llm.rb 0.2.1 → 0.3.0

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

@@ -0,0 +1,160 @@
+# frozen_string_literal: true
+
+class LLM::Gemini
+  ##
+  # The {LLM::Gemini::Files LLM::Gemini::Files} class provides a files
+  # object for interacting with [Gemini's Files API](https://ai.google.dev/gemini-api/docs/files).
+  # The files API allows a client to reference media files in prompts
+  # where they can be referenced by their URL.
+  #
+  # The files API is intended to preserve bandwidth and latency,
+  # especially for large files but it can be helpful for smaller files
+  # as well because it does not require the client to include a file
+  # in the prompt over and over again (which could be the case in a
+  # multi-turn conversation).
+  #
+  # @example
+  #   #!/usr/bin/env ruby
+  #   require "llm"
+  #
+  #   llm  = LLM.gemini(ENV["KEY"])
+  #   file = llm.files.create file: LLM::File("/audio/haiku.mp3")
+  #   bot = LLM::Chat.new(llm).lazy
+  #   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.")
+  #   bot.messages.select(&:assistant?).each { print "[#{_1.role}]", _1.content, "\n" }
+  # @example
+  #   #!/usr/bin/env ruby
+  #   require "llm"
+  #
+  #   llm  = LLM.gemini(ENV["KEY"])
+  #   file = llm.files.create file: LLM::File("/audio/haiku.mp3")
+  #   bot = LLM::Chat.new(llm).lazy
+  #   bot.chat(["Describe the audio file I sent to you", file])
+  #   bot.messages.select(&:assistant?).each { print "[#{_1.role}]", _1.content, "\n" }
+  class Files
+    ##
+    # Returns a new Files object
+    # @param provider [LLM::Provider]
+    # @return [LLM::Gemini::Files]
+    def initialize(provider)
+      @provider = provider
+    end
+
+    ##
+    # List all files
+    # @example
+    #   llm = LLM.gemini(ENV["KEY"])
+    #   res = llm.files.all
+    #   res.each do |file|
+    #     print "name: ", file.name, "\n"
+    #   end
+    # @see https://ai.google.dev/gemini-api/docs/files Gemini docs
+    # @param [Hash] params Other parameters (see Gemini docs)
+    # @raise (see LLM::HTTPClient#request)
+    # @return [LLM::Response::FileList]
+    def all(**params)
+      query = URI.encode_www_form(params.merge!(key: secret))
+      req = Net::HTTP::Get.new("/v1beta/files?#{query}", headers)
+      res = request(http, req)
+      LLM::Response::FileList.new(res).tap { |filelist|
+        files = filelist.body["files"]&.map do |file|
+          file = file.transform_keys { snakecase(_1) }
+          OpenStruct.from_hash(file)
+        end || []
+        filelist.files = files
+      }
+    end
+
+    ##
+    # Create a file
+    # @example
+    #   llm = LLM.gemini(ENV["KEY"])
+    #   res = llm.files.create file: LLM::File("/audio/haiku.mp3"),
+    # @see https://ai.google.dev/gemini-api/docs/files Gemini docs
+    # @param [File] file The file
+    # @param [Hash] params Other parameters (see Gemini docs)
+    # @raise (see LLM::HTTPClient#request)
+    # @return [LLM::Response::File]
+    def create(file:, **params)
+      req = Net::HTTP::Post.new(request_upload_url(file:), {})
+      req["content-length"] = file.bytesize
+      req["X-Goog-Upload-Offset"] = 0
+      req["X-Goog-Upload-Command"] = "upload, finalize"
+      req.body = File.binread(file.path)
+      res = request(http, req)
+      LLM::Response::File.new(res)
+    end
+
+    ##
+    # Get a file
+    # @example
+    #   llm = LLM.gemini(ENV["KEY"])
+    #   res = llm.files.get(file: "files/1234567890")
+    #   print "name: ", res.name, "\n"
+    # @see https://ai.google.dev/gemini-api/docs/files Gemini docs
+    # @param [#name, String] file The file to get
+    # @param [Hash] params Other parameters (see Gemini docs)
+    # @raise (see LLM::HTTPClient#request)
+    # @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))
+      req = Net::HTTP::Get.new("/v1beta/#{file_id}?#{query}", headers)
+      res = request(http, req)
+      LLM::Response::File.new(res)
+    end
+
+    ##
+    # Delete a file
+    # @example
+    #   llm = LLM.gemini(ENV["KEY"])
+    #   res = llm.files.delete(file: "files/1234567890")
+    # @see https://ai.google.dev/gemini-api/docs/files Gemini docs
+    # @param [#name, String] file The file to delete
+    # @param [Hash] params Other parameters (see Gemini docs)
+    # @raise (see LLM::HTTPClient#request)
+    # @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))
+      req = Net::HTTP::Delete.new("/v1beta/#{file_id}?#{query}", headers)
+      request(http, req)
+    end
+
+    ##
+    # @raise [NotImplementedError]
+    #  This method is not implemented by Gemini
+    def download
+      raise NotImplementedError
+    end
+
+    private
+
+    include LLM::Utils
+
+    def request_upload_url(file:)
+      req = Net::HTTP::Post.new("/upload/v1beta/files?key=#{secret}", headers)
+      req["X-Goog-Upload-Protocol"] = "resumable"
+      req["X-Goog-Upload-Command"] = "start"
+      req["X-Goog-Upload-Header-Content-Length"] = file.bytesize
+      req["X-Goog-Upload-Header-Content-Type"] = file.mime_type
+      req.body = JSON.dump(file: {display_name: File.basename(file.path)})
+      res = request(http, req)
+      res["x-goog-upload-url"]
+    end
+
+    def http
+      @provider.instance_variable_get(:@http)
+    end
+
+    def secret
+      @provider.instance_variable_get(:@secret)
+    end
+
+    [:headers, :request].each do |m|
+      define_method(m) { |*args, &b| @provider.send(m, *args, &b) }
+    end
+  end
+end

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

@@ -1,6 +1,8 @@
 # frozen_string_literal: true
 
 class LLM::Gemini
+  ##
+  # @private
   module Format
     ##
     # @param [Array<LLM::Message>] messages
@@ -19,16 +21,20 @@ class LLM::Gemini
     private
 
     ##
-    # @param [String, LLM::File] content
+    # @param [String, Array, LLM::Response::File, LLM::File] content
     #  The content to format
-    # @return [String, Hash]
+    # @return [Hash]
     #  The formatted content
     def format_content(content)
-      if LLM::File === content
+      case content
+      when Array
+        content.map { format_content(_1) }
+      when LLM::Response::File
         file = content
-        {
-          inline_data: {mime_type: file.mime_type, data: [File.binread(file.path)].pack("m0")}
-        }
+        {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}}
       else
         {text: content}
       end

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

@@ -0,0 +1,99 @@
+# frozen_string_literal: true
+
+class LLM::Gemini
+  ##
+  # The {LLM::Gemini::Images LLM::Gemini::Images} class provides an images
+  # object for interacting with [Gemini's images API](https://ai.google.dev/gemini-api/docs/image-generation).
+  # Please note that unlike OpenAI, which can return either URLs or base64-encoded strings,
+  # Gemini's images API will always return an image as a base64 encoded string that
+  # can be decoded into binary.
+  # @example
+  #   #!/usr/bin/env ruby
+  #   require "llm"
+  #
+  #   llm = LLM.gemini(ENV["KEY"])
+  #   res = llm.images.create prompt: "A dog on a rocket to the moon"
+  #   File.binwrite "rocket.png", res.images[0].binary
+  class Images
+    include Format
+
+    ##
+    # Returns a new Images object
+    # @param provider [LLM::Provider]
+    # @return [LLM::Gemini::Responses]
+    def initialize(provider)
+      @provider = provider
+    end
+
+    ##
+    # Create an image
+    # @example
+    #   llm = LLM.gemini(ENV["KEY"])
+    #   res = llm.images.create prompt: "A dog on a rocket to the moon"
+    #   File.binwrite "rocket.png", res.images[0].binary
+    # @see https://ai.google.dev/gemini-api/docs/image-generation Gemini docs
+    # @param [String] prompt The prompt
+    # @param [Hash] params Other parameters (see Gemini docs)
+    # @raise (see LLM::HTTPClient#request)
+    # @note
+    #  The prompt should make it clear you want to generate an image, or you
+    #  might unexpectedly receive a purely textual response. This is due to how
+    #  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.body = JSON.dump({
+        contents: [{parts: {text: prompt}}],
+        generationConfig: {responseModalities: ["TEXT", "IMAGE"]}
+      }.merge!(params))
+      res = request(http, req)
+      LLM::Response::Image.new(res).extend(response_parser)
+    end
+
+    ##
+    # Edit an image
+    # @example
+    #   llm = LLM.gemini(ENV["KEY"])
+    #   res = llm.images.edit image: LLM::File("cat.png"), prompt: "Add a hat to the cat"
+    #   File.binwrite "hatoncat.png", res.images[0].binary
+    # @see https://ai.google.dev/gemini-api/docs/image-generation Gemini docs
+    # @param [LLM::File] image The image to edit
+    # @param [String] prompt The prompt
+    # @param [Hash] params Other parameters (see Gemini docs)
+    # @raise (see LLM::HTTPClient#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)
+      req.body = JSON.dump({
+        contents: [
+          {parts: [{text: prompt}, format_content(image)]}
+        ],
+        generationConfig: {responseModalities: ["TEXT", "IMAGE"]}
+      }.merge!(params))
+      res = request(http, req)
+      LLM::Response::Image.new(res).extend(response_parser)
+    end
+
+    ##
+    # @raise [NotImplementedError]
+    #  This method is not implemented by Gemini
+    def create_variation
+      raise NotImplementedError
+    end
+
+    private
+
+    def secret
+      @provider.instance_variable_get(:@secret)
+    end
+
+    def http
+      @provider.instance_variable_get(:@http)
+    end
+
+    [:response_parser, :headers, :request].each do |m|
+      define_method(m) { |*args, &b| @provider.send(m, *args, &b) }
+    end
+  end
+end

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

@@ -1,7 +1,13 @@
 # frozen_string_literal: true
 
 class LLM::Gemini
+  ##
+  # @private
   module ResponseParser
+    ##
+    # @param [Hash] body
+    #  The response body from the LLM provider
+    # @return [Hash]
     def parse_embedding(body)
       {
         model: "text-embedding-004",
@@ -20,12 +26,32 @@ class LLM::Gemini
           LLM::Message.new(
             _1.dig("content", "role"),
             _1.dig("content", "parts", 0, "text"),
-            {completion: self}
+            {response: self}
           )
         end,
         prompt_tokens: body.dig("usageMetadata", "promptTokenCount"),
         completion_tokens: body.dig("usageMetadata", "candidatesTokenCount")
       }
     end
+
+    ##
+    # @param [Hash] body
+    #  The response body from the LLM provider
+    # @return [Hash]
+    def parse_image(body)
+      {
+        urls: [],
+        images: body["candidates"].flat_map do |candidate|
+          candidate["content"]["parts"].filter_map do
+            next unless _1.dig("inlineData", "data")
+            OpenStruct.from_hash(
+              mime_type: _1["inlineData"]["mimeType"],
+              encoded: _1["inlineData"]["data"],
+              binary: _1["inlineData"]["data"].unpack1("m0")
+            )
+          end
+        end
+      }
+    end
   end
 end

data/lib/llm/providers/gemini.rb

@@ -3,11 +3,37 @@
 module LLM
   ##
   # The Gemini class implements a provider for
-  # [Gemini](https://ai.google.dev/)
+  # [Gemini](https://ai.google.dev/).
+  #
+  # The Gemini 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 that are over 20MB
+  #
+  # @example
+  #   #!/usr/bin/env ruby
+  #   require "llm"
+  #
+  #   llm = LLM.gemini(ENV["KEY"])
+  #   bot = LLM::Chat.new(llm).lazy
+  #   bot.chat LLM::File("/images/capybara.png")
+  #   bot.chat "Describe the image"
+  #   bot.messages.select(&:assistant?).each { print "[#{_1.role}]", _1.content, "\n" }
+  # @example
+  #   #!/usr/bin/env ruby
+  #   require "llm"
+  #
+  #   llm = LLM.gemini(ENV["KEY"])
+  #   bot = LLM::Chat.new(llm).lazy
+  #   bot.chat ["Describe the image", LLM::File("/images/capybara.png")]
+  #   bot.messages.select(&:assistant?).each { print "[#{_1.role}]", _1.content, "\n" }
   class Gemini < Provider
     require_relative "gemini/error_handler"
     require_relative "gemini/response_parser"
     require_relative "gemini/format"
+    require_relative "gemini/images"
+    require_relative "gemini/files"
+    require_relative "gemini/audio"
     include Format
 
     HOST = "generativelanguage.googleapis.com"
@@ -19,10 +45,14 @@ module LLM
     end
 
     ##
+    # Provides an embedding
     # @param input (see LLM::Provider#embed)
+    # @param model (see LLM::Provider#embed)
+    # @param params (see LLM::Provider#embed)
+    # @raise (see LLM::HTTPClient#request)
     # @return (see LLM::Provider#embed)
-    def embed(input, **params)
-      path = ["/v1beta/models/text-embedding-004", "embedContent?key=#{@secret}"].join(":")
+    def embed(input, model: "text-embedding-004", **params)
+      path = ["/v1beta/models/#{model}", "embedContent?key=#{@secret}"].join(":")
       req = Net::HTTP::Post.new(path, headers)
       req.body = JSON.dump({content: {parts: [{text: input}]}})
       res = request(@http, req)
@@ -30,13 +60,17 @@ module LLM
     end
 
     ##
+    # Provides an interface to the chat completions API
     # @see https://ai.google.dev/api/generate-content#v1beta.models.generateContent Gemini docs
     # @param prompt (see LLM::Provider#complete)
     # @param role (see LLM::Provider#complete)
+    # @param model (see LLM::Provider#complete)
+    # @param params (see LLM::Provider#complete)
+    # @example (see LLM::Provider#complete)
+    # @raise (see LLM::HTTPClient#request)
     # @return (see LLM::Provider#complete)
-    def complete(prompt, role = :user, **params)
-      params   = {model: "gemini-1.5-flash"}.merge!(params)
-      path     = ["/v1beta/models/#{params.delete(:model)}", "generateContent?key=#{@secret}"].join(":")
+    def complete(prompt, role = :user, model: "gemini-1.5-flash", **params)
+      path     = ["/v1beta/models/#{model}", "generateContent?key=#{@secret}"].join(":")
       req      = Net::HTTP::Post.new(path, headers)
       messages = [*(params.delete(:messages) || []), LLM::Message.new(role, prompt)]
       req.body = JSON.dump({contents: format(messages)})
@@ -44,6 +78,28 @@ module LLM
       Response::Completion.new(res).extend(response_parser)
     end
 
+    ##
+    # Provides an interface to Gemini's audio API
+    # @see https://ai.google.dev/gemini-api/docs/audio Gemini docs
+    def audio
+      LLM::Gemini::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]
+    def images
+      LLM::Gemini::Images.new(self)
+    end
+
+    ##
+    # Provides an interface to Gemini's file management API
+    # @see https://ai.google.dev/gemini-api/docs/files Gemini docs
+    def files
+      LLM::Gemini::Files.new(self)
+    end
+
     ##
     # @return (see LLM::Provider#assistant_role)
     def assistant_role

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

@@ -1,6 +1,8 @@
 # frozen_string_literal: true
 
 class LLM::Ollama
+  ##
+  # @private
   class ErrorHandler
     ##
     # @return [Net::HTTPResponse]

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

@@ -1,6 +1,8 @@
 # frozen_string_literal: true
 
 class LLM::Ollama
+  ##
+  # @private
   module Format
     ##
     # @param [Array<LLM::Message>] messages
@@ -9,9 +11,11 @@ class LLM::Ollama
     def format(messages)
       messages.map do
         if Hash === _1
-          {role: _1[:role], content: format_content(_1[:content])}
+          {role: _1[:role]}
+            .merge!(_1)
+            .merge!(format_content(_1[:content]))
         else
-          {role: _1.role, content: format_content(_1.content)}
+          {role: _1.role}.merge! format_content(_1.content)
         end
       end
     end
@@ -24,10 +28,14 @@ class LLM::Ollama
     # @return [String, Hash]
     #  The formatted content
     def format_content(content)
-      if URI === content
-        [{type: :image_url, image_url: {url: content.to_s}}]
+      if LLM::File === content
+        if content.image?
+          {content: "This message has an image associated with it", images: [content.to_b64]}
+        else
+          raise TypeError, "'#{content.path}' was not recognized as an image file."
+        end
       else
-        content
+        {content:}
       end
     end
   end

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

@@ -1,6 +1,8 @@
 # frozen_string_literal: true
 
 class LLM::Ollama
+  ##
+  # @private
   module ResponseParser
     ##
     # @param [Hash] body
@@ -22,7 +24,7 @@ class LLM::Ollama
     def parse_completion(body)
       {
         model: body["model"],
-        choices: [LLM::Message.new(*body["message"].values_at("role", "content"), {completion: self})],
+        choices: [LLM::Message.new(*body["message"].values_at("role", "content"), {response: self})],
         prompt_tokens: body.dig("prompt_eval_count"),
         completion_tokens: body.dig("eval_count")
       }

data/lib/llm/providers/ollama.rb

@@ -2,8 +2,22 @@
 
 module LLM
   ##
-  # The Ollama class implements a provider for
-  # [Ollama](https://ollama.ai/)
+  # The Ollama class implements a provider for [Ollama](https://ollama.ai/).
+  #
+  # This provider supports a wide range of models, it is relatively
+  # straight forward to run on your own hardware, and includes multi-modal
+  # models that can process images and text. See the example for a demonstration
+  # of a multi-modal model by the name `llava`
+  #
+  # @example
+  #   #!/usr/bin/env ruby
+  #   require "llm"
+  #
+  #   llm = LLM.ollama(nil)
+  #   bot = LLM::Chat.new(llm, model: "llava").lazy
+  #   bot.chat LLM::File("/images/capybara.png")
+  #   bot.chat "Describe the image"
+  #   bot.messages.select(&:assistant?).each { print "[#{_1.role}]", _1.content, "\n" }
   class Ollama < Provider
     require_relative "ollama/error_handler"
     require_relative "ollama/response_parser"
@@ -19,10 +33,14 @@ module LLM
     end
 
     ##
+    # Provides an embedding
     # @param input (see LLM::Provider#embed)
+    # @param model (see LLM::Provider#embed)
+    # @param params (see LLM::Provider#embed)
+    # @raise (see LLM::HTTPClient#request)
     # @return (see LLM::Provider#embed)
-    def embed(input, **params)
-      params   = {model: "llama3.2"}.merge!(params)
+    def embed(input, model: "llama3.2", **params)
+      params   = {model:}.merge!(params)
       req      = Net::HTTP::Post.new("/v1/embeddings", headers)
       req.body = JSON.dump({input:}.merge!(params))
       res      = request(@http, req)
@@ -30,15 +48,20 @@ module LLM
     end
 
     ##
+    # Provides an interface to the chat completions API
     # @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)
+    # @param model (see LLM::Provider#complete)
+    # @param params (see LLM::Provider#complete)
+    # @example (see LLM::Provider#complete)
+    # @raise (see LLM::HTTPClient#request)
     # @return (see LLM::Provider#complete)
-    def complete(prompt, role = :user, **params)
-      params   = {model: "llama3.2", stream: false}.merge!(params)
+    def complete(prompt, role = :user, model: "llama3.2", **params)
+      params   = {model:, stream: false}.merge!(params)
       req      = Net::HTTP::Post.new("/api/chat", headers)
       messages = [*(params.delete(:messages) || []), LLM::Message.new(role, prompt)]
-      req.body = JSON.dump({messages: messages.map(&:to_h)}.merge!(params))
+      req.body = JSON.dump({messages: format(messages)}.merge!(params))
       res      = request(@http, req)
       Response::Completion.new(res).extend(response_parser)
     end

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

@@ -0,0 +1,97 @@
+# frozen_string_literal: true
+
+class LLM::OpenAI
+  ##
+  # The {LLM::OpenAI::Audio LLM::OpenAI::Audio} class provides an audio
+  # object for interacting with [OpenAI's audio API](https://platform.openai.com/docs/api-reference/audio/createSpeech).
+  # @example
+  #   llm = LLM.openai(ENV["KEY"])
+  #   res = llm.audio.create_speech(input: "A dog on a rocket to the moon")
+  #   File.binwrite("rocket.mp3", res.audio.string)
+  class Audio
+    require "stringio"
+
+    ##
+    # Returns a new Audio object
+    # @param provider [LLM::Provider]
+    # @return [LLM::OpenAI::Responses]
+    def initialize(provider)
+      @provider = provider
+    end
+
+    ##
+    # Create an audio track
+    # @example
+    #   llm = LLM.openai(ENV["KEY"])
+    #   res = llm.images.create_speech(input: "A dog on a rocket to the moon")
+    #   File.binwrite("rocket.mp3", res.audio.string)
+    # @see https://platform.openai.com/docs/api-reference/audio/createSpeech OpenAI docs
+    # @param [String] input The text input
+    # @param [String] voice The voice to use
+    # @param [String] model The model to use
+    # @param [String] response_format The response format
+    # @param [Hash] params Other parameters (see OpenAI docs)
+    # @raise (see LLM::HTTPClient#request)
+    # @return [LLM::Response::Audio]
+    def create_speech(input:, voice: "alloy", model: "gpt-4o-mini-tts", response_format: "mp3", **params)
+      req = Net::HTTP::Post.new("/v1/audio/speech", headers)
+      req.body = JSON.dump({input:, voice:, model:, response_format:}.merge!(params))
+      io = StringIO.new("".b)
+      res = request(http, req) { _1.read_body { |chunk| io << chunk } }
+      LLM::Response::Audio.new(res).tap { _1.audio = io }
+    end
+
+    ##
+    # Create an audio transcription
+    # @example
+    #   llm = LLM.openai(ENV["KEY"])
+    #   res = llm.audio.create_transcription(file: LLM::File("/rocket.mp3"))
+    #   res.text # => "A dog on a rocket to the moon"
+    # @see https://platform.openai.com/docs/api-reference/audio/createTranscription OpenAI docs
+    # @param [LLM::File] file The input audio
+    # @param [String] model The model to use
+    # @param [Hash] params Other parameters (see OpenAI docs)
+    # @raise (see LLM::HTTPClient#request)
+    # @return [LLM::Response::AudioTranscription]
+    def create_transcription(file:, model: "whisper-1", **params)
+      multi = LLM::Multipart.new(params.merge!(file:, model:))
+      req = Net::HTTP::Post.new("/v1/audio/transcriptions", headers)
+      req["content-type"] = multi.content_type
+      req.body = multi.body
+      res = request(http, req)
+      LLM::Response::AudioTranscription.new(res).tap { _1.text = _1.body["text"] }
+    end
+
+    ##
+    # Create an audio translation (in English)
+    # @example
+    #   # Arabic => English
+    #   llm = LLM.openai(ENV["KEY"])
+    #   res = llm.audio.create_translation(file: LLM::File("/bismillah.mp3"))
+    #   res.text # => "In the name of Allah, the Beneficent, the Merciful."
+    # @see https://platform.openai.com/docs/api-reference/audio/createTranslation OpenAI docs
+    # @param [LLM::File] file The input audio
+    # @param [String] model The model to use
+    # @param [Hash] params Other parameters (see OpenAI docs)
+    # @raise (see LLM::HTTPClient#request)
+    # @return [LLM::Response::AudioTranslation]
+    def create_translation(file:, model: "whisper-1", **params)
+      multi = LLM::Multipart.new(params.merge!(file:, model:))
+      req = Net::HTTP::Post.new("/v1/audio/translations", headers)
+      req["content-type"] = multi.content_type
+      req.body = multi.body
+      res = request(http, req)
+      LLM::Response::AudioTranslation.new(res).tap { _1.text = _1.body["text"] }
+    end
+
+    private
+
+    def http
+      @provider.instance_variable_get(:@http)
+    end
+
+    [:headers, :request].each do |m|
+      define_method(m) { |*args, &b| @provider.send(m, *args, &b) }
+    end
+  end
+end