llm.rb 0.7.2 → 0.9.0

data/lib/llm/providers/llamacpp.rb

@@ -1,5 +1,7 @@
 # frozen_string_literal: true
 
+require_relative "openai" unless defined?(LLM::OpenAI)
+
 module LLM
   ##
   # The LlamaCpp class implements a provider for
@@ -32,12 +34,24 @@ module LLM
       raise NotImplementedError
     end
 
+    ##
+    # @raise [NotImplementedError]
+    def moderations
+      raise NotImplementedError
+    end
+
+    ##
+    # @raise [NotImplementedError]
+    def responses
+      raise NotImplementedError
+    end
+
     ##
     # Returns the default model for chat completions
-    # @see https://ollama.com/library llama3.2
+    # @see https://ollama.com/library/qwen3 qwen3
     # @return [String]
     def default_model
-      "llama3.2"
+      "qwen3"
     end
   end
 end

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

@@ -23,11 +23,11 @@ class LLM::Ollama
     def raise_error!
       case res
       when Net::HTTPUnauthorized
-        raise LLM::Error::Unauthorized.new { _1.response = res }, "Authentication error"
+        raise LLM::UnauthorizedError.new { _1.response = res }, "Authentication error"
       when Net::HTTPTooManyRequests
-        raise LLM::Error::RateLimit.new { _1.response = res }, "Too many requests"
+        raise LLM::RateLimitError.new { _1.response = res }, "Too many requests"
       else
-        raise LLM::Error::ResponseError.new { _1.response = res }, "Unexpected response"
+        raise LLM::ResponseError.new { _1.response = res }, "Unexpected response"
       end
     end
   end

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

@@ -63,7 +63,7 @@ module LLM::Ollama::Format
       elsif returns.any?
         returns.map { {role: "tool", tool_call_id: _1.id, content: JSON.dump(_1.value)} }
       else
-        [{role: message.role, content: content.flat_map { format_content(_1) }}]
+        content.flat_map { {role: message.role}.merge(format_content(_1)) }
       end
     end
 

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

@@ -43,7 +43,7 @@ class LLM::Ollama
     def all(**params)
       query = URI.encode_www_form(params)
       req = Net::HTTP::Get.new("/api/tags?#{query}", headers)
-      res = request(http, req)
+      res = execute(request: req)
       LLM::Response::ModelList.new(res).tap { |modellist|
         models = modellist.body["models"].map do |model|
           model = model.transform_keys { snakecase(_1) }
@@ -55,12 +55,8 @@ class LLM::Ollama
 
     private
 
-    def http
-      @provider.instance_variable_get(:@http)
-    end
-
-    [:headers, :request].each do |m|
-      define_method(m) { |*args, &b| @provider.send(m, *args, &b) }
+    [:headers, :execute].each do |m|
+      define_method(m) { |*args, **kwargs, &b| @provider.send(m, *args, **kwargs, &b) }
     end
   end
 end

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

@@ -5,7 +5,7 @@ module LLM::Ollama::ResponseParser
   # @private
   class CompletionParser
     def initialize(body)
-      @body = OpenStruct.from_hash(body)
+      @body = LLM::Object.from_hash(body)
     end
 
     def format(response)
@@ -29,7 +29,7 @@ module LLM::Ollama::ResponseParser
       return [] unless tools
       tools.filter_map do |tool|
         next unless tool["function"]
-        OpenStruct.new(tool["function"])
+        LLM::Object.new(tool["function"])
       end
     end
 

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

@@ -0,0 +1,44 @@
+# frozen_string_literal: true
+
+class LLM::Ollama
+  ##
+  # @private
+  class StreamParser
+    ##
+    # Returns the fully constructed response body
+    # @return [LLM::Object]
+    attr_reader :body
+
+    ##
+    # @return [LLM::OpenAI::Chunk]
+    def initialize(io)
+      @body = LLM::Object.new
+      @io = io
+    end
+
+    ##
+    # @param [Hash] chunk
+    # @return [LLM::OpenAI::Chunk]
+    def parse!(chunk)
+      tap { merge!(chunk) }
+    end
+
+    private
+
+    def merge!(chunk)
+      chunk.each do |key, value|
+        if key == "message"
+          if @body[key]
+            @body[key]["content"] << value["content"]
+            @io << value["content"] if @io.respond_to?(:<<)
+          else
+            @body[key] = value
+            @io << value["content"] if @io.respond_to?(:<<)
+          end
+        else
+          @body[key] = value
+        end
+      end
+    end
+  end
+end

data/lib/llm/providers/ollama.rb

@@ -14,15 +14,17 @@ module LLM
   #   require "llm"
   #
   #   llm = LLM.ollama(nil)
-  #   bot = LLM::Chat.new(llm, model: "llava").lazy
+  #   bot = LLM::Bot.new(llm, model: "llava")
   #   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"
     require_relative "ollama/format"
+    require_relative "ollama/stream_parser"
+    require_relative "ollama/response_parser"
     require_relative "ollama/models"
+
     include Format
 
     HOST = "localhost"
@@ -40,11 +42,11 @@ module LLM
     # @param params (see LLM::Provider#embed)
     # @raise (see LLM::Provider#request)
     # @return (see LLM::Provider#embed)
-    def embed(input, model: "llama3.2", **params)
+    def embed(input, model: default_model, **params)
       params   = {model:}.merge!(params)
       req      = Net::HTTP::Post.new("/v1/embeddings", headers)
       req.body = JSON.dump({input:}.merge!(params))
-      res      = request(@http, req)
+      res      = execute(request: req)
       Response::Embedding.new(res).extend(response_parser)
     end
 
@@ -59,14 +61,15 @@ module LLM
     #  When given an object a provider does not understand
     # @return (see LLM::Provider#complete)
     def complete(prompt, params = {})
-      params = {role: :user, model: default_model, stream: false}.merge!(params)
+      params = {role: :user, model: default_model, stream: true}.merge!(params)
       params = [params, {format: params[:schema]}, format_tools(params)].inject({}, &:merge!).compact
-      role = params.delete(:role)
+      role, stream = params.delete(:role), params.delete(:stream)
+      params[:stream] = true if stream.respond_to?(:<<) || stream == true
       req = Net::HTTP::Post.new("/api/chat", headers)
       messages = [*(params.delete(:messages) || []), LLM::Message.new(role, prompt)]
       body = JSON.dump({messages: [format(messages)].flatten}.merge!(params))
       set_body_stream(req, StringIO.new(body))
-      res = request(@http, req)
+      res = execute(request: req, stream:)
       Response::Completion.new(res).extend(response_parser)
     end
 
@@ -86,10 +89,10 @@ module LLM
 
     ##
     # Returns the default model for chat completions
-    # @see https://ollama.com/library llama3.2
+    # @see https://ollama.com/library/qwen3 qwen3
     # @return [String]
     def default_model
-      "llama3.2"
+      "qwen3:latest"
     end
 
     private
@@ -105,6 +108,10 @@ module LLM
       LLM::Ollama::ResponseParser
     end
 
+    def stream_parser
+      LLM::Ollama::StreamParser
+    end
+
     def error_handler
       LLM::Ollama::ErrorHandler
     end

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

@@ -35,7 +35,7 @@ class LLM::OpenAI
       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 } }
+      res = execute(request: req) { _1.read_body { |chunk| io << chunk } }
       LLM::Response::Audio.new(res).tap { _1.audio = io }
     end
 
@@ -56,7 +56,7 @@ class LLM::OpenAI
       req = Net::HTTP::Post.new("/v1/audio/transcriptions", headers)
       req["content-type"] = multi.content_type
       set_body_stream(req, multi.body)
-      res = request(http, req)
+      res = execute(request: req)
       LLM::Response::AudioTranscription.new(res).tap { _1.text = _1.body["text"] }
     end
 
@@ -78,18 +78,14 @@ class LLM::OpenAI
       req = Net::HTTP::Post.new("/v1/audio/translations", headers)
       req["content-type"] = multi.content_type
       set_body_stream(req, multi.body)
-      res = request(http, req)
+      res = execute(request: req)
       LLM::Response::AudioTranslation.new(res).tap { _1.text = _1.body["text"] }
     end
 
     private
 
-    def http
-      @provider.instance_variable_get(:@http)
-    end
-
-    [:headers, :request, :set_body_stream].each do |m|
-      define_method(m) { |*args, &b| @provider.send(m, *args, &b) }
+    [:headers, :execute, :set_body_stream].each do |m|
+      define_method(m) { |*args, **kwargs, &b| @provider.send(m, *args, **kwargs, &b) }
     end
   end
 end

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

@@ -23,11 +23,11 @@ class LLM::OpenAI
     def raise_error!
       case res
       when Net::HTTPUnauthorized
-        raise LLM::Error::Unauthorized.new { _1.response = res }, "Authentication error"
+        raise LLM::UnauthorizedError.new { _1.response = res }, "Authentication error"
       when Net::HTTPTooManyRequests
-        raise LLM::Error::RateLimit.new { _1.response = res }, "Too many requests"
+        raise LLM::RateLimitError.new { _1.response = res }, "Too many requests"
       else
-        raise LLM::Error::ResponseError.new { _1.response = res }, "Unexpected response"
+        raise LLM::ResponseError.new { _1.response = res }, "Unexpected response"
       end
     end
   end

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

@@ -8,22 +8,23 @@ class LLM::OpenAI
   # and API endpoints. OpenAI supports multiple file formats, including text
   # files, CSV files, JSON files, and more.
   #
-  # @example
+  # @example example #1
   #   #!/usr/bin/env ruby
   #   require "llm"
   #
   #   llm = LLM.openai(ENV["KEY"])
-  #   bot = LLM::Chat.new(llm).lazy
+  #   bot = LLM::Bot.new(llm)
   #   file = llm.files.create file: "/documents/freebsd.pdf"
   #   bot.chat(file)
   #   bot.chat("Describe the document")
   #   bot.messages.select(&:assistant?).each { print "[#{_1.role}]", _1.content, "\n" }
-  # @example
+  #
+  # @example example #2
   #   #!/usr/bin/env ruby
   #   require "llm"
   #
   #   llm = LLM.openai(ENV["KEY"])
-  #   bot = LLM::Chat.new(llm).lazy
+  #   bot = LLM::Bot.new(llm)
   #   file = llm.files.create file: "/documents/openbsd.pdf"
   #   bot.chat(["Describe the document I sent to you", file])
   #   bot.messages.select(&:assistant?).each { print "[#{_1.role}]", _1.content, "\n" }
@@ -51,9 +52,9 @@ class LLM::OpenAI
     def all(**params)
       query = URI.encode_www_form(params)
       req = Net::HTTP::Get.new("/v1/files?#{query}", headers)
-      res = request(http, req)
+      res = execute(request: req)
       LLM::Response::FileList.new(res).tap { |filelist|
-        files = filelist.body["data"].map { OpenStruct.from_hash(_1) }
+        files = filelist.body["data"].map { LLM::Object.from_hash(_1) }
         filelist.files = files
       }
     end
@@ -74,7 +75,7 @@ class LLM::OpenAI
       req = Net::HTTP::Post.new("/v1/files", headers)
       req["content-type"] = multi.content_type
       set_body_stream(req, multi.body)
-      res = request(http, req)
+      res = execute(request: req)
       LLM::Response::File.new(res)
     end
 
@@ -93,7 +94,7 @@ class LLM::OpenAI
       file_id = file.respond_to?(:id) ? file.id : file
       query = URI.encode_www_form(params)
       req = Net::HTTP::Get.new("/v1/files/#{file_id}?#{query}", headers)
-      res = request(http, req)
+      res = execute(request: req)
       LLM::Response::File.new(res)
     end
 
@@ -114,7 +115,7 @@ class LLM::OpenAI
       file_id = file.respond_to?(:id) ? file.id : file
       req = Net::HTTP::Get.new("/v1/files/#{file_id}/content?#{query}", headers)
       io = StringIO.new("".b)
-      res = request(http, req) { |res| res.read_body { |chunk| io << chunk } }
+      res = execute(request: req) { |res| res.read_body { |chunk| io << chunk } }
       LLM::Response::DownloadFile.new(res).tap { _1.file = io }
     end
 
@@ -127,22 +128,18 @@ class LLM::OpenAI
     # @see https://platform.openai.com/docs/api-reference/files/delete OpenAI docs
     # @param [#id, #to_s] file The file ID
     # @raise (see LLM::Provider#request)
-    # @return [OpenStruct] Response body
+    # @return [LLM::Object] Response body
     def delete(file:)
       file_id = file.respond_to?(:id) ? file.id : file
       req = Net::HTTP::Delete.new("/v1/files/#{file_id}", headers)
-      res = request(http, req)
-      OpenStruct.from_hash JSON.parse(res.body)
+      res = execute(request: req)
+      LLM::Object.from_hash JSON.parse(res.body)
     end
 
     private
 
-    def http
-      @provider.instance_variable_get(:@http)
-    end
-
-    [:headers, :request, :set_body_stream].each do |m|
-      define_method(m) { |*args, &b| @provider.send(m, *args, &b) }
+    [:headers, :execute, :set_body_stream].each do |m|
+      define_method(m) { |*args, **kwargs, &b| @provider.send(m, *args, **kwargs, &b) }
     end
   end
 end

data/lib/llm/providers/openai/format/moderation_format.rb

@@ -0,0 +1,35 @@
+# frozen_string_literal: true
+
+module LLM::OpenAI::Format
+  ##
+  # @private
+  class ModerationFormat
+    ##
+    # @param [String, URI, Array<String, URI>] inputs
+    #  The inputs to format
+    # @return [LLM::OpenAI::Format::ModerationFormat]
+    def initialize(inputs)
+      @inputs = inputs
+    end
+
+    ##
+    # Formats the inputs for the OpenAI moderations API
+    # @return [Array<Hash>]
+    def format
+      [*inputs].flat_map do |input|
+        if String === input
+          {type: :text, text: input}
+        elsif URI === input
+          {type: :image_url, url: input.to_s}
+        else
+          raise LLM::Error::FormatError, "The given object (an instance of #{input.class}) " \
+                                         "is not supported by OpenAI moderations API"
+        end
+      end
+    end
+
+    private
+
+    attr_reader :inputs
+  end
+end

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

@@ -6,6 +6,7 @@ class LLM::OpenAI
   module Format
     require_relative "format/completion_format"
     require_relative "format/respond_format"
+    require_relative "format/moderation_format"
 
     ##
     # @param [Array<LLM::Message>] messages
@@ -43,9 +44,8 @@ class LLM::OpenAI
     # @param [Hash] params
     # @return [Hash]
     def format_tools(params)
-      return {} unless params and params[:tools]&.any?
-      tools = params[:tools]
-      {tools: tools.map { _1.format(self) }}
+      tools = params.delete(:tools)
+      (tools.nil? || tools.empty?) ? {} : {tools: tools.map { _1.format(self) }}
     end
   end
 end

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

@@ -7,7 +7,7 @@ class LLM::OpenAI
   # OpenAI supports multiple response formats: temporary URLs, or binary strings
   # encoded in base64. The default is to return temporary URLs.
   #
-  # @example
+  # @example example #1
   #   #!/usr/bin/env ruby
   #   require "llm"
   #   require "open-uri"
@@ -17,7 +17,8 @@ class LLM::OpenAI
   #   res = llm.images.create prompt: "A dog on a rocket to the moon"
   #   FileUtils.mv OpenURI.open_uri(res.urls[0]).path,
   #                "rocket.png"
-  # @example
+  #
+  # @example example #2
   #   #!/usr/bin/env ruby
   #   require "llm"
   #
@@ -49,7 +50,7 @@ class LLM::OpenAI
     def create(prompt:, model: "dall-e-3", **params)
       req = Net::HTTP::Post.new("/v1/images/generations", headers)
       req.body = JSON.dump({prompt:, n: 1, model:}.merge!(params))
-      res = request(http, req)
+      res = execute(request: req)
       LLM::Response::Image.new(res).extend(response_parser)
     end
 
@@ -71,7 +72,7 @@ class LLM::OpenAI
       req = Net::HTTP::Post.new("/v1/images/variations", headers)
       req["content-type"] = multi.content_type
       set_body_stream(req, multi.body)
-      res = request(http, req)
+      res = execute(request: req)
       LLM::Response::Image.new(res).extend(response_parser)
     end
 
@@ -94,18 +95,14 @@ class LLM::OpenAI
       req = Net::HTTP::Post.new("/v1/images/edits", headers)
       req["content-type"] = multi.content_type
       set_body_stream(req, multi.body)
-      res = request(http, req)
+      res = execute(request: req)
       LLM::Response::Image.new(res).extend(response_parser)
     end
 
     private
 
-    def http
-      @provider.instance_variable_get(:@http)
-    end
-
-    [:response_parser, :headers, :request, :set_body_stream].each do |m|
-      define_method(m) { |*args, &b| @provider.send(m, *args, &b) }
+    [:response_parser, :headers, :execute, :set_body_stream].each do |m|
+      define_method(m) { |*args, **kwargs, &b| @provider.send(m, *args, **kwargs, &b) }
     end
   end
 end

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

@@ -40,7 +40,7 @@ class LLM::OpenAI
     def all(**params)
       query = URI.encode_www_form(params)
       req = Net::HTTP::Get.new("/v1/models?#{query}", headers)
-      res = request(http, req)
+      res = execute(request: req)
       LLM::Response::ModelList.new(res).tap { |modellist|
         models = modellist.body["data"].map do |model|
           LLM::Model.from_hash(model).tap { _1.provider = @provider }
@@ -51,12 +51,8 @@ class LLM::OpenAI
 
     private
 
-    def http
-      @provider.instance_variable_get(:@http)
-    end
-
-    [:headers, :request, :set_body_stream].each do |m|
-      define_method(m) { |*args, &b| @provider.send(m, *args, &b) }
+    [:headers, :execute, :set_body_stream].each do |m|
+      define_method(m) { |*args, **kwargs, &b| @provider.send(m, *args, **kwargs, &b) }
     end
   end
 end

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

@@ -0,0 +1,67 @@
+# frozen_string_literal: true
+
+class LLM::OpenAI
+  ##
+  # The {LLM::OpenAI::Moderations LLM::OpenAI::Moderations} class provides a moderations
+  # object for interacting with [OpenAI's moderations API](https://platform.openai.com/docs/api-reference/moderations).
+  # The moderations API can categorize content into different categories, such as
+  # hate speech, self-harm, and sexual content. It can also provide a confidence score
+  # for each category.
+  #
+  # @example
+  #   #!/usr/bin/env ruby
+  #   require "llm"
+  #
+  #   llm = LLM.openai(key: ENV["KEY"])
+  #   mod = llm.moderations.create input: "I hate you"
+  #   print "categories: #{mod.categories}", "\n"
+  #   print "scores: #{mod.scores}", "\n"
+  #
+  # @example
+  #   #!/usr/bin/env ruby
+  #   require "llm"
+  #
+  #   llm = LLM.openai(key: ENV["KEY"])
+  #   mod = llm.moderations.create input: URI.parse("https://example.com/image.png")
+  #   print "categories: #{mod.categories}", "\n"
+  #   print "scores: #{mod.scores}", "\n"
+  #
+  # @see https://platform.openai.com/docs/api-reference/moderations/create OpenAI docs
+  # @see https://platform.openai.com/docs/models#moderation OpenAI moderation models
+  class Moderations
+    ##
+    # Returns a new Moderations object
+    # @param [LLM::Provider] provider
+    # @return [LLM::OpenAI::Moderations]
+    def initialize(provider)
+      @provider = provider
+    end
+
+    ##
+    # Create a moderation
+    # @see https://platform.openai.com/docs/api-reference/moderations/create OpenAI docs
+    # @see https://platform.openai.com/docs/models#moderation OpenAI moderation models
+    # @note
+    # Although OpenAI mentions an array as a valid input, and that it can return one
+    # or more moderations, in practice the API only returns one moderation object. We
+    # recommend using a single input string or URI, and to keep in mind that llm.rb
+    # returns a Moderation object but has code in place to return multiple objects in
+    # the future (in case OpenAI documentation ever matches the actual API).
+    # @param [String, URI, Array<String, URI>] input
+    # @param [String, LLM::Model] model The model to use
+    # @return [LLM::Response::ModerationList::Moderation]
+    def create(input:, model: "omni-moderation-latest", **params)
+      req = Net::HTTP::Post.new("/v1/moderations", headers)
+      input = Format::ModerationFormat.new(input).format
+      req.body = JSON.dump({input:, model:}.merge!(params))
+      res = execute(request: req)
+      LLM::Response::ModerationList.new(res).extend(response_parser).first
+    end
+
+    private
+
+    [:response_parser, :headers, :execute].each do |m|
+      define_method(m) { |*args, **kwargs, &b| @provider.send(m, *args, **kwargs, &b) }
+    end
+  end
+end

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

@@ -5,7 +5,7 @@ module LLM::OpenAI::ResponseParser
   # @private
   class CompletionParser
     def initialize(body)
-      @body = OpenStruct.from_hash(body)
+      @body = LLM::Object.from_hash(body)
     end
 
     def format(response)
@@ -41,15 +41,15 @@ module LLM::OpenAI::ResponseParser
           name: tool.function.name,
           arguments: JSON.parse(tool.function.arguments)
         }
-        OpenStruct.new(tool)
+        LLM::Object.new(tool)
       end
     end
 
     def body = @body
     def model = body.model
-    def prompt_tokens = body.usage.prompt_tokens
-    def completion_tokens = body.usage.completion_tokens
-    def total_tokens = body.usage.total_tokens
+    def prompt_tokens = body.usage&.prompt_tokens
+    def completion_tokens = body.usage&.completion_tokens
+    def total_tokens = body.usage&.total_tokens
     def choices = body.choices
   end
 end

data/lib/llm/providers/openai/response_parser/respond_parser.rb

@@ -5,7 +5,7 @@ module LLM::OpenAI::ResponseParser
   # @private
   class RespondParser
     def initialize(body)
-      @body = OpenStruct.from_hash(body)
+      @body = LLM::Object.from_hash(body)
     end
 
     def format(response)
@@ -37,7 +37,7 @@ module LLM::OpenAI::ResponseParser
     end
 
     def format_tool(tool)
-      OpenStruct.new(
+      LLM::Object.new(
         id: tool.call_id,
         name: tool.name,
         arguments: JSON.parse(tool.arguments)

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

@@ -4,6 +4,9 @@ class LLM::OpenAI
   ##
   # @private
   module ResponseParser
+    require_relative "response_parser/completion_parser"
+    require_relative "response_parser/respond_parser"
+
     ##
     # @param [Hash] body
     #  The response body from the LLM provider
@@ -20,6 +23,18 @@ class LLM::OpenAI
       RespondParser.new(body).format(self)
     end
 
+    ##
+    # @param [Hash] body
+    #  The response body from the LLM provider
+    # @return [Hash]
+    def parse_moderation_list(body)
+      {
+        id: body["id"],
+        model: body["model"],
+        moderations: body["results"].map { LLM::Response::ModerationList::Moderation.new(_1) }
+      }
+    end
+
     ##
     # @param [Hash] body
     #  The response body from the LLM provider