llm.rb 0.10.1 → 0.11.0

data/lib/llm/providers/anthropic/response/completion.rb

@@ -0,0 +1,39 @@
+# frozen_string_literal: true
+
+module LLM::Anthropic::Response
+  module Completion
+    def choices = format_choices
+    def role = body.role
+    def model = body.model
+    def prompt_tokens = body.usage&.input_tokens || 0
+    def completion_tokens = body.usage&.output_tokens || 0
+    def total_tokens = prompt_tokens + completion_tokens
+
+    private
+
+    def format_choices
+      texts.map.with_index do |choice, index|
+        extra = {
+          index:, response: self,
+          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
+        }
+        LLM::Object.new(tool)
+      end
+    end
+
+    def parts = body.content
+    def texts = @texts ||= LLM::Object.from_hash(parts.select { _1["type"] == "text" })
+    def tools = @tools ||= LLM::Object.from_hash(parts.select { _1["type"] == "tool_use" })
+  end
+end

data/lib/llm/providers/anthropic.rb

@@ -5,10 +5,10 @@ module LLM
   # The Anthropic class implements a provider for
   # [Anthropic](https://www.anthropic.com)
   class Anthropic < Provider
+    require_relative "anthropic/response/completion"
     require_relative "anthropic/format"
     require_relative "anthropic/error_handler"
     require_relative "anthropic/stream_parser"
-    require_relative "anthropic/response_parser"
     require_relative "anthropic/models"
     include Format
 
@@ -20,23 +20,6 @@ module LLM
       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] 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, key:, model: "voyage-2", **params)
-      llm = LLM.voyageai(key:)
-      llm.embed(input, **params.merge(model:))
-    end
-
     ##
     # Provides an interface to the chat completions API
     # @see https://docs.anthropic.com/en/api/messages Anthropic docs
@@ -44,7 +27,7 @@ module LLM
     # @param params (see LLM::Provider#complete)
     # @example (see LLM::Provider#complete)
     # @raise (see LLM::Provider#request)
-    # @raise [LLM::Error::PromptError]
+    # @raise [LLM::PromptError]
     #  When given an object a provider does not understand
     # @return (see LLM::Provider#complete)
     def complete(prompt, params = {})
@@ -57,7 +40,7 @@ module LLM
       body = JSON.dump({messages: [format(messages)].flatten}.merge!(params))
       set_body_stream(req, StringIO.new(body))
       res = execute(request: req, stream:)
-      Response::Completion.new(res).extend(response_parser)
+      LLM::Response.new(res).extend(LLM::Anthropic::Response::Completion)
     end
 
     ##
@@ -92,10 +75,6 @@ module LLM
       )
     end
 
-    def response_parser
-      LLM::Anthropic::ResponseParser
-    end
-
     def stream_parser
       LLM::Anthropic::StreamParser
     end

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

@@ -12,7 +12,7 @@ module LLM::DeepSeek::Format
     end
 
     ##
-    # Formats the message for the OpenAI chat completions API
+    # Formats the message for the DeepSeek chat completions API
     # @return [Hash]
     def format
       catch(:abort) do
@@ -37,8 +37,8 @@ module LLM::DeepSeek::Format
       when LLM::Function::Return
         throw(:abort, {role: "tool", tool_call_id: content.id, content: JSON.dump(content.value)})
       else
-        raise LLM::Error::PromptError, "The given object (an instance of #{content.class}) " \
-                                       "is not supported by the DeepSeek chat completions API"
+        raise LLM::PromptError, "The given object (an instance of #{content.class}) " \
+                                "is not supported by the DeepSeek chat completions API"
       end
     end
 

data/lib/llm/providers/deepseek.rb

@@ -49,6 +49,12 @@ module LLM
       raise NotImplementedError
     end
 
+    ##
+    # @raise [NotImplementedError]
+    def vector_stores
+      raise NotImplementedError
+    end
+
     ##
     # Returns the default model for chat completions
     # @see https://api-docs.deepseek.com/quick_start/pricing deepseek-chat

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

@@ -34,20 +34,18 @@ class LLM::Gemini
     #   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 [String, LLM::File, LLM::Response::File] file The input audio
+    # @param [String, LLM::File, LLM::Response] file The input audio
     # @param [String] model The model to use
     # @param [Hash] params Other parameters (see Gemini docs)
     # @raise (see LLM::Provider#request)
-    # @return [LLM::Response::AudioTranscription]
+    # @return [LLM::Response]
     def create_transcription(file:, model: "gemini-1.5-flash", **params)
       res = @provider.complete [
         "Your task is to transcribe the contents of an audio file",
         "Your response should include the transcription, and nothing else",
         LLM.File(file)
       ], params.merge(role: :user, model:)
-      LLM::Response::AudioTranscription
-        .new(res)
-        .tap { _1.text = res.choices[0].content }
+      res.tap { _1.define_singleton_method(:text) { choices[0].content } }
     end
 
     ##
@@ -58,20 +56,18 @@ class LLM::Gemini
     #   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 [String, LLM::File, LLM::Response::File] file The input audio
+    # @param [String, LLM::File, LLM::Response] file The input audio
     # @param [String] model The model to use
     # @param [Hash] params Other parameters (see Gemini docs)
     # @raise (see LLM::Provider#request)
-    # @return [LLM::Response::AudioTranslation]
+    # @return [LLM::Response]
     def create_translation(file:, model: "gemini-1.5-flash", **params)
       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",
         LLM.File(file)
       ], params.merge(role: :user, model:)
-      LLM::Response::AudioTranslation
-        .new(res)
-        .tap { _1.text = res.choices[0].content }
+      res.tap { _1.define_singleton_method(:text) { choices[0].content } }
     end
   end
 end

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

@@ -35,6 +35,8 @@ class LLM::Gemini
   #   bot.chat(["Describe the audio file I sent to you", file])
   #   bot.messages.select(&:assistant?).each { print "[#{_1.role}]", _1.content, "\n" }
   class Files
+    require_relative "response/file"
+
     ##
     # Returns a new Files object
     # @param provider [LLM::Provider]
@@ -54,18 +56,12 @@ class LLM::Gemini
     # @see https://ai.google.dev/gemini-api/docs/files Gemini docs
     # @param [Hash] params Other parameters (see Gemini docs)
     # @raise (see LLM::Provider#request)
-    # @return [LLM::Response::FileList]
+    # @return [LLM::Response]
     def all(**params)
       query = URI.encode_www_form(params.merge!(key: key))
       req = Net::HTTP::Get.new("/v1beta/files?#{query}", headers)
       res = execute(request: req)
-      LLM::Response::FileList.new(res).tap { |filelist|
-        files = filelist.body["files"]&.map do |file|
-          file = file.transform_keys { snakecase(_1) }
-          LLM::Object.from_hash(file)
-        end || []
-        filelist.files = files
-      }
+      LLM::Response.new(res)
     end
 
     ##
@@ -77,7 +73,7 @@ class LLM::Gemini
     # @param [String, LLM::File] file The file
     # @param [Hash] params Other parameters (see Gemini docs)
     # @raise (see LLM::Provider#request)
-    # @return [LLM::Response::File]
+    # @return [LLM::Response]
     def create(file:, **params)
       file = LLM.File(file)
       req = Net::HTTP::Post.new(request_upload_url(file:), {})
@@ -87,7 +83,7 @@ class LLM::Gemini
       file.with_io do |io|
         set_body_stream(req, io)
         res = execute(request: req)
-        LLM::Response::File.new(res)
+        LLM::Response.new(res).extend(LLM::Gemini::Response::File)
       end
     end
 
@@ -101,13 +97,13 @@ class LLM::Gemini
     # @param [#name, String] file The file to get
     # @param [Hash] params Other parameters (see Gemini docs)
     # @raise (see LLM::Provider#request)
-    # @return [LLM::Response::File]
+    # @return [LLM::Response]
     def get(file:, **params)
       file_id = file.respond_to?(:name) ? file.name : file.to_s
       query = URI.encode_www_form(params.merge!(key: key))
       req = Net::HTTP::Get.new("/v1beta/#{file_id}?#{query}", headers)
       res = execute(request: req)
-      LLM::Response::File.new(res)
+      LLM::Response.new(res).extend(LLM::Gemini::Response::File)
     end
 
     ##
@@ -119,12 +115,13 @@ class LLM::Gemini
     # @param [#name, String] file The file to delete
     # @param [Hash] params Other parameters (see Gemini docs)
     # @raise (see LLM::Provider#request)
-    # @return [LLM::Response::File]
+    # @return [LLM::Response]
     def delete(file:, **params)
       file_id = file.respond_to?(:name) ? file.name : file.to_s
       query = URI.encode_www_form(params.merge!(key: key))
       req = Net::HTTP::Delete.new("/v1beta/#{file_id}?#{query}", headers)
-      execute(request: req)
+      res = execute(request: req)
+      LLM::Response.new(res)
     end
 
     ##

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

@@ -30,9 +30,11 @@ module LLM::Gemini::Format
       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::Response
+        format_response(content)
+      when File
+        content.close unless content.closed?
+        format_content(LLM.File(content.path))
       when LLM::File
         file = content
         [{inline_data: {mime_type: file.mime_type, data: file.to_b64}}]
@@ -43,11 +45,24 @@ module LLM::Gemini::Format
       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"
+        prompt_error!(content)
       end
     end
 
+    def format_response(response)
+      if response.file?
+        file = response
+        [{file_data: {mime_type: file.mime_type, file_uri: file.uri}}]
+      else
+        prompt_error!(content)
+      end
+    end
+
+    def prompt_error!(object)
+        raise LLM::PromptError, "The given object (an instance of #{object.class}) " \
+                                "is not supported by the Gemini API"
+    end
+
     def message = @message
     def content = message.content
   end

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

@@ -15,6 +15,7 @@ class LLM::Gemini
   #   res = llm.images.create prompt: "A dog on a rocket to the moon"
   #   IO.copy_stream res.images[0], "rocket.png"
   class Images
+    require_relative "response/image"
     include Format
 
     ##
@@ -39,16 +40,16 @@ class LLM::Gemini
     #  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]
+    # @return [LLM::Response]
     def create(prompt:, model: "gemini-2.0-flash-exp-image-generation", **params)
       req  = Net::HTTP::Post.new("/v1beta/models/#{model}:generateContent?key=#{key}", headers)
       body = JSON.dump({
-        contents: [{parts: [{text: create_prompt}, {text: prompt}]}],
+        contents: [{parts: [{text: system_prompt}, {text: prompt}]}],
         generationConfig: {responseModalities: ["TEXT", "IMAGE"]}
       }.merge!(params))
       req.body = body
       res = execute(request: req)
-      LLM::Response::Image.new(res).extend(response_parser)
+      LLM::Response.new(res).extend(LLM::Gemini::Response::Image)
     end
 
     ##
@@ -63,7 +64,7 @@ class LLM::Gemini
     # @param [Hash] params Other parameters (see Gemini docs)
     # @raise (see LLM::Provider#request)
     # @note (see LLM::Gemini::Images#create)
-    # @return [LLM::Response::Image]
+    # @return [LLM::Response]
     def edit(image:, prompt:, model: "gemini-2.0-flash-exp-image-generation", **params)
       req   = Net::HTTP::Post.new("/v1beta/models/#{model}:generateContent?key=#{key}", headers)
       image = LLM.File(image)
@@ -73,7 +74,7 @@ class LLM::Gemini
       }.merge!(params)).b
       set_body_stream(req, StringIO.new(body))
       res = execute(request: req)
-      LLM::Response::Image.new(res).extend(response_parser)
+      LLM::Response.new(res).extend(LLM::Gemini::Response::Image)
     end
 
     ##
@@ -93,7 +94,7 @@ class LLM::Gemini
       @provider.instance_variable_get(:@key)
     end
 
-    def create_prompt
+    def system_prompt
       <<~PROMPT
         Your task is to generate one or more image(s) from
         text I will provide to you. Your response *MUST* include
@@ -102,7 +103,7 @@ class LLM::Gemini
       PROMPT
     end
 
-    [:response_parser, :headers, :execute, :set_body_stream].each do |m|
+    [:headers, :execute, :set_body_stream].each do |m|
       define_method(m) { |*args, **kwargs, &b| @provider.send(m, *args, **kwargs, &b) }
     end
   end

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

@@ -38,18 +38,12 @@ class LLM::Gemini
     # @see https://ai.google.dev/api/models?hl=en#method:-models.list Gemini docs
     # @param [Hash] params Other parameters (see Gemini docs)
     # @raise (see LLM::Provider#request)
-    # @return [LLM::Response::ModelList]
+    # @return [LLM::Response]
     def all(**params)
       query = URI.encode_www_form(params.merge!(key: key))
       req = Net::HTTP::Get.new("/v1beta/models?#{query}", headers)
       res = execute(request: req)
-      LLM::Response::ModelList.new(res).tap { |modellist|
-        models = modellist.body["models"].map do |model|
-          model = model.transform_keys { snakecase(_1) }
-          LLM::Model.from_hash(model).tap { _1.provider = @provider }
-        end
-        modellist.models = models
-      }
+      LLM::Response.new(res)
     end
 
     private

data/lib/llm/providers/gemini/{response_parser/completion_parser.rb → response/completion.rb}

@@ -1,30 +1,23 @@
 # frozen_string_literal: true
 
-module LLM::Gemini::ResponseParser
-  class CompletionParser
-    def initialize(body)
-      @body = LLM::Object.from_hash(body)
-    end
-
-    def format(response)
-      {
-        model:,
-        prompt_tokens:,
-        completion_tokens:,
-        total_tokens:,
-        choices: format_choices(response)
-      }
-    end
+module LLM::Gemini::Response
+  module Completion
+    def model = body.modelVersion
+    def prompt_tokens = body.usageMetadata.promptTokenCount
+    def completion_tokens = body.usageMetadata.candidatesTokenCount
+    def total_tokens = body.usageMetadata.totalTokenCount
+    def choices = format_choices
 
     private
 
-    def format_choices(response)
+    def format_choices
       candidates.map.with_index do |choice, index|
+        choice = LLM::Object.from_hash(choice)
         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}
+        extra = {index:, response: self, tool_calls: format_tool_calls(tools), original_tool_calls: tools}
         LLM::Message.new(role, text, extra)
       end
     end
@@ -35,12 +28,5 @@ module LLM::Gemini::ResponseParser
         LLM::Object.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

data/lib/llm/providers/gemini/response/embedding.rb

@@ -0,0 +1,8 @@
+# frozen_string_literal: true
+
+module LLM::Gemini::Response
+  module Embedding
+    def model = "text-embedding-004"
+    def embeddings = body.dig("embedding", "values")
+  end
+end

data/lib/llm/providers/gemini/response/file.rb

@@ -0,0 +1,11 @@
+# frozen_string_literal: true
+
+module LLM::Gemini::Response
+  module File
+    def name = respond_to?(:file) ? file.name : body.name
+    def display_name = respond_to?(:file) ? file.displayName : body.displayName
+    def mime_type = respond_to?(:file) ? file.mimeType : body.mimeType
+    def uri = respond_to?(:file) ? file.uri : body.uri
+    def file? = true
+  end
+end

data/lib/llm/providers/gemini/response/image.rb

@@ -0,0 +1,26 @@
+# frozen_string_literal: true
+
+module LLM::Gemini::Response
+  module Image
+    ##
+    # @return [Array<StringIO>]
+    def images
+      candidates.flat_map do |candidate|
+        parts = candidate["content"]["parts"]
+        parts.filter_map do
+          data = _1.dig(:inlineData, :data)
+          next unless data
+          StringIO.new(data.unpack1("m0"))
+        end
+      end
+    end
+
+    ##
+    # Returns one or more image URLs, or an empty array
+    # @note
+    #  Gemini's image generation API does not return URLs, so this method
+    #  will always return an empty array.
+    # @return [Array<String>]
+    def urls = []
+  end
+end

data/lib/llm/providers/gemini.rb

@@ -29,10 +29,11 @@ module LLM
   #   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/response/embedding"
+    require_relative "gemini/response/completion"
     require_relative "gemini/error_handler"
     require_relative "gemini/format"
     require_relative "gemini/stream_parser"
-    require_relative "gemini/response_parser"
     require_relative "gemini/models"
     require_relative "gemini/images"
     require_relative "gemini/files"
@@ -61,7 +62,7 @@ module LLM
       req = Net::HTTP::Post.new(path, headers)
       req.body = JSON.dump({content: {parts: [{text: input}]}})
       res = execute(request: req)
-      Response::Embedding.new(res).extend(response_parser)
+      LLM::Response.new(res).extend(LLM::Gemini::Response::Embedding)
     end
 
     ##
@@ -71,7 +72,7 @@ module LLM
     # @param params (see LLM::Provider#complete)
     # @example (see LLM::Provider#complete)
     # @raise (see LLM::Provider#request)
-    # @raise [LLM::Error::PromptError]
+    # @raise [LLM::PromptError]
     #  When given an object a provider does not understand
     # @return (see LLM::Provider#complete)
     def complete(prompt, params = {})
@@ -86,7 +87,7 @@ module LLM
       body = JSON.dump({contents: format(messages)}.merge!(params))
       set_body_stream(req, StringIO.new(body))
       res = execute(request: req, stream:)
-      Response::Completion.new(res).extend(response_parser)
+      LLM::Response.new(res).extend(LLM::Gemini::Response::Completion)
     end
 
     ##
@@ -140,10 +141,6 @@ module LLM
       )
     end
 
-    def response_parser
-      LLM::Gemini::ResponseParser
-    end
-
     def stream_parser
       LLM::Gemini::StreamParser
     end

data/lib/llm/providers/llamacpp.rb

@@ -46,6 +46,12 @@ module LLM
       raise NotImplementedError
     end
 
+    ##
+    # @raise [NotImplementedError]
+    def vector_stores
+      raise NotImplementedError
+    end
+
     ##
     # Returns the default model for chat completions
     # @see https://ollama.com/library/qwen3 qwen3

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

@@ -28,13 +28,16 @@ module LLM::Ollama::Format
 
     def format_content(content)
       case content
+      when File
+        content.close unless content.closed?
+        format_content(LLM.File(content.path))
       when LLM::File
         if content.image?
           {content: "This message has an image associated with it", images: [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 " \
-                                         "Ollama API"
+          raise LLM::PromptError, "The given object (an instance of #{content.class}) " \
+                                  "is not an image, and therefore not supported by the " \
+                                  "Ollama API"
         end
       when String
         {content:}
@@ -43,8 +46,8 @@ module LLM::Ollama::Format
       when LLM::Function::Return
         throw(:abort, {role: "tool", tool_call_id: content.id, content: JSON.dump(content.value)})
       else
-        raise LLM::Error::PromptError, "The given object (an instance of #{content.class}) " \
-                                       "is not supported by the Ollama API"
+        raise LLM::PromptError, "The given object (an instance of #{content.class}) " \
+                                "is not supported by the Ollama API"
       end
     end
 

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

@@ -39,18 +39,12 @@ class LLM::Ollama
     # @see https://ollama.com/library Ollama library
     # @param [Hash] params Other parameters (see Ollama docs)
     # @raise (see LLM::Provider#request)
-    # @return [LLM::Response::ModelList]
+    # @return [LLM::Response]
     def all(**params)
       query = URI.encode_www_form(params)
       req = Net::HTTP::Get.new("/api/tags?#{query}", headers)
       res = execute(request: req)
-      LLM::Response::ModelList.new(res).tap { |modellist|
-        models = modellist.body["models"].map do |model|
-          model = model.transform_keys { snakecase(_1) }
-          LLM::Model.from_hash(model).tap { _1.provider = @provider }
-        end
-        modellist.models = models
-      }
+      LLM::Response.new(res)
     end
 
     private

data/lib/llm/providers/ollama/response/completion.rb

@@ -0,0 +1,28 @@
+# frozen_string_literal: true
+
+module LLM::Ollama::Response
+  module Completion
+    def model = body.model
+    def prompt_tokens = body.prompt_eval_count || 0
+    def completion_tokens = body.eval_count || 0
+    def total_tokens = prompt_tokens + completion_tokens
+    def message = body.message
+    def choices = [format_choices]
+
+    private
+
+    def format_choices
+      role, content, calls = message.to_h.values_at("role", "content", "tool_calls")
+      extra = {response: self, tool_calls: format_tool_calls(calls)}
+      LLM::Message.new(role, content, extra)
+    end
+
+    def format_tool_calls(tools)
+      return [] unless tools
+      tools.filter_map do |tool|
+        next unless tool["function"]
+        LLM::Object.new(tool["function"])
+      end
+    end
+  end
+end

data/lib/llm/providers/ollama/response/embedding.rb

@@ -0,0 +1,10 @@
+
+# frozen_string_literal: true
+
+module LLM::Ollama::Response
+  module Embedding
+    def embeddings = data.map { _1["embedding"] }
+    def prompt_tokens = body.dig("usage", "prompt_tokens") || 0
+    def total_tokens = body.dig("usage", "total_tokens") || 0
+  end
+end