wrapping your head around multiple APIs - just clean Ruby code that works.
 
 <p align="center">
   <img src="https://upload.wikimedia.org/wikipedia/commons/4/4d/OpenAI_Logo.svg" alt="OpenAI" height="40" width="120">
   &nbsp;&nbsp;&nbsp;&nbsp;
   <img src="https://upload.wikimedia.org/wikipedia/commons/7/78/Anthropic_logo.svg" alt="Anthropic" height="40" width="120">
   &nbsp;&nbsp;&nbsp;&nbsp;
-<img src="https://upload.wikimedia.org/wikipedia/commons/8/8a/Google_Gemini_logo.svg" alt="Google" height="40" width="120">
+  <img src="https://upload.wikimedia.org/wikipedia/commons/8/8a/Google_Gemini_logo.svg" alt="Google" height="40" width="120">
   &nbsp;&nbsp;&nbsp;&nbsp;
-  <img src="https://upload.wikimedia.org/wikipedia/commons/e/ec/DeepSeek_logo.svg" alt="DeepSeek" height="40" width="120"]>
+  <img src="https://upload.wikimedia.org/wikipedia/commons/e/ec/DeepSeek_logo.svg" alt="DeepSeek" height="40" width="120">
 </p>
 
 [![Gem Version](https://badge.fury.io/rb/ruby_llm.svg)](https://badge.fury.io/rb/ruby_llm)
 [![Ruby Style Guide](https://img.shields.io/badge/code_style-standard-brightgreen.svg)](https://github.com/testdouble/standard)
 
+## Features
+
+- 💬 **Beautiful Chat Interface** - Converse with AI models as easily as `RubyLLM.chat.ask "teach me Ruby"`
+- 🎵 **Audio Analysis** - Get audio transcription and understanding with `chat.ask "what's said here?", with: { audio: "clip.wav" }`
+- 👁️ **Vision Understanding** - Let AIs analyze images with a simple `chat.ask "what's in this?", with: { image: "photo.jpg" }`
+- 🌊 **Streaming** - Real-time responses with proper Ruby streaming with `chat.ask "hello" do |chunk| puts chunk.content end`
+- 🚂 **Rails Integration** - Persist chats and messages with ActiveRecord with `acts_as_{chat|message|tool_call}`
+- 🛠️ **Tool Support** - Give AIs access to your Ruby code with `chat.with_tool(Calculator).ask "what's 2+2?"`
+- 🎨 **Paint with AI** - Create images as easily as `RubyLLM.paint "a sunset over mountains"`
+- 📊 **Embeddings** - Generate vector embeddings for your text with `RubyLLM.embed "hello"`
+- 🔄 **Multi-Provider Support** - Works with OpenAI, Anthropic, Google, and DeepSeek
+- 🎯 **Token Tracking** - Automatic usage tracking across providers
+
 ## Installation
 
 Add it to your Gemfile:
@@ -87,11 +100,47 @@ chat.ask "Tell me a story about a Ruby programmer" do |chunk|
   print chunk.content
 end
 
+# Ask about images
+chat.ask "What do you see in this image?", with: { image: "ruby_logo.png" }
+
+# Get analysis of audio content
+chat.ask "What's being said in this recording?", with: { audio: "meeting.wav" }
+
+# Combine multiple pieces of content
+chat.ask "Compare these diagrams", with: { image: ["diagram1.png", "diagram2.png"] }
+
 # Check token usage
 last_message = chat.messages.last
 puts "Conversation used #{last_message.input_tokens} input tokens and #{last_message.output_tokens} output tokens"
 ```
 
+You can provide content as local files or URLs - RubyLLM handles the rest. Vision and audio capabilities are available with compatible models (Claude 3, GPT-4V, Gemini Pro Vision). The API stays clean and consistent whether you're working with text, images, or audio.
+
+## Image Generation
+
+Want to create AI-generated images? RubyLLM makes it super simple:
+
+```ruby
+# Paint a picture!
+image = RubyLLM.paint "a starry night over San Francisco in Van Gogh's style"
+image.url         # => "https://..."
+image.revised_prompt  # Shows how DALL-E interpreted your prompt
+
+# Choose size and model
+image = RubyLLM.paint(
+  "a cyberpunk cityscape at sunset",
+  model: "dall-e-3",
+  size: "1792x1024"
+)
+
+# Set your default model
+RubyLLM.configure do |config|
+  config.default_image_model = "dall-e-3"
+end
+```
+
+RubyLLM automatically handles all the complexities of the DALL-E API, token/credit management, and error handling, so you can focus on being creative.
+
 ## Text Embeddings
 
 Need vector embeddings for your text? RubyLLM makes it simple:

data/lib/ruby_llm/active_record/acts_as.rb

@@ -73,22 +73,22 @@ module RubyLLM
       end
 
       def to_llm
-        chat = RubyLLM.chat(model: model_id)
+        @chat ||= RubyLLM.chat(model: model_id)
 
         # Load existing messages into chat
         messages.each do |msg|
-          chat.add_message(msg.to_llm)
+          @chat.add_message(msg.to_llm)
         end
 
         # Set up message persistence
-        chat.on_new_message { persist_new_message }
-            .on_end_message { |msg| persist_message_completion(msg) }
+        @chat.on_new_message { persist_new_message }
+             .on_end_message { |msg| persist_message_completion(msg) }
       end
 
       def ask(message, &block)
         message = { role: :user, content: message }
         messages.create!(**message)
-        chat.complete(&block)
+        to_llm.complete(&block)
       end
 
       alias say ask

data/lib/ruby_llm/chat.rb

@@ -25,8 +25,8 @@ module RubyLLM
       }
     end
 
-    def ask(message, &block)
-      add_message role: :user, content: message
+    def ask(message = nil, with: {}, &block)
+      add_message role: :user, content: Content.new(message, with)
       complete(&block)
     end
 

data/lib/ruby_llm/configuration.rb

@@ -16,12 +16,16 @@ module RubyLLM
                   :deepseek_api_key,
                   :default_model,
                   :default_embedding_model,
-                  :request_timeout
+                  :default_image_model,
+                  :request_timeout,
+                  :max_retries
 
     def initialize
       @request_timeout = 120
+      @max_retries = 3
       @default_model = 'gpt-4o-mini'
       @default_embedding_model = 'text-embedding-3-small'
+      @default_image_model = 'dall-e-3'
     end
   end
 end

data/lib/ruby_llm/content.rb

@@ -0,0 +1,81 @@
+# frozen_string_literal: true
+
+module RubyLLM
+  # Represents the content sent to or received from an LLM.
+  # Stores data in a standard internal format, letting providers
+  # handle their own formatting needs.
+  class Content
+    def initialize(text = nil, attachments = {})
+      @parts = []
+      @parts << { type: 'text', text: text } unless text.nil? || text.empty?
+
+      Array(attachments[:image]).each do |source|
+        @parts << attach_image(source)
+      end
+
+      Array(attachments[:audio]).each do |source|
+        @parts << attach_audio(source)
+      end
+    end
+
+    def to_a
+      return if @parts.empty?
+
+      @parts
+    end
+
+    def format
+      return @parts.first[:text] if @parts.size == 1 && @parts.first[:type] == 'text'
+
+      to_a
+    end
+
+    private
+
+    def attach_image(source) # rubocop:disable Metrics/MethodLength
+      source = File.expand_path(source) unless source.start_with?('http')
+
+      return { type: 'image', source: { url: source } } if source.start_with?('http')
+
+      data = Base64.strict_encode64(File.read(source))
+      mime_type = mime_type_for(source)
+
+      {
+        type: 'image',
+        source: {
+          type: 'base64',
+          media_type: mime_type,
+          data: data
+        }
+      }
+    end
+
+    def attach_audio(source)
+      source = File.expand_path(source) unless source.start_with?('http')
+      data = encode_file(source)
+      format = File.extname(source).delete('.') || 'wav'
+
+      {
+        type: 'input_audio',
+        input_audio: {
+          data: data,
+          format: format
+        }
+      }
+    end
+
+    def encode_file(source)
+      if source.start_with?('http')
+        response = Faraday.get(source)
+        Base64.strict_encode64(response.body)
+      else
+        Base64.strict_encode64(File.read(source))
+      end
+    end
+
+    def mime_type_for(path)
+      ext = File.extname(path).delete('.')
+      "image/#{ext}"
+    end
+  end
+end

data/lib/ruby_llm/embedding.rb

@@ -3,10 +3,16 @@
 module RubyLLM
   # Core embedding interface. Provides a clean way to generate embeddings
   # from text using various provider models.
-  module Embedding
-    module_function
+  class Embedding
+    attr_reader :vectors, :model, :input_tokens
 
-    def embed(text, model: nil)
+    def initialize(vectors:, model:, input_tokens: 0)
+      @vectors = vectors
+      @model = model
+      @input_tokens = input_tokens
+    end
+
+    def self.embed(text, model: nil)
       model_id = model || RubyLLM.config.default_embedding_model
       Models.find(model_id)
 

data/lib/ruby_llm/image.rb

@@ -0,0 +1,24 @@
+# frozen_string_literal: true
+
+module RubyLLM
+  # Represents a generated image from an AI model.
+  # Provides an interface to image generation capabilities
+  # from providers like DALL-E.
+  class Image
+    attr_reader :url, :revised_prompt, :model_id
+
+    def initialize(url:, revised_prompt: nil, model_id: nil)
+      @url = url
+      @revised_prompt = revised_prompt
+      @model_id = model_id
+    end
+
+    def self.paint(prompt, model: nil, size: '1024x1024')
+      model_id = model || RubyLLM.config.default_image_model
+      Models.find(model_id) # Validate model exists
+
+      provider = Provider.for(model_id)
+      provider.paint(prompt, model: model_id, size: size)
+    end
+  end
+end

data/lib/ruby_llm/message.rb

@@ -11,7 +11,7 @@ module RubyLLM
 
     def initialize(options = {})
       @role = options[:role].to_sym
-      @content = options[:content]
+      @content = normalize_content(options[:content])
       @tool_calls = options[:tool_calls]
       @input_tokens = options[:input_tokens]
       @output_tokens = options[:output_tokens]
@@ -47,6 +47,14 @@ module RubyLLM
 
     private
 
+    def normalize_content(content)
+      case content
+      when Content then content.format
+      when String then Content.new(content).format
+      else content
+      end
+    end
+
     def ensure_valid_role
       raise InvalidRoleError, "Expected role to be one of: #{ROLES.join(', ')}" unless ROLES.include?(role)
     end

data/lib/ruby_llm/models.json

@@ -727,7 +727,7 @@
     "family": "gemini20_flash",
     "supports_vision": true,
     "supports_functions": true,
-    "supports_json_mode": false,
+    "supports_json_mode": true,
     "input_price_per_million": 0.1,
     "output_price_per_million": 0.4,
     "metadata": {
@@ -746,7 +746,7 @@
     "family": "gemini20_flash",
     "supports_vision": true,
     "supports_functions": true,
-    "supports_json_mode": false,
+    "supports_json_mode": true,
     "input_price_per_million": 0.1,
     "output_price_per_million": 0.4,
     "metadata": {
@@ -765,7 +765,7 @@
     "family": "gemini20_flash",
     "supports_vision": true,
     "supports_functions": true,
-    "supports_json_mode": false,
+    "supports_json_mode": true,
     "input_price_per_million": 0.1,
     "output_price_per_million": 0.4,
     "metadata": {
@@ -784,7 +784,7 @@
     "family": "gemini20_flash",
     "supports_vision": true,
     "supports_functions": true,
-    "supports_json_mode": false,
+    "supports_json_mode": true,
     "input_price_per_million": 0.1,
     "output_price_per_million": 0.4,
     "metadata": {
@@ -803,7 +803,7 @@
     "family": "gemini20_flash",
     "supports_vision": true,
     "supports_functions": true,
-    "supports_json_mode": false,
+    "supports_json_mode": true,
     "input_price_per_million": 0.1,
     "output_price_per_million": 0.4,
     "metadata": {
@@ -822,7 +822,7 @@
     "family": "gemini20_flash",
     "supports_vision": true,
     "supports_functions": true,
-    "supports_json_mode": false,
+    "supports_json_mode": true,
     "input_price_per_million": 0.1,
     "output_price_per_million": 0.4,
     "metadata": {
@@ -841,7 +841,7 @@
     "family": "gemini20_flash_lite",
     "supports_vision": true,
     "supports_functions": false,
-    "supports_json_mode": false,
+    "supports_json_mode": true,
     "input_price_per_million": 0.075,
     "output_price_per_million": 0.3,
     "metadata": {
@@ -860,7 +860,7 @@
     "family": "gemini20_flash_lite",
     "supports_vision": true,
     "supports_functions": false,
-    "supports_json_mode": false,
+    "supports_json_mode": true,
     "input_price_per_million": 0.075,
     "output_price_per_million": 0.3,
     "metadata": {
@@ -879,7 +879,7 @@
     "family": "gemini20_flash",
     "supports_vision": true,
     "supports_functions": true,
-    "supports_json_mode": false,
+    "supports_json_mode": true,
     "input_price_per_million": 0.1,
     "output_price_per_million": 0.4,
     "metadata": {
@@ -898,7 +898,7 @@
     "family": "gemini20_flash",
     "supports_vision": true,
     "supports_functions": true,
-    "supports_json_mode": false,
+    "supports_json_mode": true,
     "input_price_per_million": 0.1,
     "output_price_per_million": 0.4,
     "metadata": {
@@ -917,7 +917,7 @@
     "family": "gemini20_flash",
     "supports_vision": true,
     "supports_functions": true,
-    "supports_json_mode": false,
+    "supports_json_mode": true,
     "input_price_per_million": 0.1,
     "output_price_per_million": 0.4,
     "metadata": {
@@ -936,7 +936,7 @@
     "family": "gemini20_flash",
     "supports_vision": true,
     "supports_functions": true,
-    "supports_json_mode": false,
+    "supports_json_mode": true,
     "input_price_per_million": 0.1,
     "output_price_per_million": 0.4,
     "metadata": {
@@ -955,7 +955,7 @@
     "family": "other",
     "supports_vision": true,
     "supports_functions": true,
-    "supports_json_mode": false,
+    "supports_json_mode": true,
     "input_price_per_million": 0.075,
     "output_price_per_million": 0.3,
     "metadata": {
@@ -974,7 +974,7 @@
     "family": "other",
     "supports_vision": true,
     "supports_functions": true,
-    "supports_json_mode": false,
+    "supports_json_mode": true,
     "input_price_per_million": 0.075,
     "output_price_per_million": 0.3,
     "metadata": {

data/lib/ruby_llm/provider.rb

@@ -5,15 +5,11 @@ module RubyLLM
   # Handles the complexities of API communication, streaming responses,
   # and error handling so individual providers can focus on their unique features.
   module Provider
-    def self.included(base)
-      base.include(InstanceMethods)
-    end
-
     # Common functionality for all LLM providers. Implements the core provider
     # interface so specific providers only need to implement a few key methods.
-    module InstanceMethods
+    module Methods # rubocop:disable Metrics/ModuleLength
       def complete(messages, tools:, temperature:, model:, &block)
-        payload = build_payload messages, tools: tools, temperature: temperature, model: model, stream: block_given?
+        payload = render_payload messages, tools: tools, temperature: temperature, model: model, stream: block_given?
 
         if block_given?
           stream_response payload, &block
@@ -31,11 +27,18 @@ module RubyLLM
       end
 
       def embed(text, model:)
-        payload = build_embedding_payload text, model: model
+        payload = render_embedding_payload text, model: model
         response = post embedding_url, payload
         parse_embedding_response response
       end
 
+      def paint(prompt, model:, size:)
+        payload = render_image_payload(prompt, model:, size:)
+
+        response = post(images_url, payload)
+        parse_image_response(response)
+      end
+
       private
 
       def sync_response(payload)
@@ -63,14 +66,43 @@ module RubyLLM
         end
       end
 
-      def connection
-        @connection ||= Faraday.new(api_base) do |f|
+      def connection # rubocop:disable Metrics/MethodLength,Metrics/AbcSize
+        @connection ||= Faraday.new(api_base) do |f| # rubocop:disable Metrics/BlockLength
           f.options.timeout = RubyLLM.config.request_timeout
+
+          f.response :logger,
+                     RubyLLM.logger,
+                     bodies: true,
+                     response: true,
+                     errors: true,
+                     headers: false,
+                     log_level: :debug do |logger|
+            logger.filter(%r{"[A-Za-z0-9+/=]{100,}"}, 'data":"[BASE64 DATA]"')
+            logger.filter(/[-\d.e,\s]{100,}/, '[EMBEDDINGS ARRAY]')
+          end
+
+          f.request :retry, {
+            max: RubyLLM.config.max_retries,
+            interval: 0.05,
+            interval_randomness: 0.5,
+            backoff_factor: 2,
+            exceptions: [
+              Errno::ETIMEDOUT,
+              Timeout::Error,
+              Faraday::TimeoutError,
+              Faraday::ConnectionFailed,
+              Faraday::RetriableResponse,
+              RubyLLM::RateLimitError,
+              RubyLLM::ServerError,
+              RubyLLM::ServiceUnavailableError
+            ],
+            retry_statuses: [429, 500, 502, 503, 504]
+          }
+
           f.request :json
           f.response :json
           f.adapter Faraday.default_adapter
           f.use :llm_errors, provider: self
-          f.response :logger, RubyLLM.logger, { headers: false, bodies: true, errors: true, log_level: :debug }
         end
       end
 
@@ -111,9 +143,16 @@ module RubyLLM
       maybe_json
     end
 
+    def parse_error(response)
+      return if response.body.empty?
+
+      body = try_parse_json(response.body)
+      body.is_a?(Hash) ? body.dig('error', 'message') : body
+    end
+
     def capabilities
       provider_name = self.class.name.split('::').last
-      RubyLLM.const_get "ModelCapabilities::#{provider_name}"
+      provider_name::Capabilities
     end
 
     def slug
@@ -121,15 +160,17 @@ module RubyLLM
     end
 
     class << self
-      def register(name, provider_class)
-        providers[name.to_sym] = provider_class
+      def extended(base)
+        base.extend(Methods)
+      end
+
+      def register(name, provider_module)
+        providers[name.to_sym] = provider_module
       end
 
       def for(model)
         model_info = Models.find(model)
-        provider_class = providers[model_info.provider.to_sym]
-
-        provider_class.new
+        providers[model_info.provider.to_sym]
       end
 
       def providers