llm.rb 0.1.0 → 0.2.0

data/lib/llm/providers/ollama/response_parser.rb

@@ -2,6 +2,19 @@
 
 class LLM::Ollama
   module ResponseParser
+    ##
+    # @param [Hash] body
+    #  The response body from the LLM provider
+    # @return [Hash]
+    def parse_embedding(body)
+      {
+        model: body["model"],
+        embeddings: body["data"].map { _1["embedding"] },
+        prompt_tokens: body.dig("usage", "prompt_tokens"),
+        total_tokens: body.dig("usage", "total_tokens")
+      }
+    end
+
     ##
     # @param [Hash] body
     #  The response body from the LLM provider

data/lib/llm/providers/ollama.rb

@@ -11,7 +11,6 @@ module LLM
     include Format
 
     HOST = "localhost"
-    DEFAULT_PARAMS = {model: "llama3.2", stream: false}.freeze
 
     ##
     # @param secret (see LLM::Provider#initialize)
@@ -19,25 +18,50 @@ module LLM
       super(secret, host: HOST, port: 11434, ssl: false, **)
     end
 
+    ##
+    # @param input (see LLM::Provider#embed)
+    # @return (see LLM::Provider#embed)
+    def embed(input, **params)
+      params   = {model: "llama3.2"}.merge!(params)
+      req      = Net::HTTP::Post.new("/v1/embeddings", headers)
+      req.body = JSON.dump({input:}.merge!(params))
+      res      = request(@http, req)
+      Response::Embedding.new(res).extend(response_parser)
+    end
+
     ##
     # @see https://github.com/ollama/ollama/blob/main/docs/api.md#generate-a-chat-completion Ollama docs
     # @param prompt (see LLM::Provider#complete)
     # @param role (see LLM::Provider#complete)
     # @return (see LLM::Provider#complete)
     def complete(prompt, role = :user, **params)
-      req = Net::HTTP::Post.new ["/api", "chat"].join("/")
+      params   = {model: "llama3.2", stream: false}.merge!(params)
+      req      = Net::HTTP::Post.new("/api/chat", headers)
       messages = [*(params.delete(:messages) || []), LLM::Message.new(role, prompt)]
-      params = DEFAULT_PARAMS.merge(params)
-      body = {messages: messages.map(&:to_h)}.merge!(params)
-      req = preflight(req, body)
-      res = request(@http, req)
+      req.body = JSON.dump({messages: messages.map(&:to_h)}.merge!(params))
+      res      = request(@http, req)
       Response::Completion.new(res).extend(response_parser)
     end
 
+    ##
+    # @return (see LLM::Provider#assistant_role)
+    def assistant_role
+      "assistant"
+    end
+
+    ##
+    # @return (see LLM::Provider#models)
+    def models
+      @models ||= load_models!("ollama")
+    end
+
     private
 
-    def auth(req)
-      req["Authorization"] = "Bearer #{@secret}"
+    def headers
+      {
+        "Content-Type" => "application/json",
+        "Authorization" => "Bearer #{@secret}"
+      }
     end
 
     def response_parser

data/lib/llm/providers/openai/response_parser.rb

@@ -2,12 +2,14 @@
 
 class LLM::OpenAI
   module ResponseParser
+    ##
+    # @param [Hash] body
+    #  The response body from the LLM provider
+    # @return [Hash]
     def parse_embedding(body)
       {
         model: body["model"],
-        embeddings: body.dig("data").map do |data|
-          data["embedding"]
-        end,
+        embeddings: body["data"].map { _1["embedding"] },
         prompt_tokens: body.dig("usage", "prompt_tokens"),
         total_tokens: body.dig("usage", "total_tokens")
       }

data/lib/llm/providers/openai.rb

@@ -11,7 +11,6 @@ module LLM
     include Format
 
     HOST = "api.openai.com"
-    DEFAULT_PARAMS = {model: "gpt-4o-mini"}.freeze
 
     ##
     # @param secret (see LLM::Provider#initialize)
@@ -23,10 +22,9 @@ module LLM
     # @param input (see LLM::Provider#embed)
     # @return (see LLM::Provider#embed)
     def embed(input, **params)
-      req = Net::HTTP::Post.new ["/v1", "embeddings"].join("/")
-      body = {input:, model: "text-embedding-3-small"}.merge!(params)
-      req = preflight(req, body)
-      res = request @http, req
+      req = Net::HTTP::Post.new("/v1/embeddings", headers)
+      req.body = JSON.dump({input:, model: "text-embedding-3-small"}.merge!(params))
+      res = request(@http, req)
       Response::Embedding.new(res).extend(response_parser)
     end
 
@@ -36,19 +34,31 @@ module LLM
     # @param role (see LLM::Provider#complete)
     # @return (see LLM::Provider#complete)
     def complete(prompt, role = :user, **params)
-      req = Net::HTTP::Post.new ["/v1", "chat", "completions"].join("/")
+      params   = {model: "gpt-4o-mini"}.merge!(params)
+      req      = Net::HTTP::Post.new("/v1/chat/completions", headers)
       messages = [*(params.delete(:messages) || []), Message.new(role, prompt)]
-      params = DEFAULT_PARAMS.merge(params)
-      body = {messages: format(messages)}.merge!(params)
-      req = preflight(req, body)
-      res = request(@http, req)
+      req.body = JSON.dump({messages: format(messages)}.merge!(params))
+      res      = request(@http, req)
       Response::Completion.new(res).extend(response_parser)
     end
 
+    ##
+    # @return (see LLM::Provider#assistant_role)
+    def assistant_role
+      "assistant"
+    end
+
+    def models
+      @models ||= load_models!("openai")
+    end
+
     private
 
-    def auth(req)
-      req["Authorization"] = "Bearer #{@secret}"
+    def headers
+      {
+        "Content-Type" => "application/json",
+        "Authorization" => "Bearer #{@secret}"
+      }
     end
 
     def response_parser

data/lib/llm/providers/voyageai/error_handler.rb

@@ -0,0 +1,32 @@
+# frozen_string_literal: true
+
+class LLM::VoyageAI
+  class ErrorHandler
+    ##
+    # @return [Net::HTTPResponse]
+    #  Non-2XX response from the server
+    attr_reader :res
+
+    ##
+    # @param [Net::HTTPResponse] res
+    #  The response from the server
+    # @return [LLM::OpenAI::ErrorHandler]
+    def initialize(res)
+      @res = res
+    end
+
+    ##
+    # @raise [LLM::Error]
+    #  Raises a subclass of {LLM::Error LLM::Error}
+    def raise_error!
+      case res
+      when Net::HTTPUnauthorized
+        raise LLM::Error::Unauthorized.new { _1.response = res }, "Authentication error"
+      when Net::HTTPTooManyRequests
+        raise LLM::Error::RateLimit.new { _1.response = res }, "Too many requests"
+      else
+        raise LLM::Error::BadResponse.new { _1.response = res }, "Unexpected response"
+      end
+    end
+  end
+end

data/lib/llm/providers/voyageai/response_parser.rb

@@ -0,0 +1,13 @@
+# frozen_string_literal: true
+
+class LLM::VoyageAI
+  module ResponseParser
+    def parse_embedding(body)
+      {
+        model: body["model"],
+        embeddings: body["data"].map { _1["embedding"] },
+        total_tokens: body.dig("usage", "total_tokens")
+      }
+    end
+  end
+end

data/lib/llm/providers/voyageai.rb

@@ -0,0 +1,44 @@
+# frozen_string_literal: true
+
+module LLM
+  class VoyageAI < Provider
+    require_relative "voyageai/error_handler"
+    require_relative "voyageai/response_parser"
+    HOST = "api.voyageai.com"
+
+    ##
+    # @param secret (see LLM::Provider#initialize)
+    def initialize(secret, **)
+      super(secret, host: HOST, **)
+    end
+
+    ##
+    # Provides an embedding via VoyageAI per
+    # [Anthropic's recommendation](https://docs.anthropic.com/en/docs/build-with-claude/embeddings)
+    # @param input (see LLM::Provider#embed)
+    # @return (see LLM::Provider#embed)
+    def embed(input, **params)
+      req = Net::HTTP::Post.new("/v1/embeddings", headers)
+      req.body = JSON.dump({input:, model: "voyage-2"}.merge!(params))
+      res = request(@http, req)
+      Response::Embedding.new(res).extend(response_parser)
+    end
+
+    private
+
+    def headers
+      {
+        "Content-Type" => "application/json",
+        "Authorization" => "Bearer #{@secret}"
+      }
+    end
+
+    def response_parser
+      LLM::VoyageAI::ResponseParser
+    end
+
+    def error_handler
+      LLM::VoyageAI::ErrorHandler
+    end
+  end
+end

data/lib/llm/version.rb

@@ -1,5 +1,5 @@
 # frozen_string_literal: true
 
 module LLM
-  VERSION = "0.1.0"
+  VERSION = "0.2.0"
 end

data/lib/llm.rb

@@ -6,6 +6,7 @@ module LLM
   require_relative "llm/message"
   require_relative "llm/response"
   require_relative "llm/file"
+  require_relative "llm/model"
   require_relative "llm/provider"
   require_relative "llm/conversation"
   require_relative "llm/lazy_conversation"
@@ -16,32 +17,41 @@ module LLM
   ##
   # @param secret (see LLM::Anthropic#initialize)
   # @return (see LLM::Anthropic#initialize)
-  def anthropic(secret, **)
+  def anthropic(secret, options = {})
     require_relative "llm/providers/anthropic" unless defined?(LLM::Anthropic)
-    LLM::Anthropic.new(secret, **)
+    require_relative "llm/providers/voyageai" unless defined?(LLM::VoyageAI)
+    LLM::Anthropic.new(secret, **options)
+  end
+
+  ##
+  # @param secret (see LLM::VoyageAI#initialize)
+  # @return (see LLM::VoyageAI#initialize)
+  def voyageai(secret, options = {})
+    require_relative "llm/providers/voyageai" unless defined?(LLM::VoyageAI)
+    LLM::VoyageAI.new(secret, **options)
   end
 
   ##
   # @param secret (see LLM::Gemini#initialize)
   # @return (see LLM::Gemini#initialize)
-  def gemini(secret, **)
+  def gemini(secret, options = {})
     require_relative "llm/providers/gemini" unless defined?(LLM::Gemini)
-    LLM::Gemini.new(secret, **)
+    LLM::Gemini.new(secret, **options)
   end
 
   ##
   # @param host (see LLM::Ollama#initialize)
   # @return (see LLM::Ollama#initialize)
-  def ollama(secret)
+  def ollama(secret, options = {})
     require_relative "llm/providers/ollama" unless defined?(LLM::Ollama)
-    LLM::Ollama.new(secret)
+    LLM::Ollama.new(secret, **options)
   end
 
   ##
   # @param secret (see LLM::OpenAI#initialize)
   # @return (see LLM::OpenAI#initialize)
-  def openai(secret, **)
+  def openai(secret, options = {})
     require_relative "llm/providers/openai" unless defined?(LLM::OpenAI)
-    LLM::OpenAI.new(secret, **)
+    LLM::OpenAI.new(secret, **options)
   end
 end

data/llm.gemspec

@@ -25,16 +25,21 @@ Gem::Specification.new do |spec|
     "README.md", "LICENSE.txt",
     "lib/*.rb", "lib/**/*.rb",
     "spec/*.rb", "spec/**/*.rb",
-    "llm.gemspec"
+    "share/llm/models/*.yml", "llm.gemspec"
   ]
   spec.require_paths = ["lib"]
 
   spec.add_runtime_dependency "net-http", "~> 0.6.0"
   spec.add_runtime_dependency "json"
+  spec.add_runtime_dependency "yaml"
 
   spec.add_development_dependency "webmock", "~> 3.24.0"
   spec.add_development_dependency "yard", "~> 0.9.37"
   spec.add_development_dependency "kramdown", "~> 2.4"
   spec.add_development_dependency "webrick", "~> 1.8"
   spec.add_development_dependency "test-cmd.rb", "~> 0.12.0"
+  spec.add_development_dependency "rake", "~> 13.0"
+  spec.add_development_dependency "rspec", "~> 3.0"
+  spec.add_development_dependency "standard", "~> 1.40"
+  spec.add_development_dependency "vcr", "~> 6.0"
 end

data/share/llm/models/anthropic.yml

@@ -0,0 +1,35 @@
+claude-3-7-sonnet-20250219:
+  name: Claude 3.7 Sonnet
+  parameters: Unknown
+  description: Most intelligent Claude model with extended thinking and high capability
+  to_param: claude-3-7-sonnet-20250219
+
+claude-3-5-sonnet-20241022:
+  name: Claude 3.5 Sonnet (v2)
+  parameters: Unknown
+  description: High intelligence and capability; upgraded from previous Sonnet
+  to_param: claude-3-5-sonnet-20241022
+
+claude-3-5-sonnet-20240620:
+  name: Claude 3.5 Sonnet
+  parameters: Unknown
+  description: Intelligent and capable general-purpose model
+  to_param: claude-3-5-sonnet-20240620
+
+claude-3-5-haiku-20241022:
+  name: Claude 3.5 Haiku
+  parameters: Unknown
+  description: Blazing fast model for low-latency text generation
+  to_param: claude-3-5-haiku-20241022
+
+claude-3-opus-20240229:
+  name: Claude 3 Opus
+  parameters: Unknown
+  description: Top-level intelligence, fluency, and reasoning for complex tasks
+  to_param: claude-3-opus-20240229
+
+claude-3-haiku-20240307:
+  name: Claude 3 Haiku
+  parameters: Unknown
+  description: Fastest and most compact Claude model for near-instant responsiveness
+  to_param: claude-3-haiku-20240307

data/share/llm/models/gemini.yml

@@ -0,0 +1,35 @@
+gemini-2.5-pro-exp-03-25:
+  name: Gemini
+  parameters: Unknown
+  description: Enhanced thinking and reasoning, multimodal understanding, advanced coding, and more
+  to_param: gemini-2.5-pro-exp-03-25
+
+gemini-2.0-flash:
+  name: Gemini
+  parameters: Unknown
+  description: Next generation features, speed, thinking, realtime streaming, and multimodal generation
+  to_param: gemini-2.0-flash
+
+gemini-2.0-flash-lite:
+  name: Gemini
+  parameters: Unknown
+  description: Cost efficiency and low latency
+  to_param: gemini-2.0-flash-lite
+
+gemini-1.5-flash:
+  name: Gemini
+  parameters: Unknown
+  description: Fast and versatile performance across a diverse variety of tasks
+  to_param: gemini-1.5-flash
+
+gemini-1.5-flash-8b:
+  name: Gemini
+  parameters: 8B
+  description: High volume and lower intelligence tasks
+  to_param: gemini-1.5-flash-8b
+
+gemini-1.5-pro:
+  name: Gemini
+  parameters: Unknown
+  description: Complex reasoning tasks requiring more intelligence
+  to_param: gemini-1.5-pro

data/share/llm/models/ollama.yml

@@ -0,0 +1,155 @@
+---
+gemma3:1b:
+  name: Gemma
+  parameters: 1B
+  description: Lightweight version of Google's Gemma 3 language model, suitable for
+    low-resource environments
+  to_param: gemma3:1b
+gemma3:
+  name: Gemma
+  parameters: 4B
+  description: Balanced Gemma 3 model providing good accuracy with reasonable size
+  to_param: gemma3
+gemma3:12b:
+  name: Gemma
+  parameters: 12B
+  description: Larger Gemma 3 model offering improved reasoning and generation abilities
+  to_param: gemma3:12b
+gemma3:27b:
+  name: Gemma
+  parameters: 27B
+  description: High-end Gemma 3 model focused on top-tier performance and accuracy
+  to_param: gemma3:27b
+
+qwq:
+  name: QwQ
+  parameters: 32B
+  description: Large-scale model with high parameter count for complex tasks and
+    high-quality generation
+  to_param: qwq
+
+deepseek-r1:
+  name: DeepSeek-R1
+  parameters: 7B
+  description: Compact DeepSeek model optimized for research and experimentation
+  to_param: deepseek-r1
+deepseek-r1:671b:
+  name: DeepSeek-R1
+  parameters: 671B
+  description: Massive-scale DeepSeek model focused on advanced AI reasoning and
+    capabilities
+  to_param: deepseek-r1:671b
+deepseek-coder:
+  name: DeepSeek-Coder
+  parameters: 1.3B
+  description: Lightweight code generation model trained on 2T tokens of code and natural language
+  to_param: deepseek-coder
+deepseek-coder:6.7b:
+  name: DeepSeek-Coder
+  parameters: 6.7B
+  description: Mid-sized DeepSeek-Coder model offering a strong balance between speed and capability for code-related tasks
+  to_param: deepseek-coder:6.7b
+deepseek-coder:33b:
+  name: DeepSeek-Coder
+  parameters: 33B
+  description: Large DeepSeek-Coder model with high performance for code generation, understanding, and multilingual coding tasks
+  to_param: deepseek-coder:33b
+
+llama3.3:
+  name: Llama
+  parameters: 70B
+  description: Latest large Llama model designed for high-end performance in reasoning
+    and language tasks
+  to_param: llama3.3
+llama3.2:
+  name: Llama
+  parameters: 3B
+  description: Small but capable version of Llama 3.2 for lightweight applications
+  to_param: llama3.2
+llama3.2:1b:
+  name: Llama
+  parameters: 1B
+  description: Tiny version of Llama 3.2, extremely lightweight and fast
+  to_param: llama3.2:1b
+llama3.2-vision:
+  name: Llama Vision
+  parameters: 11B
+  description: Multimodal Llama 3.2 model with vision capabilities (images + text)
+  to_param: llama3.2-vision
+llama3.2-vision:90b:
+  name: Llama Vision
+  parameters: 90B
+  description: Large-scale vision-capable Llama model for advanced multimodal tasks
+  to_param: llama3.2-vision:90b
+llama3.1:
+  name: Llama
+  parameters: 8B
+  description: General-purpose Llama model designed for good accuracy and performance
+    balance
+  to_param: llama3.1
+llama3.1:405b:
+  name: Llama
+  parameters: 405B
+  description: Extremely large-scale version of Llama 3.1, suitable for advanced tasks
+  to_param: llama3.1:405b
+
+phi4:
+  name: Phi
+  parameters: 14B
+  description: Phi 4 is known for compact size and competitive performance in general
+    tasks
+  to_param: phi4
+phi4-mini:
+  name: Phi Mini
+  parameters: 3.8B
+  description: Lightweight variant of Phi 4 ideal for quick inference on constrained systems
+  to_param: phi4-mini
+
+mistral:
+  name: Mistral
+  parameters: 7B
+  description: Popular and versatile open model for general language tasks
+  to_param: mistral
+
+moondream:
+  name: Moondream
+  parameters: 1.4B
+  description: Compact vision-enabled model with strong general performance
+  to_param: moondream
+
+neural-chat:
+  name: Neural Chat
+  parameters: 7B
+  description: Chat-focused model fine-tuned for natural conversations
+  to_param: neural-chat
+
+starling-lm:
+  name: Starling
+  parameters: 7B
+  description: Model focused on instruction-following and conversational performance
+  to_param: starling-lm
+
+codellama:
+  name: Code Llama
+  parameters: 7B
+  description: Llama model variant fine-tuned specifically for code understanding
+    and generation
+  to_param: codellama
+
+llama2-uncensored:
+  name: Llama 2 Uncensored
+  parameters: 7B
+  description: Unfiltered version of Llama 2 for unrestricted language modeling
+  to_param: llama2-uncensored
+
+llava:
+  name: LLaVA
+  parameters: 7B
+  description: Multimodal model combining vision and language understanding
+  to_param: llava
+
+granite3.2:
+  name: Granite
+  parameters: 8B
+  description: IBM’s Granite model for enterprise-grade language applications
+  to_param: granite3.2

data/share/llm/models/openai.yml

@@ -0,0 +1,46 @@
+---
+o3-mini:
+  name: OpenAI o3-mini
+  parameters: Unknown
+  description: Fast, flexible, intelligent reasoning model
+  to_param: o3-mini
+o1:
+  name: OpenAI o1
+  parameters: Unknown
+  description: High-intelligence reasoning model
+  to_param: o1
+o1-mini:
+  name: OpenAI o1-mini
+  parameters: Unknown
+  description: Faster, more affordable reasoning model than o1
+  to_param: o1-mini
+o1-pro:
+  name: OpenAI o1-pro
+  parameters: Unknown
+  description: More compute than o1 for better responses
+  to_param: o1-pro
+gpt-4.5-preview:
+  name: GPT-4.5 Preview
+  parameters: Unknown
+  description: Largest and most capable GPT model
+  to_param: gpt-4.5-preview
+gpt-4o:
+  name: GPT-4o
+  parameters: Unknown
+  description: Fast, intelligent, flexible GPT model
+  to_param: gpt-4o
+gpt-4o-mini:
+  name: GPT-4o Mini
+  parameters: Mini
+  description: Fast, affordable small model for focused tasks
+  to_param: gpt-4o-mini
+gpt-4o-realtime-preview:
+  name: GPT-4o Realtime
+  parameters: Unknown
+  description: Realtime model for text and audio inputs/outputs
+  to_param: gpt-4o-realtime-preview
+gpt-3.5-turbo:
+  name: GPT-3.5 Turbo
+  parameters: Unknown
+  description: Legacy GPT model for cheaper chat and non-chat tasks
+  to_param: gpt-3.5-turbo