llm.rb 4.7.0 → 4.9.0

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

@@ -1,6 +1,6 @@
 # frozen_string_literal: true
 
-module LLM::Gemini::ResponseAdapter
+module LLM::Google::ResponseAdapter
   module Completion
     ##
     # (see LLM::Contract::Completion#messages)
@@ -64,17 +64,17 @@ module LLM::Gemini::ResponseAdapter
         content = choice.content || LLM::Object.new
         role = content.role || "model"
         parts = content.parts || [{"text" => choice.finishReason}]
-        text  = parts.filter_map { _1["text"] }.join
-        tools = parts.filter_map { _1["functionCall"] }
+        text = parts.filter_map { _1["text"] }.join
+        tools = parts.select { _1["functionCall"] }
         extra = {index:, response: self, tool_calls: adapt_tool_calls(tools), original_tool_calls: tools}
         LLM::Message.new(role, text, extra)
       end
     end
 
-    def adapt_tool_calls(tools)
-      (tools || []).map do |tool|
-        function = {name: tool.name, arguments: tool.args}
-        function
+    def adapt_tool_calls(parts)
+      (parts || []).map do |part|
+        tool = part["functionCall"]
+        {name: tool.name, arguments: tool.args}
       end
     end
 

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

@@ -1,6 +1,6 @@
 # frozen_string_literal: true
 
-module LLM::Gemini::ResponseAdapter
+module LLM::Google::ResponseAdapter
   module Embedding
     def model = "text-embedding-004"
     def embeddings = body.dig("embedding", "values")

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

@@ -1,6 +1,6 @@
 # frozen_string_literal: true
 
-module LLM::Gemini::ResponseAdapter
+module LLM::Google::ResponseAdapter
   module File
     def name = respond_to?(:file) ? file.name : body.name
     def display_name = respond_to?(:file) ? file.displayName : body.displayName

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

@@ -1,6 +1,6 @@
 # frozen_string_literal: true
 
-module LLM::Gemini::ResponseAdapter
+module LLM::Google::ResponseAdapter
   module Files
     include ::Enumerable
     def each(&)

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

@@ -1,6 +1,6 @@
 # frozen_string_literal: true
 
-module LLM::Gemini::ResponseAdapter
+module LLM::Google::ResponseAdapter
   module Image
     ##
     # @return [Array<StringIO>]

data/lib/llm/providers/google/response_adapter/models.rb

@@ -0,0 +1,13 @@
+# frozen_string_literal: true
+
+module LLM::Google::ResponseAdapter
+  module Models
+    include LLM::Model::Collection
+
+    private
+
+    def raw_models
+      body.models || []
+    end
+  end
+end

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

@@ -1,8 +1,8 @@
 # frozen_string_literal: true
 
-module LLM::Gemini::ResponseAdapter
+module LLM::Google::ResponseAdapter
   ##
-  # The {LLM::Gemini::ResponseAdapter::WebSearch LLM::Gemini::ResponseAdapter::WebSearch}
+  # The {LLM::Google::ResponseAdapter::WebSearch LLM::Google::ResponseAdapter::WebSearch}
   # module provides methods for accessing web search results from a web search
   # tool call made via the {LLM::Provider#web_search LLM::Provider#web_search}
   # method.

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

@@ -1,6 +1,6 @@
 # frozen_string_literal: true
 
-class LLM::Gemini
+class LLM::Google
   ##
   # @private
   module ResponseAdapter
@@ -27,13 +27,13 @@ class LLM::Gemini
     # @api private
     def select(type)
       case type
-      when :completion then LLM::Gemini::ResponseAdapter::Completion
-      when :embedding then LLM::Gemini::ResponseAdapter::Embedding
-      when :file then LLM::Gemini::ResponseAdapter::File
-      when :files then LLM::Gemini::ResponseAdapter::Files
-      when :image then LLM::Gemini::ResponseAdapter::Image
-      when :models then LLM::Gemini::ResponseAdapter::Models
-      when :web_search then LLM::Gemini::ResponseAdapter::WebSearch
+      when :completion then LLM::Google::ResponseAdapter::Completion
+      when :embedding then LLM::Google::ResponseAdapter::Embedding
+      when :file then LLM::Google::ResponseAdapter::File
+      when :files then LLM::Google::ResponseAdapter::Files
+      when :image then LLM::Google::ResponseAdapter::Image
+      when :models then LLM::Google::ResponseAdapter::Models
+      when :web_search then LLM::Google::ResponseAdapter::WebSearch
       else
         raise ArgumentError, "Unknown response adapter type: #{type.inspect}"
       end

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

@@ -1,6 +1,6 @@
 # frozen_string_literal: true
 
-class LLM::Gemini
+class LLM::Google
   ##
   # @private
   class StreamParser
@@ -11,7 +11,7 @@ class LLM::Gemini
 
     ##
     # @param [#<<] io An IO-like object
-    # @return [LLM::Gemini::StreamParser]
+    # @return [LLM::Google::StreamParser]
     def initialize(io)
       @body = {"candidates" => []}
       @io = io
@@ -19,7 +19,7 @@ class LLM::Gemini
 
     ##
     # @param [Hash] chunk
-    # @return [LLM::Gemini::StreamParser]
+    # @return [LLM::Google::StreamParser]
     def parse!(chunk)
       tap { merge_chunk!(chunk) }
     end

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/error_handler.rb

@@ -35,7 +35,7 @@ class LLM::Ollama
       ex = error
       @tracer.on_request_error(ex:, span:)
     ensure
-      raise(ex)
+      raise(ex) if ex
     end
 
     private

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

@@ -44,7 +44,7 @@ class LLM::Ollama
       query = URI.encode_www_form(params)
       req = Net::HTTP::Get.new("/api/tags?#{query}", headers)
       res, span, tracer = execute(request: req, operation: "request")
-      res = LLM::Response.new(res)
+      res = ResponseAdapter.adapt(res, type: :models)
       tracer.on_request_finish(operation: "request", res:, span:)
       res
     end

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

@@ -0,0 +1,13 @@
+# frozen_string_literal: true
+
+module LLM::Ollama::ResponseAdapter
+  module Models
+    include LLM::Model::Collection
+
+    private
+
+    def raw_models
+      body.models || []
+    end
+  end
+end

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

@@ -6,6 +6,7 @@ class LLM::Ollama
   module ResponseAdapter
     require_relative "response_adapter/completion"
     require_relative "response_adapter/embedding"
+    require_relative "response_adapter/models"
 
     module_function
 
@@ -24,6 +25,7 @@ class LLM::Ollama
       case type
       when :completion then LLM::Ollama::ResponseAdapter::Completion
       when :embedding then LLM::Ollama::ResponseAdapter::Embedding
+      when :models then LLM::Ollama::ResponseAdapter::Models
       else
         raise ArgumentError, "Unknown response adapter type: #{type.inspect}"
       end

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/error_handler.rb

@@ -35,15 +35,15 @@ class LLM::OpenAI
       ex = error
       @tracer.on_request_error(ex:, span:)
     ensure
-      raise(ex)
+      raise(ex) if ex
     end
 
     private
 
     ##
-    # @return [LLM::Object]
+    # @return [String, LLM::Object]
     def body
-      @body ||= LLM.json.load(res.body)
+      @body ||= parse_body!
     end
 
     ##
@@ -79,5 +79,20 @@ class LLM::OpenAI
         LLM::InvalidRequestError.new(error["message"]).tap { _1.response = res }
       end
     end
+
+    ##
+    # Tries to parse the response body as a LLM::Object
+    # @return [String, LLM::Object]
+    def parse_body!
+      if String === res.body
+        LLM::Object.from LLM.json.load(res.body)
+      elsif Hash === res.body
+        LLM::Object.from(res.body)
+      else
+        res.body
+      end
+    rescue
+      res.body
+    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/images.rb

@@ -5,7 +5,7 @@ class LLM::OpenAI
   # The {LLM::OpenAI::Images LLM::OpenAI::Images} class provides an interface
   # for [OpenAI's images API](https://platform.openai.com/docs/api-reference/images).
   # OpenAI supports multiple response formats: temporary URLs, or binary strings
-  # encoded in base64. The default is to return temporary URLs.
+  # encoded in base64. The default is to return base64-encoded image data.
   #
   # @example Temporary URLs
   #   #!/usr/bin/env ruby
@@ -14,7 +14,8 @@ class LLM::OpenAI
   #   require "fileutils"
   #
   #   llm = LLM.openai(key: ENV["KEY"])
-  #   res = llm.images.create prompt: "A dog on a rocket to the moon"
+  #   res = llm.images.create prompt: "A dog on a rocket to the moon",
+  #                           response_format: "url"
   #   FileUtils.mv OpenURI.open_uri(res.urls[0]).path,
   #                "rocket.png"
   #
@@ -40,16 +41,17 @@ class LLM::OpenAI
     # @example
     #   llm = LLM.openai(key: ENV["KEY"])
     #   res = llm.images.create prompt: "A dog on a rocket to the moon"
-    #   res.urls.each { print _1, "\n" }
+    #   IO.copy_stream res.images[0], "rocket.png"
     # @see https://platform.openai.com/docs/api-reference/images/create OpenAI docs
     # @param [String] prompt The prompt
     # @param [String] model The model to use
+    # @param [String] response_format The response format ("b64_json" or "url")
     # @param [Hash] params Other parameters (see OpenAI docs)
     # @raise (see LLM::Provider#request)
     # @return [LLM::Response]
-    def create(prompt:, model: "dall-e-3", **params)
+    def create(prompt:, model: "dall-e-3", response_format: "b64_json", **params)
       req = Net::HTTP::Post.new("/v1/images/generations", headers)
-      req.body = LLM.json.dump({prompt:, n: 1, model:}.merge!(params))
+      req.body = LLM.json.dump({prompt:, n: 1, model:, response_format:}.merge!(params))
       res, span, tracer = execute(request: req, operation: "request")
       res = ResponseAdapter.adapt(res, type: :image)
       tracer.on_request_finish(operation: "request", model:, res:, span:)
@@ -61,16 +63,19 @@ class LLM::OpenAI
     # @example
     #   llm = LLM.openai(key: ENV["KEY"])
     #   res = llm.images.create_variation(image: "/images/hat.png", n: 5)
-    #   p res.urls
+    #   res.images.each.with_index do |image, index|
+    #     IO.copy_stream image, "variation#{index}.png"
+    #   end
     # @see https://platform.openai.com/docs/api-reference/images/createVariation OpenAI docs
     # @param [File] image The image to create variations from
     # @param [String] model The model to use
+    # @param [String] response_format The response format ("b64_json" or "url")
     # @param [Hash] params Other parameters (see OpenAI docs)
     # @raise (see LLM::Provider#request)
     # @return [LLM::Response]
-    def create_variation(image:, model: "dall-e-2", **params)
+    def create_variation(image:, model: "dall-e-2", response_format: "b64_json", **params)
       image = LLM.File(image)
-      multi = LLM::Multipart.new(params.merge!(image:, model:))
+      multi = LLM::Multipart.new(params.merge!(image:, model:, response_format:))
       req = Net::HTTP::Post.new("/v1/images/variations", headers)
       req["content-type"] = multi.content_type
       set_body_stream(req, multi.body)
@@ -85,17 +90,18 @@ class LLM::OpenAI
     # @example
     #   llm = LLM.openai(key: ENV["KEY"])
     #   res = llm.images.edit(image: "/images/hat.png", prompt: "A cat wearing this hat")
-    #   p res.urls
+    #   IO.copy_stream res.images[0], "hatoncat.png"
     # @see https://platform.openai.com/docs/api-reference/images/createEdit OpenAI docs
     # @param [File] image The image to edit
     # @param [String] prompt The prompt
     # @param [String] model The model to use
+    # @param [String] response_format The response format ("b64_json" or "url")
     # @param [Hash] params Other parameters (see OpenAI docs)
     # @raise (see LLM::Provider#request)
     # @return [LLM::Response]
-    def edit(image:, prompt:, model: "dall-e-2", **params)
+    def edit(image:, prompt:, model: "dall-e-2", response_format: "b64_json", **params)
       image = LLM.File(image)
-      multi = LLM::Multipart.new(params.merge!(image:, prompt:, model:))
+      multi = LLM::Multipart.new(params.merge!(image:, prompt:, model:, response_format:))
       req = Net::HTTP::Post.new("/v1/images/edits", headers)
       req["content-type"] = multi.content_type
       set_body_stream(req, multi.body)

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

@@ -41,7 +41,7 @@ class LLM::OpenAI
       query = URI.encode_www_form(params)
       req = Net::HTTP::Get.new("/v1/models?#{query}", headers)
       res, span, tracer = execute(request: req, operation: "request")
-      res = ResponseAdapter.adapt(res, type: :enumerable)
+      res = ResponseAdapter.adapt(res, type: :models)
       tracer.on_request_finish(operation: "request", res:, span:)
       res
     end

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/models.rb

@@ -0,0 +1,13 @@
+# frozen_string_literal: true
+
+module LLM::OpenAI::ResponseAdapter
+  module Models
+    include LLM::Model::Collection
+
+    private
+
+    def raw_models
+      data || []
+    end
+  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/response_adapter.rb

@@ -11,6 +11,7 @@ class LLM::OpenAI
     require_relative "response_adapter/file"
     require_relative "response_adapter/image"
     require_relative "response_adapter/moderations"
+    require_relative "response_adapter/models"
     require_relative "response_adapter/responds"
     require_relative "response_adapter/web_search"
 
@@ -37,6 +38,7 @@ class LLM::OpenAI
       when :file then LLM::OpenAI::ResponseAdapter::File
       when :image then LLM::OpenAI::ResponseAdapter::Image
       when :moderations then LLM::OpenAI::ResponseAdapter::Moderations
+      when :models then LLM::OpenAI::ResponseAdapter::Models
       when :responds then LLM::OpenAI::ResponseAdapter::Responds
       when :web_search then LLM::OpenAI::ResponseAdapter::WebSearch
       else