llm.rb 4.8.0 → 4.10.0

data/lib/llm/providers/{gemini.rb → google.rb}

@@ -2,8 +2,8 @@
 
 module LLM
   ##
-  # The Gemini class implements a provider for
-  # [Gemini](https://ai.google.dev/). The Gemini provider
+  # The Google class implements a provider for
+  # [Gemini](https://ai.google.dev/). The Google provider
   # can accept multiple inputs (text, images, audio, and video).
   # The inputs can be provided inline via the prompt for files
   # under 20MB or via the Gemini Files API for files
@@ -13,19 +13,19 @@ module LLM
   #   #!/usr/bin/env ruby
   #   require "llm"
   #
-  #   llm = LLM.gemini(key: ENV["KEY"])
-  #   ses = LLM::Session.new(llm)
-  #   ses.talk ["Tell me about this photo", ses.local_file("/images/photo.png")]
-  #   ses.messages.select(&:assistant?).each { print "[#{_1.role}]", _1.content, "\n" }
-  class Gemini < Provider
-    require_relative "gemini/error_handler"
-    require_relative "gemini/request_adapter"
-    require_relative "gemini/response_adapter"
-    require_relative "gemini/stream_parser"
-    require_relative "gemini/models"
-    require_relative "gemini/images"
-    require_relative "gemini/audio"
-    require_relative "gemini/files"
+  #   llm = LLM.google(key: ENV["KEY"])
+  #   ctx = LLM::Context.new(llm)
+  #   ctx.talk ["Tell me about this photo", ctx.local_file("/images/photo.png")]
+  #   ctx.messages.select(&:assistant?).each { print "[#{_1.role}]", _1.content, "\n" }
+  class Google < Provider
+    require_relative "google/error_handler"
+    require_relative "google/request_adapter"
+    require_relative "google/response_adapter"
+    require_relative "google/stream_parser"
+    require_relative "google/models"
+    require_relative "google/images"
+    require_relative "google/audio"
+    require_relative "google/files"
 
     include RequestAdapter
 
@@ -37,6 +37,13 @@ module LLM
       super(host: HOST, **)
     end
 
+    ##
+    # @return [Symbol]
+    #  Returns the provider's name
+    def name
+      :google
+    end
+
     ##
     # Provides an embedding
     # @param input (see LLM::Provider#embed)
@@ -78,33 +85,33 @@ module LLM
     ##
     # Provides an interface to Gemini's audio API
     # @see https://ai.google.dev/gemini-api/docs/audio Gemini docs
-    # @return [LLM::Gemini::Audio]
+    # @return [LLM::Google::Audio]
     def audio
-      LLM::Gemini::Audio.new(self)
+      LLM::Google::Audio.new(self)
     end
 
     ##
     # Provides an interface to Gemini's image generation API
     # @see https://ai.google.dev/gemini-api/docs/image-generation Gemini docs
-    # @return [see LLM::Gemini::Images]
+    # @return [see LLM::Google::Images]
     def images
-      LLM::Gemini::Images.new(self)
+      LLM::Google::Images.new(self)
     end
 
     ##
     # Provides an interface to Gemini's file management API
     # @see https://ai.google.dev/gemini-api/docs/files Gemini docs
-    # @return [LLM::Gemini::Files]
+    # @return [LLM::Google::Files]
     def files
-      LLM::Gemini::Files.new(self)
+      LLM::Google::Files.new(self)
     end
 
     ##
     # Provides an interface to Gemini's models API
     # @see https://ai.google.dev/gemini-api/docs/models Gemini docs
-    # @return [LLM::Gemini::Models]
+    # @return [LLM::Google::Models]
     def models
-      LLM::Gemini::Models.new(self)
+      LLM::Google::Models.new(self)
     end
 
     ##
@@ -177,11 +184,11 @@ module LLM
     end
 
     def stream_parser
-      LLM::Gemini::StreamParser
+      LLM::Google::StreamParser
     end
 
     def error_handler
-      LLM::Gemini::ErrorHandler
+      LLM::Google::ErrorHandler
     end
 
     def normalize_complete_params(params)
@@ -197,10 +204,18 @@ module LLM
       model.respond_to?(:id) ? model.id : model
       path = ["/v1beta/models/#{model}", action].join(":")
       req  = Net::HTTP::Post.new(path, headers)
-      messages = [*(params.delete(:messages) || []), LLM::Message.new(role, prompt)]
+      messages = build_complete_messages(prompt, params, role)
       body = LLM.json.dump({contents: adapt(messages)}.merge!(params))
       set_body_stream(req, StringIO.new(body))
       req
     end
+
+    def build_complete_messages(prompt, params, role)
+      if LLM::Prompt === prompt
+        [*(params.delete(:messages) || []), *prompt.to_a]
+      else
+        [*(params.delete(:messages) || []), LLM::Message.new(role, prompt)]
+      end
+    end
   end
 end

data/lib/llm/providers/llamacpp.rb

@@ -16,9 +16,9 @@ module LLM
   #   require "llm"
   #
   #   llm = LLM.llamacpp(key: nil)
-  #   ses = LLM::Session.new(llm)
-  #   ses.talk ["Tell me about this photo", ses.local_file("/images/photo.png")]
-  #   ses.messages.select(&:assistant?).each { print "[#{_1.role}]", _1.content, "\n" }
+  #   ctx = LLM::Context.new(llm)
+  #   ctx.talk ["Tell me about this photo", ctx.local_file("/images/photo.png")]
+  #   ctx.messages.select(&:assistant?).each { print "[#{_1.role}]", _1.content, "\n" }
   class LlamaCpp < OpenAI
     ##
     # @param (see LLM::Provider#initialize)
@@ -27,6 +27,13 @@ module LLM
       super
     end
 
+    ##
+    # @return [Symbol]
+    #  Returns the provider's name
+    def name
+      :llamacpp
+    end
+
     ##
     # @raise [NotImplementedError]
     def files

data/lib/llm/providers/ollama.rb

@@ -12,9 +12,9 @@ module LLM
   #   require "llm"
   #
   #   llm = LLM.ollama(key: nil)
-  #   ses = LLM::Session.new(llm, model: "llava")
-  #   ses.talk ["Tell me about this image", ses.local_file("/images/photo.png")]
-  #   ses.messages.select(&:assistant?).each { print "[#{_1.role}]", _1.content, "\n" }
+  #   ctx = LLM::Context.new(llm, model: "llava")
+  #   ctx.talk ["Tell me about this image", ctx.local_file("/images/photo.png")]
+  #   ctx.messages.select(&:assistant?).each { print "[#{_1.role}]", _1.content, "\n" }
   class Ollama < Provider
     require_relative "ollama/error_handler"
     require_relative "ollama/request_adapter"
@@ -32,6 +32,13 @@ module LLM
       super(host: HOST, port: 11434, ssl: false, **)
     end
 
+    ##
+    # @return [Symbol]
+    #  Returns the provider's name
+    def name
+      :ollama
+    end
+
     ##
     # Provides an embedding
     # @param input (see LLM::Provider#embed)
@@ -120,11 +127,19 @@ module LLM
     end
 
     def build_complete_request(prompt, params, role)
-      messages = [*(params.delete(:messages) || []), LLM::Message.new(role, prompt)]
+      messages = build_complete_messages(prompt, params, role)
       body = LLM.json.dump({messages: [adapt(messages)].flatten}.merge!(params))
       req = Net::HTTP::Post.new("/api/chat", headers)
       set_body_stream(req, StringIO.new(body))
       req
     end
+
+    def build_complete_messages(prompt, params, role)
+      if LLM::Prompt === prompt
+        [*(params.delete(:messages) || []), *prompt.to_a]
+      else
+        [*(params.delete(:messages) || []), LLM::Message.new(role, prompt)]
+      end
+    end
   end
 end

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

@@ -13,10 +13,10 @@ class LLM::OpenAI
   #   require "llm"
   #
   #   llm = LLM.openai(key: ENV["KEY"])
-  #   ses = LLM::Session.new(llm)
+  #   ctx = LLM::Context.new(llm)
   #   file = llm.files.create file: "/books/goodread.pdf"
-  #   ses.talk ["Tell me about this PDF", file]
-  #   ses.messages.select(&:assistant?).each { print "[#{_1.role}]", _1.content, "\n" }
+  #   ctx.talk ["Tell me about this PDF", file]
+  #   ctx.messages.select(&:assistant?).each { print "[#{_1.role}]", _1.content, "\n" }
   class Files
     ##
     # Returns a new Files object

data/lib/llm/providers/openai/response_adapter/completion.rb

@@ -74,10 +74,18 @@ module LLM::OpenAI::ResponseAdapter
     def adapt_tool_calls(tools)
       (tools || []).filter_map do |tool|
         next unless tool.function
-        {id: tool.id, name: tool.function.name, arguments: LLM.json.load(tool.function.arguments)}
+        {id: tool.id, name: tool.function.name, arguments: parse_tool_arguments(tool.function.arguments)}
       end
     end
 
+    def parse_tool_arguments(arguments)
+      return {} if arguments.to_s.empty?
+      parsed = LLM.json.load(arguments)
+      Hash === parsed ? parsed : {}
+    rescue *LLM.json.parser_error
+      {}
+    end
+
     include LLM::Contract::Completion
   end
 end

data/lib/llm/providers/openai/response_adapter/responds.rb

@@ -38,7 +38,15 @@ module LLM::OpenAI::ResponseAdapter
     end
 
     def adapt_tool(tool)
-      {id: tool.call_id, name: tool.name, arguments: LLM.json.load(tool.arguments)}
+      {id: tool.call_id, name: tool.name, arguments: parse_tool_arguments(tool.arguments)}
+    end
+
+    def parse_tool_arguments(arguments)
+      return {} if arguments.to_s.empty?
+      parsed = LLM.json.load(arguments)
+      Hash === parsed ? parsed : {}
+    rescue *LLM.json.parser_error
+      {}
     end
   end
 end

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

@@ -41,7 +41,7 @@ class LLM::OpenAI
       role, stream = params.delete(:role), params.delete(:stream)
       params[:stream] = true if stream.respond_to?(:<<) || stream == true
       req = Net::HTTP::Post.new("/v1/responses", headers)
-      messages = [*(params.delete(:input) || []), LLM::Message.new(role, prompt)]
+      messages = build_complete_messages(prompt, params, role)
       @provider.tracer.set_request_metadata(user_input: extract_user_input(messages, fallback: prompt))
       body = LLM.json.dump({input: [adapt(messages, mode: :response)].flatten}.merge!(params))
       set_body_stream(req, StringIO.new(body))
@@ -89,6 +89,14 @@ class LLM::OpenAI
       define_method(m) { |*args, **kwargs, &b| @provider.send(m, *args, **kwargs, &b) }
     end
 
+    def build_complete_messages(prompt, params, role)
+      if LLM::Prompt === prompt
+        [*(params.delete(:input) || []), *prompt]
+      else
+        [*(params.delete(:input) || []), LLM::Message.new(role, prompt)]
+      end
+    end
+
     def adapt_schema(params)
       return {} unless params && params[:schema]
       schema = params.delete(:schema)

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

@@ -43,6 +43,7 @@ class LLM::OpenAI
           target_message = @body["choices"][index]["message"]
           delta = choice["delta"] || {}
           delta.each do |key, value|
+            next if value.nil?
             if key == "content"
               target_message[key] ||= +""
               target_message[key] << value
@@ -57,6 +58,7 @@ class LLM::OpenAI
           message_hash = {"role" => "assistant"}
           @body["choices"][index] = {"message" => message_hash}
           (choice["delta"] || {}).each do |key, value|
+            next if value.nil?
             if key == "content"
               @io << value if @io.respond_to?(:<<)
               message_hash[key] = value

data/lib/llm/providers/openai.rb

@@ -10,9 +10,9 @@ module LLM
   #   require "llm"
   #
   #   llm = LLM.openai(key: ENV["KEY"])
-  #   ses = LLM::Session.new(llm)
-  #   ses.talk ["Tell me about this photo", ses.local_file("/images/photo.png")]
-  #   ses.messages.select(&:assistant?).each { print "[#{_1.role}]", _1.content, "\n" }
+  #   ctx = LLM::Context.new(llm)
+  #   ctx.talk ["Tell me about this photo", ctx.local_file("/images/photo.png")]
+  #   ctx.messages.select(&:assistant?).each { print "[#{_1.role}]", _1.content, "\n" }
   class OpenAI < Provider
     require_relative "openai/error_handler"
     require_relative "openai/request_adapter"
@@ -36,6 +36,13 @@ module LLM
       super(host: HOST, **)
     end
 
+    ##
+    # @return [Symbol]
+    #  Returns the provider's name
+    def name
+      :openai
+    end
+
     ##
     # Provides an embedding
     # @see https://platform.openai.com/docs/api-reference/embeddings/create OpenAI docs
@@ -213,13 +220,21 @@ module LLM
     end
 
     def build_complete_request(prompt, params, role)
-      messages = [*(params.delete(:messages) || []), Message.new(role, prompt)]
+      messages = build_complete_messages(prompt, params, role)
       body = LLM.json.dump({messages: adapt(messages, mode: :complete).flatten}.merge!(params))
       req = Net::HTTP::Post.new(completions_path, headers)
       set_body_stream(req, StringIO.new(body))
       [req, messages]
     end
 
+    def build_complete_messages(prompt, params, role)
+      if LLM::Prompt === prompt
+        [*(params.delete(:messages) || []), *prompt]
+      else
+        [*(params.delete(:messages) || []), Message.new(role, prompt)]
+      end
+    end
+
     def extract_user_input(messages, fallback:)
       message = messages.reverse.find(&:user?) || messages.last
       value = message&.content || fallback

data/lib/llm/providers/xai.rb

@@ -11,9 +11,9 @@ module LLM
   #   require "llm"
   #
   #   llm = LLM.xai(key: ENV["KEY"])
-  #   ses = LLM::Session.new(llm)
-  #   ses.talk ["Tell me about this photo", ses.local_file("/images/photo.png")]
-  #   ses.messages.select(&:assistant?).each { print "[#{_1.role}]", _1.content, "\n" }
+  #   ctx = LLM::Context.new(llm)
+  #   ctx.talk ["Tell me about this photo", ctx.local_file("/images/photo.png")]
+  #   ctx.messages.select(&:assistant?).each { print "[#{_1.role}]", _1.content, "\n" }
   class XAI < OpenAI
     require_relative "xai/images"
 
@@ -25,6 +25,13 @@ module LLM
       super
     end
 
+    ##
+    # @return [Symbol]
+    #  Returns the provider's name
+    def name
+      :xai
+    end
+
     ##
     # @raise [NotImplementedError]
     def files

data/lib/llm/providers/zai.rb

@@ -11,8 +11,8 @@ module LLM
   #   require "llm"
   #
   #   llm = LLM.zai(key: ENV["KEY"])
-  #   ses = LLM::Session.new(llm, stream: $stdout)
-  #   ses.talk "Hello"
+  #   ctx = LLM::Context.new(llm, stream: $stdout)
+  #   ctx.talk "Hello"
   class ZAI < OpenAI
     ##
     # @param [String] host A regional host or the default ("api.z.ai")
@@ -21,6 +21,13 @@ module LLM
       super
     end
 
+    ##
+    # @return [Symbol]
+    #  Returns the provider's name
+    def name
+      :zai
+    end
+
     ##
     # @raise [NotImplementedError]
     def files

data/lib/llm/registry.rb

@@ -0,0 +1,81 @@
+# frozen_string_literal: true
+
+##
+# The {LLM::Registry LLM::Registry} class provides a small API over
+# provider model data. It exposes model metadata such as pricing,
+# capabilities, modalities, and limits from the registry files
+# stored under `data/`. The data is provided by https://models.dev
+# and shipped with llm.rb.
+class LLM::Registry
+  @root = File.join(__dir__, "..", "..")
+
+  ##
+  # @raise [LLM::Error]
+  #  Might raise an error
+  # @param [Symbol]
+  #  A provider name
+  # @return [LLM::Registry]
+  def self.for(name)
+    path = File.join @root, "data", "#{name}.json"
+    if File.file?(path)
+      new LLM.json.load(File.binread(path))
+    else
+      raise LLM::NoSuchRegistryError, "no registry found for #{name}"
+    end
+  end
+
+  ##
+  # @param [Hash] blob
+  #  A model registry
+  # @return [LLM::Registry]
+  def initialize(blob)
+    @registry = LLM::Object.from(blob)
+    @models = @registry.models
+  end
+
+  ##
+  # @return [LLM::Object]
+  #  Returns model costs
+  def cost(model:)
+    lookup(model:).cost
+  end
+
+  ##
+  # @return [LLM::Object]
+  #  Returns model modalities
+  def modalities(model:)
+    lookup(model:).modalities
+  end
+
+  ##
+  # @return [LLM::Object]
+  #  Returns model limits such as the context window size
+  def limit(model:)
+    lookup(model:).limit
+  end
+
+  private
+
+  def lookup(model:)
+    if @models.key?(model)
+      @models[model]
+    else
+      patterns = {/-\d{4}-\d{2}-\d{2}$/ => "", /\A(gpt-.*)-\d{4}$/ => "\\1"}
+      fallback = find_map(patterns) { model.dup.sub!(_1, _2) } || "none"
+      if @models.key?(fallback)
+        @models[fallback]
+      else
+        raise LLM::NoSuchModelError, "no such model: #{model} (fallback: #{fallback})"
+      end
+    end
+  end
+
+  ##
+  # Similar to #{find} but returns the block's return value
+  # @return [Object, nil]
+  def find_map(pair)
+    result = nil
+    pair.each_pair { break if result = yield(_1, _2) }
+    result
+  end
+end

data/lib/llm/schema/all_of.rb

@@ -0,0 +1,31 @@
+# frozen_string_literal: true
+
+class LLM::Schema
+  ##
+  # The {LLM::Schema::AllOf LLM::Schema::AllOf} class represents an
+  # allOf union in a JSON schema. It is a subclass of
+  # {LLM::Schema::Leaf LLM::Schema::Leaf}.
+  class AllOf < Leaf
+    ##
+    # Returns an allOf union for the given types.
+    # @return [LLM::Schema::AllOf]
+    def self.[](*types)
+      schema = LLM::Schema.new
+      new(types.map { LLM::Schema::Utils.resolve(schema, _1) })
+    end
+
+    ##
+    # @param [Array<LLM::Schema::Leaf>] values
+    #  The values required by the union
+    # @return [LLM::Schema::AllOf]
+    def initialize(values)
+      @values = values
+    end
+
+    ##
+    # @return [Hash]
+    def to_h
+      super.merge!(allOf: @values)
+    end
+  end
+end

data/lib/llm/schema/any_of.rb

@@ -0,0 +1,31 @@
+# frozen_string_literal: true
+
+class LLM::Schema
+  ##
+  # The {LLM::Schema::AnyOf LLM::Schema::AnyOf} class represents an
+  # anyOf union in a JSON schema. It is a subclass of
+  # {LLM::Schema::Leaf LLM::Schema::Leaf}.
+  class AnyOf < Leaf
+    ##
+    # Returns an anyOf union for the given types.
+    # @return [LLM::Schema::AnyOf]
+    def self.[](*types)
+      schema = LLM::Schema.new
+      new(types.map { LLM::Schema::Utils.resolve(schema, _1) })
+    end
+
+    ##
+    # @param [Array<LLM::Schema::Leaf>] values
+    #  The values allowed by the union
+    # @return [LLM::Schema::AnyOf]
+    def initialize(values)
+      @values = values
+    end
+
+    ##
+    # @return [Hash]
+    def to_h
+      super.merge!(anyOf: @values)
+    end
+  end
+end

data/lib/llm/schema/one_of.rb

@@ -0,0 +1,31 @@
+# frozen_string_literal: true
+
+class LLM::Schema
+  ##
+  # The {LLM::Schema::OneOf LLM::Schema::OneOf} class represents an
+  # oneOf union in a JSON schema. It is a subclass of
+  # {LLM::Schema::Leaf LLM::Schema::Leaf}.
+  class OneOf < Leaf
+    ##
+    # Returns a oneOf union for the given types.
+    # @return [LLM::Schema::OneOf]
+    def self.[](*types)
+      schema = LLM::Schema.new
+      new(types.map { LLM::Schema::Utils.resolve(schema, _1) })
+    end
+
+    ##
+    # @param [Array<LLM::Schema::Leaf>] values
+    #  The values allowed by the union
+    # @return [LLM::Schema::OneOf]
+    def initialize(values)
+      @values = values
+    end
+
+    ##
+    # @return [Hash]
+    def to_h
+      super.merge!(oneOf: @values)
+    end
+  end
+end