llm.rb 0.4.2 → 0.6.0

data/lib/llm/providers/anthropic/format/completion_format.rb

@@ -0,0 +1,73 @@
+# frozen_string_literal: true
+
+module LLM::Anthropic::Format
+  ##
+  # @private
+  class CompletionFormat
+    ##
+    # @param [LLM::Message, Hash] message
+    #  The message to format
+    def initialize(message)
+      @message = message
+    end
+
+    ##
+    # Formats the message for the Anthropic chat completions API
+    # @return [Hash]
+    def format
+      catch(:abort) do
+        if Hash === message
+          {role: message[:role], content: format_content(message[:content])}
+        else
+          format_message
+        end
+      end
+    end
+
+    private
+
+    def format_message
+      if message.tool_call?
+        {role: message.role, content: message.extra[:original_tool_calls]}
+      else
+        {role: message.role, content: format_content(content)}
+      end
+    end
+
+    ##
+    # @param [String, URI] content
+    #  The content to format
+    # @return [String, Hash]
+    #  The formatted content
+    def format_content(content)
+      case content
+      when Hash
+        content.empty? ? throw(:abort, nil) : [content]
+      when Array
+        content.empty? ? throw(:abort, nil) : content.flat_map { format_content(_1) }
+      when URI
+        [{type: :image, source: {type: "url", url: content.to_s}}]
+      when LLM::File
+        if content.image?
+          [{type: :image, source: {type: "base64", media_type: content.mime_type, data: content.to_b64}}]
+        else
+          raise LLM::Error::PromptError, "The given object (an instance of #{content.class}) " \
+                                          "is not an image, and therefore not supported by the " \
+                                          "Anthropic API"
+        end
+      when String
+        [{type: :text, text: content}]
+      when LLM::Message
+        format_content(content.content)
+      when LLM::Function::Return
+        [{type: "tool_result", tool_use_id: content.id, content: [{type: :text, text: JSON.dump(content.value)}]}]
+      else
+        raise LLM::Error::PromptError, "The given object (an instance of #{content.class}) " \
+                                       "is not supported by the Anthropic API"
+      end
+    end
+
+    def message = @message
+    def content = message.content
+  end
+end

data/lib/llm/providers/anthropic/format.rb

@@ -4,49 +4,24 @@ class LLM::Anthropic
   ##
   # @private
   module Format
+    require_relative "format/completion_format"
+
     ##
     # @param [Array<LLM::Message>] messages
     #  The messages to format
     # @return [Array<Hash>]
     def format(messages)
-      messages.map do
-        if Hash === _1
-          {role: _1[:role], content: format_content(_1[:content])}
-        else
-          {role: _1.role, content: format_content(_1.content)}
-        end
+      messages.filter_map do
+        CompletionFormat.new(_1).format
       end
     end
 
     private
 
-    ##
-    # @param [String, URI] content
-    #  The content to format
-    # @return [String, Hash]
-    #  The formatted content
-    def format_content(content)
-      case content
-      when Array
-        content.flat_map { format_content(_1) }
-      when URI
-        [{type: :image, source: {type: "url", url: content.to_s}}]
-      when LLM::File
-        if content.image?
-          [{type: :image, source: {type: "base64", media_type: content.mime_type, data: content.to_b64}}]
-        else
-          raise LLM::Error::PromptError, "The given object (an instance of #{content.class}) " \
-                                          "is not an image, and therefore not supported by the " \
-                                          "Anthropic API"
-        end
-      when String
-        [{type: :text, text: content}]
-      when LLM::Message
-        format_content(content.content)
-      else
-        raise LLM::Error::PromptError, "The given object (an instance of #{content.class}) " \
-                                       "is not supported by the Anthropic API"
-      end
+    def format_tools(params)
+      return {} unless params and params[:tools]&.any?
+      tools = params[:tools]
+      {tools: tools.map { _1.format(self) }}
     end
   end
 end

data/lib/llm/providers/anthropic/response_parser/completion_parser.rb

@@ -0,0 +1,51 @@
+# frozen_string_literal: true
+
+module LLM::Anthropic::ResponseParser
+  ##
+  # @private
+  class CompletionParser
+    def initialize(body)
+      @body = OpenStruct.from_hash(body)
+    end
+
+    def format(response)
+      {
+        model:,
+        prompt_tokens:,
+        completion_tokens:,
+        total_tokens:,
+        choices: format_choices(response)
+      }
+    end
+
+    private
+
+    def format_choices(response)
+      texts.map.with_index do |choice, index|
+        extra = {index:, response:, tool_calls: format_tool_calls(tools), original_tool_calls: tools}
+        LLM::Message.new(role, choice.text, extra)
+      end
+    end
+
+    def format_tool_calls(tools)
+      (tools || []).filter_map do |tool|
+        tool = {
+          id: tool.id,
+          name: tool.name,
+          arguments: tool.input
+        }
+        OpenStruct.new(tool)
+      end
+    end
+
+    def body = @body
+    def role = body.role
+    def model = body.model
+    def prompt_tokens = body.usage.input_tokens
+    def completion_tokens = body.usage.output_tokens
+    def total_tokens = body.usage.total_tokens
+    def parts = body.content
+    def texts = parts.select { _1["type"] == "text" }
+    def tools = parts.select { _1["type"] == "tool_use" }
+  end
+end

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

@@ -17,15 +17,7 @@ class LLM::Anthropic
     #  The response body from the LLM provider
     # @return [Hash]
     def parse_completion(body)
-      {
-        model: body["model"],
-        choices: body["content"].map do
-          # TODO: don't hardcode role
-          LLM::Message.new("assistant", _1["text"], {response: self})
-        end,
-        prompt_tokens: body.dig("usage", "input_tokens"),
-        completion_tokens: body.dig("usage", "output_tokens")
-      }
+      CompletionParser.new(body).format(self)
     end
   end
 end

data/lib/llm/providers/anthropic.rb

@@ -7,6 +7,7 @@ module LLM
   class Anthropic < Provider
     require_relative "anthropic/error_handler"
     require_relative "anthropic/response_parser"
+    require_relative "anthropic/response_parser/completion_parser"
     require_relative "anthropic/format"
     require_relative "anthropic/models"
     include Format
@@ -14,25 +15,25 @@ module LLM
     HOST = "api.anthropic.com"
 
     ##
-    # @param secret (see LLM::Provider#initialize)
-    def initialize(secret, **)
-      super(secret, host: HOST, **)
+    # @param key (see LLM::Provider#initialize)
+    def initialize(**)
+      super(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)
-    # @param [String] token
-    #  Valid token for the VoyageAI API
+    # @param [String] key
+    #  Valid key for the VoyageAI API
     # @param [String] model
     #  The embedding model to use
     # @param [Hash] params
     #  Other embedding parameters
     # @raise (see LLM::Provider#request)
     # @return (see LLM::Provider#embed)
-    def embed(input, token:, model: "voyage-2", **params)
-      llm = LLM.voyageai(token)
+    def embed(input, key:, model: "voyage-2", **params)
+      llm = LLM.voyageai(key:)
       llm.embed(input, **params.merge(model:))
     end
 
@@ -40,20 +41,19 @@ module LLM
     # Provides an interface to the chat completions API
     # @see https://docs.anthropic.com/en/api/messages Anthropic docs
     # @param prompt (see LLM::Provider#complete)
-    # @param role (see LLM::Provider#complete)
-    # @param model (see LLM::Provider#complete)
-    # @param max_tokens The maximum number of tokens to generate
     # @param params (see LLM::Provider#complete)
     # @example (see LLM::Provider#complete)
     # @raise (see LLM::Provider#request)
     # @raise [LLM::Error::PromptError]
     #  When given an object a provider does not understand
     # @return (see LLM::Provider#complete)
-    def complete(prompt, role = :user, model: default_model, max_tokens: 1024, **params)
-      params = {max_tokens:, model:}.merge!(params)
+    def complete(prompt, params = {})
+      params = {role: :user, model: default_model, max_tokens: 1024}.merge!(params)
+      params = [params, format_tools(params)].inject({}, &:merge!).compact
+      role = params.delete(:role)
       req = Net::HTTP::Post.new("/v1/messages", headers)
       messages = [*(params.delete(:messages) || []), Message.new(role, prompt)]
-      body = JSON.dump({messages: format(messages)}.merge!(params))
+      body = JSON.dump({messages: [format(messages)].flatten}.merge!(params))
       set_body_stream(req, StringIO.new(body))
       res = request(@http, req)
       Response::Completion.new(res).extend(response_parser)
@@ -86,7 +86,7 @@ module LLM
     def headers
       {
         "Content-Type" => "application/json",
-        "x-api-key" => @secret,
+        "x-api-key" => @key,
         "anthropic-version" => "2023-06-01"
       }
     end

data/lib/llm/providers/gemini/audio.rb

@@ -9,7 +9,7 @@ class LLM::Gemini
   #   require "llm"
   #
   #   llm = LLM.gemini(ENV["KEY"])
-  #   res = llm.audio.create_transcription(input: LLM::File("/rocket.mp3"))
+  #   res = llm.audio.create_transcription(input: "/audio/rocket.mp3")
   #   res.text # => "A dog on a rocket to the moon"
   class Audio
     ##
@@ -31,10 +31,10 @@ class LLM::Gemini
     # Create an audio transcription
     # @example
     #   llm = LLM.gemini(ENV["KEY"])
-    #   res = llm.audio.create_transcription(file: LLM::File("/rocket.mp3"))
+    #   res = llm.audio.create_transcription(file: "/audio/rocket.mp3")
     #   res.text # => "A dog on a rocket to the moon"
     # @see https://ai.google.dev/gemini-api/docs/audio Gemini docs
-    # @param [LLM::File, LLM::Response::File] file The input audio
+    # @param [String, LLM::File, LLM::Response::File] file The input audio
     # @param [String] model The model to use
     # @param [Hash] params Other parameters (see Gemini docs)
     # @raise (see LLM::Provider#request)
@@ -43,8 +43,8 @@ class LLM::Gemini
       res = @provider.complete [
         "Your task is to transcribe the contents of an audio file",
         "Your response should include the transcription, and nothing else",
-        file
-      ], :user, model:, **params
+        LLM.File(file)
+      ], params.merge(role: :user, model:)
       LLM::Response::AudioTranscription
         .new(res)
         .tap { _1.text = res.choices[0].content }
@@ -55,10 +55,10 @@ class LLM::Gemini
     # @example
     #   # Arabic => English
     #   llm = LLM.gemini(ENV["KEY"])
-    #   res = llm.audio.create_translation(file: LLM::File("/bismillah.mp3"))
+    #   res = llm.audio.create_translation(file: "/audio/bismillah.mp3")
     #   res.text # => "In the name of Allah, the Beneficent, the Merciful."
     # @see https://ai.google.dev/gemini-api/docs/audio Gemini docs
-    # @param [LLM::File, LLM::Response::File] file The input audio
+    # @param [String, LLM::File, LLM::Response::File] file The input audio
     # @param [String] model The model to use
     # @param [Hash] params Other parameters (see Gemini docs)
     # @raise (see LLM::Provider#request)
@@ -67,8 +67,8 @@ class LLM::Gemini
       res = @provider.complete [
         "Your task is to translate the contents of an audio file into English",
         "Your response should include the translation, and nothing else",
-        file
-      ], :user, model:, **params
+        LLM.File(file)
+      ], params.merge(role: :user, model:)
       LLM::Response::AudioTranslation
         .new(res)
         .tap { _1.text = res.choices[0].content }

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

@@ -19,7 +19,7 @@ class LLM::Gemini
   #
   #   llm = LLM.gemini(ENV["KEY"])
   #   bot = LLM::Chat.new(llm).lazy
-  #   file = llm.files.create file: LLM::File("/audio/haiku.mp3")
+  #   file = llm.files.create file: "/audio/haiku.mp3"
   #   bot.chat(file)
   #   bot.chat("Describe the audio file I sent to you")
   #   bot.chat("The audio file is the first message I sent to you.")
@@ -30,7 +30,7 @@ class LLM::Gemini
   #
   #   llm = LLM.gemini(ENV["KEY"])
   #   bot = LLM::Chat.new(llm).lazy
-  #   file = llm.files.create file: LLM::File("/audio/haiku.mp3")
+  #   file = llm.files.create file: "/audio/haiku.mp3"
   #   bot.chat(["Describe the audio file I sent to you", file])
   #   bot.messages.select(&:assistant?).each { print "[#{_1.role}]", _1.content, "\n" }
   class Files
@@ -55,7 +55,7 @@ class LLM::Gemini
     # @raise (see LLM::Provider#request)
     # @return [LLM::Response::FileList]
     def all(**params)
-      query = URI.encode_www_form(params.merge!(key: secret))
+      query = URI.encode_www_form(params.merge!(key: key))
       req = Net::HTTP::Get.new("/v1beta/files?#{query}", headers)
       res = request(http, req)
       LLM::Response::FileList.new(res).tap { |filelist|
@@ -71,13 +71,14 @@ class LLM::Gemini
     # Create a file
     # @example
     #   llm = LLM.gemini(ENV["KEY"])
-    #   res = llm.files.create file: LLM::File("/audio/haiku.mp3"),
+    #   res = llm.files.create file: "/audio/haiku.mp3"
     # @see https://ai.google.dev/gemini-api/docs/files Gemini docs
-    # @param [File] file The file
+    # @param [String, LLM::File] file The file
     # @param [Hash] params Other parameters (see Gemini docs)
     # @raise (see LLM::Provider#request)
     # @return [LLM::Response::File]
     def create(file:, **params)
+      file = LLM.File(file)
       req = Net::HTTP::Post.new(request_upload_url(file:), {})
       req["content-length"] = file.bytesize
       req["X-Goog-Upload-Offset"] = 0
@@ -102,7 +103,7 @@ class LLM::Gemini
     # @return [LLM::Response::File]
     def get(file:, **params)
       file_id = file.respond_to?(:name) ? file.name : file.to_s
-      query = URI.encode_www_form(params.merge!(key: secret))
+      query = URI.encode_www_form(params.merge!(key: key))
       req = Net::HTTP::Get.new("/v1beta/#{file_id}?#{query}", headers)
       res = request(http, req)
       LLM::Response::File.new(res)
@@ -120,7 +121,7 @@ class LLM::Gemini
     # @return [LLM::Response::File]
     def delete(file:, **params)
       file_id = file.respond_to?(:name) ? file.name : file.to_s
-      query = URI.encode_www_form(params.merge!(key: secret))
+      query = URI.encode_www_form(params.merge!(key: key))
       req = Net::HTTP::Delete.new("/v1beta/#{file_id}?#{query}", headers)
       request(http, req)
     end
@@ -137,7 +138,7 @@ class LLM::Gemini
     include LLM::Utils
 
     def request_upload_url(file:)
-      req = Net::HTTP::Post.new("/upload/v1beta/files?key=#{secret}", headers)
+      req = Net::HTTP::Post.new("/upload/v1beta/files?key=#{key}", headers)
       req["X-Goog-Upload-Protocol"] = "resumable"
       req["X-Goog-Upload-Command"] = "start"
       req["X-Goog-Upload-Header-Content-Length"] = file.bytesize
@@ -151,8 +152,8 @@ class LLM::Gemini
       @provider.instance_variable_get(:@http)
     end
 
-    def secret
-      @provider.instance_variable_get(:@secret)
+    def key
+      @provider.instance_variable_get(:@key)
     end
 
     [:headers, :request, :set_body_stream].each do |m|

data/lib/llm/providers/gemini/format/completion_format.rb

@@ -0,0 +1,54 @@
+# frozen_string_literal: true
+
+module LLM::Gemini::Format
+  ##
+  # @private
+  class CompletionFormat
+    ##
+    # @param [LLM::Message, Hash] message
+    #  The message to format
+    def initialize(message)
+      @message = message
+    end
+
+    ##
+    # Formats the message for the Gemini chat completions API
+    # @return [Hash]
+    def format
+      catch(:abort) do
+        if Hash === message
+          {role: message[:role], parts: format_content(message[:content])}
+        elsif message.tool_call?
+          {role: message.role, parts: message.extra[:original_tool_calls].map { {"functionCall" => _1} }}
+        else
+          {role: message.role, parts: format_content(message.content)}
+        end
+      end
+    end
+
+    def format_content(content)
+      case content
+      when Array
+        content.empty? ? throw(:abort, nil) : content.flat_map { format_content(_1) }
+      when LLM::Response::File
+        file = content
+        [{file_data: {mime_type: file.mime_type, file_uri: file.uri}}]
+      when LLM::File
+        file = content
+        [{inline_data: {mime_type: file.mime_type, data: file.to_b64}}]
+      when String
+        [{text: content}]
+      when LLM::Message
+        format_content(content.content)
+      when LLM::Function::Return
+        [{text: JSON.dump(content.value)}]
+      else
+        raise LLM::Error::PromptError, "The given object (an instance of #{content.class}) " \
+                                       "is not supported by the Gemini API"
+      end
+    end
+
+    def message = @message
+    def content = message.content
+  end
+end

data/lib/llm/providers/gemini/format.rb

@@ -4,45 +4,38 @@ class LLM::Gemini
   ##
   # @private
   module Format
+    require_relative "format/completion_format"
+
     ##
     # @param [Array<LLM::Message>] messages
     #  The messages to format
     # @return [Array<Hash>]
     def format(messages)
-      messages.map do
-        if Hash === _1
-          {role: _1[:role], parts: [format_content(_1[:content])]}
-        else
-          {role: _1.role, parts: [format_content(_1.content)]}
-        end
+      messages.filter_map do |message|
+        CompletionFormat.new(message).format
       end
     end
 
     private
 
     ##
-    # @param [String, Array, LLM::Response::File, LLM::File] content
-    #  The content to format
+    # @param [JSON::Schema] schema
+    #  The schema to format
     # @return [Hash]
-    #  The formatted content
-    def format_content(content)
-      case content
-      when Array
-        content.map { format_content(_1) }
-      when LLM::Response::File
-        file = content
-        {file_data: {mime_type: file.mime_type, file_uri: file.uri}}
-      when LLM::File
-        file = content
-        {inline_data: {mime_type: file.mime_type, data: file.to_b64}}
-      when String
-        {text: content}
-      when LLM::Message
-        format_content(content.content)
-      else
-        raise LLM::Error::PromptError, "The given object (an instance of #{content.class}) " \
-                                       "is not supported by the Gemini API"
-      end
+    def format_schema(params)
+      return {} unless params and params[:schema]
+      schema = params.delete(:schema)
+      {generationConfig: {response_mime_type: "application/json", response_schema: schema}}
+    end
+
+    ##
+    # @param [Array<LLM::Function>] tools
+    #  The tools to format
+    # @return [Hash]
+    def format_tools(params)
+      return {} unless params and params[:tools]&.any?
+      functions = params.delete(:tools).grep(LLM::Function)
+      {tools: {functionDeclarations: functions.map { _1.format(self) }}}
     end
   end
 end

data/lib/llm/providers/gemini/images.rb

@@ -41,7 +41,7 @@ class LLM::Gemini
     #  Gemini implements image generation under the hood.
     # @return [LLM::Response::Image]
     def create(prompt:, model: "gemini-2.0-flash-exp-image-generation", **params)
-      req  = Net::HTTP::Post.new("/v1beta/models/#{model}:generateContent?key=#{secret}", headers)
+      req  = Net::HTTP::Post.new("/v1beta/models/#{model}:generateContent?key=#{key}", headers)
       body = JSON.dump({
         contents: [{parts: {text: prompt}}],
         generationConfig: {responseModalities: ["TEXT", "IMAGE"]}
@@ -58,16 +58,17 @@ class LLM::Gemini
     #   res = llm.images.edit image: LLM::File("cat.png"), prompt: "Add a hat to the cat"
     #   IO.copy_stream res.images[0], "hatoncat.png"
     # @see https://ai.google.dev/gemini-api/docs/image-generation Gemini docs
-    # @param [LLM::File] image The image to edit
+    # @param [String, LLM::File] image The image to edit
     # @param [String] prompt The prompt
     # @param [Hash] params Other parameters (see Gemini docs)
     # @raise (see LLM::Provider#request)
     # @note (see LLM::Gemini::Images#create)
     # @return [LLM::Response::Image]
     def edit(image:, prompt:, model: "gemini-2.0-flash-exp-image-generation", **params)
-      req  = Net::HTTP::Post.new("/v1beta/models/#{model}:generateContent?key=#{secret}", headers)
-      body = JSON.dump({
-        contents: [{parts: [{text: prompt}, format_content(image)]}],
+      req   = Net::HTTP::Post.new("/v1beta/models/#{model}:generateContent?key=#{key}", headers)
+      image = LLM.File(image)
+      body  = JSON.dump({
+        contents: [{parts: [{text: prompt}, format.format_content(image)]}],
         generationConfig: {responseModalities: ["TEXT", "IMAGE"]}
       }.merge!(params)).b
       set_body_stream(req, StringIO.new(body))
@@ -84,8 +85,12 @@ class LLM::Gemini
 
     private
 
-    def secret
-      @provider.instance_variable_get(:@secret)
+    def format
+      @format ||= CompletionFormat.new(nil)
+    end
+
+    def key
+      @provider.instance_variable_get(:@key)
     end
 
     def http

data/lib/llm/providers/gemini/models.rb

@@ -40,7 +40,7 @@ class LLM::Gemini
     # @raise (see LLM::Provider#request)
     # @return [LLM::Response::ModelList]
     def all(**params)
-      query = URI.encode_www_form(params.merge!(key: secret))
+      query = URI.encode_www_form(params.merge!(key: key))
       req = Net::HTTP::Get.new("/v1beta/models?#{query}", headers)
       res = request(http, req)
       LLM::Response::ModelList.new(res).tap { |modellist|
@@ -58,8 +58,8 @@ class LLM::Gemini
       @provider.instance_variable_get(:@http)
     end
 
-    def secret
-      @provider.instance_variable_get(:@secret)
+    def key
+      @provider.instance_variable_get(:@key)
     end
 
     [:headers, :request].each do |m|

data/lib/llm/providers/gemini/response_parser/completion_parser.rb

@@ -0,0 +1,46 @@
+# frozen_string_literal: true
+
+module LLM::Gemini::ResponseParser
+  class CompletionParser
+    def initialize(body)
+      @body = OpenStruct.from_hash(body)
+    end
+
+    def format(response)
+      {
+        model:,
+        prompt_tokens:,
+        completion_tokens:,
+        total_tokens:,
+        choices: format_choices(response)
+      }
+    end
+
+    private
+
+    def format_choices(response)
+      candidates.map.with_index do |choice, index|
+        content = choice.content
+        role, parts = content.role, content.parts
+        text  = parts.filter_map { _1["text"] }.join
+        tools = parts.filter_map { _1["functionCall"] }
+        extra = {index:, response:, tool_calls: format_tool_calls(tools), original_tool_calls: tools}
+        LLM::Message.new(role, text, extra)
+      end
+    end
+
+    def format_tool_calls(tools)
+      (tools || []).map do |tool|
+        function = {name: tool.name, arguments: tool.args}
+        OpenStruct.new(function)
+      end
+    end
+
+    def body = @body
+    def model = body.modelVersion
+    def prompt_tokens = body.usageMetadata.promptTokenCount
+    def completion_tokens = body.usageMetadata.candidatesTokenCount
+    def total_tokens = body.usageMetadata.totalTokenCount
+    def candidates = body.candidates
+  end
+end