dify_llm 1.9.1 → 1.14.1

data/lib/ruby_llm/providers/vertexai.rb

@@ -10,13 +10,22 @@ module RubyLLM
       include VertexAI::Models
       include VertexAI::Transcription
 
+      SCOPES = [
+        'https://www.googleapis.com/auth/cloud-platform',
+        'https://www.googleapis.com/auth/generative-language.retriever'
+      ].freeze
+
       def initialize(config)
         super
         @authorizer = nil
       end
 
       def api_base
-        "https://#{@config.vertexai_location}-aiplatform.googleapis.com/v1beta1"
+        if @config.vertexai_location.to_s == 'global'
+          'https://aiplatform.googleapis.com/v1beta1'
+        else
+          "https://#{@config.vertexai_location}-aiplatform.googleapis.com/v1beta1"
+        end
       end
 
       def headers
@@ -31,6 +40,10 @@ module RubyLLM
       end
 
       class << self
+        def configuration_options
+          %i[vertexai_project_id vertexai_location vertexai_service_account_key]
+        end
+
         def configuration_requirements
           %i[vertexai_project_id vertexai_location]
         end
@@ -40,12 +53,15 @@ module RubyLLM
 
       def initialize_authorizer
         require 'googleauth'
-        @authorizer = ::Google::Auth.get_application_default(
-          scope: [
-            'https://www.googleapis.com/auth/cloud-platform',
-            'https://www.googleapis.com/auth/generative-language.retriever'
-          ]
-        )
+        @authorizer =
+          if @config.vertexai_service_account_key
+            ::Google::Auth::ServiceAccountCredentials.make_creds(
+              json_key_io: StringIO.new(@config.vertexai_service_account_key),
+              scope: SCOPES
+            )
+          else
+            ::Google::Auth.get_application_default(SCOPES)
+          end
       rescue LoadError
         raise Error,
               'The googleauth gem ~> 1.15 is required for Vertex AI. Please add it to your Gemfile: gem "googleauth"'

data/lib/ruby_llm/providers/xai/chat.rb

@@ -0,0 +1,15 @@
+# frozen_string_literal: true
+
+module RubyLLM
+  module Providers
+    class XAI
+      # Chat implementation for xAI
+      # https://docs.x.ai/docs/api-reference#chat-completions
+      module Chat
+        def format_role(role)
+          role.to_s
+        end
+      end
+    end
+  end
+end

data/lib/ruby_llm/providers/xai/models.rb

@@ -0,0 +1,75 @@
+# frozen_string_literal: true
+
+module RubyLLM
+  module Providers
+    class XAI
+      # Models metadata for xAI list models.
+      module Models
+        module_function
+
+        IMAGE_MODELS = %w[grok-2-image-1212].freeze
+        VISION_MODELS = %w[
+          grok-2-vision-1212
+          grok-4-0709
+          grok-4-fast-non-reasoning
+          grok-4-fast-reasoning
+          grok-4-1-fast-non-reasoning
+          grok-4-1-fast-reasoning
+        ].freeze
+        REASONING_MODELS = %w[
+          grok-3-mini
+          grok-4-0709
+          grok-4-fast-reasoning
+          grok-4-1-fast-reasoning
+          grok-code-fast-1
+        ].freeze
+
+        def parse_list_models_response(response, slug, _capabilities)
+          Array(response.body['data']).map do |model_data|
+            model_id = model_data['id']
+
+            Model::Info.new(
+              id: model_id,
+              name: format_display_name(model_id),
+              provider: slug,
+              family: 'grok',
+              created_at: model_data['created'] ? Time.at(model_data['created']) : nil,
+              context_window: nil,
+              max_output_tokens: nil,
+              modalities: modalities_for(model_id),
+              capabilities: capabilities_for(model_id),
+              pricing: {},
+              metadata: {
+                object: model_data['object'],
+                owned_by: model_data['owned_by']
+              }.compact
+            )
+          end
+        end
+
+        def modalities_for(model_id)
+          if IMAGE_MODELS.include?(model_id)
+            { input: ['text'], output: ['image'] }
+          else
+            input = ['text']
+            input << 'image' if VISION_MODELS.include?(model_id)
+            { input: input, output: ['text'] }
+          end
+        end
+
+        def capabilities_for(model_id)
+          return [] if IMAGE_MODELS.include?(model_id)
+
+          capabilities = %w[streaming function_calling structured_output]
+          capabilities << 'reasoning' if REASONING_MODELS.include?(model_id)
+          capabilities << 'vision' if VISION_MODELS.include?(model_id)
+          capabilities
+        end
+
+        def format_display_name(model_id)
+          model_id.tr('-', ' ').split.map(&:capitalize).join(' ')
+        end
+      end
+    end
+  end
+end

data/lib/ruby_llm/providers/xai.rb

@@ -0,0 +1,32 @@
+# frozen_string_literal: true
+
+module RubyLLM
+  module Providers
+    # xAI API integration
+    class XAI < OpenAI
+      include XAI::Chat
+      include XAI::Models
+
+      def api_base
+        'https://api.x.ai/v1'
+      end
+
+      def headers
+        {
+          'Authorization' => "Bearer #{@config.xai_api_key}",
+          'Content-Type' => 'application/json'
+        }
+      end
+
+      class << self
+        def configuration_options
+          %i[xai_api_key]
+        end
+
+        def configuration_requirements
+          %i[xai_api_key]
+        end
+      end
+    end
+  end
+end

data/lib/ruby_llm/stream_accumulator.rb

@@ -7,40 +7,48 @@ module RubyLLM
 
     def initialize
       @content = +''
+      @thinking_text = +''
+      @thinking_signature = nil
       @tool_calls = {}
       @input_tokens = nil
       @output_tokens = nil
       @cached_tokens = nil
       @cache_creation_tokens = nil
+      @thinking_tokens = nil
+      @inside_think_tag = false
+      @pending_think_tag = +''
       @latest_tool_call_id = nil
     end
 
     def add(chunk)
-      RubyLLM.logger.debug chunk.inspect if RubyLLM.config.log_stream_debug
+      RubyLLM.logger.debug { chunk.inspect } if RubyLLM.config.log_stream_debug
       @model_id ||= chunk.model_id
       @conversation_id ||= chunk.conversation_id
 
-      if chunk.tool_call?
-        accumulate_tool_calls chunk.tool_calls
-      else
-        @content << (chunk.content || '')
-      end
-
+      handle_chunk_content(chunk)
+      append_thinking_from_chunk(chunk)
       count_tokens chunk
-      RubyLLM.logger.debug inspect if RubyLLM.config.log_stream_debug
+      RubyLLM.logger.debug { inspect } if RubyLLM.config.log_stream_debug
     end
 
     def to_message(response)
       Message.new(
         role: :assistant,
         content: content.empty? ? nil : content,
+        thinking: Thinking.build(
+          text: @thinking_text.empty? ? nil : @thinking_text,
+          signature: @thinking_signature
+        ),
+        tokens: Tokens.build(
+          input: @input_tokens,
+          output: @output_tokens,
+          cached: @cached_tokens,
+          cache_creation: @cache_creation_tokens,
+          thinking: @thinking_tokens
+        ),
         model_id: model_id,
         conversation_id: conversation_id,
         tool_calls: tool_calls_from_stream,
-        input_tokens: @input_tokens,
-        output_tokens: @output_tokens,
-        cached_tokens: @cached_tokens,
-        cache_creation_tokens: @cache_creation_tokens,
         raw: response
       )
     end
@@ -60,26 +68,38 @@ module RubyLLM
         ToolCall.new(
           id: tc.id,
           name: tc.name,
-          arguments: arguments
+          arguments: arguments,
+          thought_signature: tc.thought_signature
         )
       end
     end
 
-    def accumulate_tool_calls(new_tool_calls)
-      RubyLLM.logger.debug "Accumulating tool calls: #{new_tool_calls}" if RubyLLM.config.log_stream_debug
+    def accumulate_tool_calls(new_tool_calls) # rubocop:disable Metrics/PerceivedComplexity
+      RubyLLM.logger.debug { "Accumulating tool calls: #{new_tool_calls}" } if RubyLLM.config.log_stream_debug
       new_tool_calls.each_value do |tool_call|
         if tool_call.id
           tool_call_id = tool_call.id.empty? ? SecureRandom.uuid : tool_call.id
-          tool_call_arguments = tool_call.arguments.empty? ? +'' : tool_call.arguments
+          tool_call_arguments = tool_call.arguments
+          if tool_call_arguments.nil? || (tool_call_arguments.respond_to?(:empty?) && tool_call_arguments.empty?)
+            tool_call_arguments = +''
+          end
           @tool_calls[tool_call.id] = ToolCall.new(
             id: tool_call_id,
             name: tool_call.name,
-            arguments: tool_call_arguments
+            arguments: tool_call_arguments,
+            thought_signature: tool_call.thought_signature
           )
           @latest_tool_call_id = tool_call.id
         else
           existing = @tool_calls[@latest_tool_call_id]
-          existing.arguments << tool_call.arguments if existing
+          if existing
+            fragment = tool_call.arguments
+            fragment = '' if fragment.nil?
+            existing.arguments << fragment
+            if tool_call.thought_signature && existing.thought_signature.nil?
+              existing.thought_signature = tool_call.thought_signature
+            end
+          end
         end
       end
     end
@@ -98,6 +118,88 @@ module RubyLLM
       @output_tokens = chunk.output_tokens if chunk.output_tokens
       @cached_tokens = chunk.cached_tokens if chunk.cached_tokens
       @cache_creation_tokens = chunk.cache_creation_tokens if chunk.cache_creation_tokens
+      @thinking_tokens = chunk.thinking_tokens if chunk.thinking_tokens
+    end
+
+    def handle_chunk_content(chunk)
+      return accumulate_tool_calls(chunk.tool_calls) if chunk.tool_call?
+
+      content_text = chunk.content || ''
+      if content_text.is_a?(String)
+        append_text_with_thinking(content_text)
+      else
+        @content << content_text.to_s
+      end
+    end
+
+    def append_text_with_thinking(text)
+      content_chunk, thinking_chunk = extract_think_tags(text)
+      @content << content_chunk
+      @thinking_text << thinking_chunk if thinking_chunk
+    end
+
+    def append_thinking_from_chunk(chunk)
+      thinking = chunk.thinking
+      return unless thinking
+
+      @thinking_text << thinking.text.to_s if thinking.text
+      @thinking_signature ||= thinking.signature # rubocop:disable Naming/MemoizedInstanceVariableName
+    end
+
+    def extract_think_tags(text)
+      start_tag = '<think>'
+      end_tag = '</think>'
+      remaining = @pending_think_tag + text
+      @pending_think_tag = +''
+
+      output = +''
+      thinking = +''
+
+      until remaining.empty?
+        remaining = if @inside_think_tag
+                      consume_think_content(remaining, end_tag, thinking)
+                    else
+                      consume_non_think_content(remaining, start_tag, output)
+                    end
+      end
+
+      [output, thinking.empty? ? nil : thinking]
+    end
+
+    def consume_think_content(remaining, end_tag, thinking)
+      end_index = remaining.index(end_tag)
+      if end_index
+        thinking << remaining.slice(0, end_index)
+        @inside_think_tag = false
+        remaining.slice((end_index + end_tag.length)..) || +''
+      else
+        suffix_len = longest_suffix_prefix(remaining, end_tag)
+        thinking << remaining.slice(0, remaining.length - suffix_len)
+        @pending_think_tag = remaining.slice(-suffix_len, suffix_len)
+        +''
+      end
+    end
+
+    def consume_non_think_content(remaining, start_tag, output)
+      start_index = remaining.index(start_tag)
+      if start_index
+        output << remaining.slice(0, start_index)
+        @inside_think_tag = true
+        remaining.slice((start_index + start_tag.length)..) || +''
+      else
+        suffix_len = longest_suffix_prefix(remaining, start_tag)
+        output << remaining.slice(0, remaining.length - suffix_len)
+        @pending_think_tag = remaining.slice(-suffix_len, suffix_len)
+        +''
+      end
+    end
+
+    def longest_suffix_prefix(text, tag)
+      max = [text.length, tag.length - 1].min
+      max.downto(1) do |len|
+        return len if text.end_with?(tag[0, len])
+      end
+      0
     end
   end
 end

data/lib/ruby_llm/streaming.rb

@@ -24,13 +24,13 @@ module RubyLLM
       end
 
       message = accumulator.to_message(response)
-      RubyLLM.logger.debug "Stream completed: #{message.content}"
+      RubyLLM.logger.debug { "Stream completed: #{message.content}" }
       message
     end
 
     def handle_stream(&block)
-      to_json_stream do |data|
-        block.call(build_chunk(data)) if data
+      build_on_data_handler do |data|
+        block.call(build_chunk(data)) if data.is_a?(Hash)
       end
     end
 
@@ -40,74 +40,52 @@ module RubyLLM
       Faraday::VERSION.start_with?('1')
     end
 
-    def to_json_stream(&)
+    def build_on_data_handler(&handler)
       buffer = +''
       parser = EventStreamParser::Parser.new
 
-      create_stream_processor(parser, buffer, &)
-    end
-
-    def create_stream_processor(parser, buffer, &)
-      if faraday_1?
-        legacy_stream_processor(parser, &)
-      else
-        stream_processor(parser, buffer, &)
-      end
+      FaradayHandlers.build(
+        faraday_v1: faraday_1?,
+        on_chunk: ->(chunk, env) { process_stream_chunk(chunk, parser, env, &handler) },
+        on_failed_response: ->(chunk, env) { handle_failed_response(chunk, buffer, env) }
+      )
     end
 
     def process_stream_chunk(chunk, parser, env, &)
-      RubyLLM.logger.debug "Received chunk: #{chunk}" if RubyLLM.config.log_stream_debug
+      RubyLLM.logger.debug { "Received chunk: #{chunk}" } if RubyLLM.config.log_stream_debug
 
       if error_chunk?(chunk)
         handle_error_chunk(chunk, env)
+      elsif json_error_payload?(chunk)
+        handle_json_error_chunk(chunk, env)
       else
         yield handle_sse(chunk, parser, env, &)
       end
     end
 
-    def legacy_stream_processor(parser, &block)
-      proc do |chunk, _size|
-        process_stream_chunk(chunk, parser, nil, &block)
-      end
+    def error_chunk?(chunk)
+      chunk.start_with?('event: error')
     end
 
-    def stream_processor(parser, buffer, &block)
-      proc do |chunk, _bytes, env|
-        if env&.status == 200
-          process_stream_chunk(chunk, parser, env, &block)
-        else
-          handle_failed_response(chunk, buffer, env)
-        end
-      end
+    def json_error_payload?(chunk)
+      chunk.lstrip.start_with?('{') && chunk.include?('"error"')
     end
 
-    def error_chunk?(chunk)
-      chunk.start_with?('event: error')
+    def handle_json_error_chunk(chunk, env)
+      parse_error_from_json(chunk, env, 'Failed to parse JSON error chunk')
     end
 
     def handle_error_chunk(chunk, env)
       error_data = chunk.split("\n")[1].delete_prefix('data: ')
-      status, _message = parse_streaming_error(error_data)
-      parsed_data = JSON.parse(error_data)
-
-      error_response = if faraday_1?
-                         Struct.new(:body, :status).new(parsed_data, status)
-                       else
-                         env.merge(body: parsed_data, status: status)
-                       end
-
-      ErrorMiddleware.parse_error(provider: self, response: error_response)
-    rescue JSON::ParserError => e
-      RubyLLM.logger.debug "Failed to parse error chunk: #{e.message}"
+      parse_error_from_json(error_data, env, 'Failed to parse error chunk')
     end
 
     def handle_failed_response(chunk, buffer, env)
       buffer << chunk
       error_data = JSON.parse(buffer)
-      error_response = env.merge(body: error_data)
-      ErrorMiddleware.parse_error(provider: self, response: error_response)
+      handle_parsed_error(error_data, env)
     rescue JSON::ParserError
-      RubyLLM.logger.debug "Accumulating error chunk: #{chunk}"
+      RubyLLM.logger.debug { "Accumulating error chunk: #{chunk}" }
     end
 
     def handle_sse(chunk, parser, env, &block)
@@ -116,38 +94,82 @@ module RubyLLM
         when :error
           handle_error_event(data, env)
         else
-          yield handle_data(data, &block) unless data == '[DONE]'
+          yield handle_data(data, env, &block) unless data == '[DONE]'
         end
       end
     end
 
-    def handle_data(data)
-      JSON.parse(data)
+    def handle_data(data, env)
+      parsed = JSON.parse(data)
+      return parsed unless parsed.is_a?(Hash) && parsed.key?('error')
+
+      handle_parsed_error(parsed, env)
     rescue JSON::ParserError => e
-      RubyLLM.logger.debug "Failed to parse data chunk: #{e.message}"
+      RubyLLM.logger.debug { "Failed to parse data chunk: #{e.message}" }
     end
 
     def handle_error_event(data, env)
-      status, _message = parse_streaming_error(data)
-      parsed_data = JSON.parse(data)
-
-      error_response = if faraday_1?
-                         Struct.new(:body, :status).new(parsed_data, status)
-                       else
-                         env.merge(body: parsed_data, status: status)
-                       end
-
-      ErrorMiddleware.parse_error(provider: self, response: error_response)
-    rescue JSON::ParserError => e
-      RubyLLM.logger.debug "Failed to parse error event: #{e.message}"
+      parse_error_from_json(data, env, 'Failed to parse error event')
     end
 
     def parse_streaming_error(data)
       error_data = JSON.parse(data)
       [500, error_data['message'] || 'Unknown streaming error']
     rescue JSON::ParserError => e
-      RubyLLM.logger.debug "Failed to parse streaming error: #{e.message}"
+      RubyLLM.logger.debug { "Failed to parse streaming error: #{e.message}" }
       [500, "Failed to parse error: #{data}"]
     end
+
+    def handle_parsed_error(parsed_data, env)
+      status, _message = parse_streaming_error(parsed_data.to_json)
+      error_response = build_stream_error_response(parsed_data, env, status)
+      ErrorMiddleware.parse_error(provider: self, response: error_response)
+    end
+
+    def parse_error_from_json(data, env, error_message)
+      parsed_data = JSON.parse(data)
+      handle_parsed_error(parsed_data, env)
+    rescue JSON::ParserError => e
+      RubyLLM.logger.debug { "#{error_message}: #{e.message}" }
+    end
+
+    def build_stream_error_response(parsed_data, env, status)
+      error_status = status || env&.status || 500
+
+      if faraday_1?
+        Struct.new(:body, :status).new(parsed_data, error_status)
+      else
+        env.merge(body: parsed_data, status: error_status)
+      end
+    end
+
+    # Builds Faraday on_data handlers for different major versions.
+    module FaradayHandlers
+      module_function
+
+      def build(faraday_v1:, on_chunk:, on_failed_response:)
+        if faraday_v1
+          v1_on_data(on_chunk)
+        else
+          v2_on_data(on_chunk, on_failed_response)
+        end
+      end
+
+      def v1_on_data(on_chunk)
+        proc do |chunk, _size|
+          on_chunk.call(chunk, nil)
+        end
+      end
+
+      def v2_on_data(on_chunk, on_failed_response)
+        proc do |chunk, _bytes, env|
+          if env&.status == 200
+            on_chunk.call(chunk, env)
+          else
+            on_failed_response.call(chunk, env)
+          end
+        end
+      end
+    end
   end
 end

data/lib/ruby_llm/thinking.rb

@@ -0,0 +1,49 @@
+# frozen_string_literal: true
+
+module RubyLLM
+  # Represents provider thinking output.
+  class Thinking
+    attr_reader :text, :signature
+
+    def initialize(text: nil, signature: nil)
+      @text = text
+      @signature = signature
+    end
+
+    def self.build(text: nil, signature: nil)
+      text = nil if text.is_a?(String) && text.empty?
+      signature = nil if signature.is_a?(String) && signature.empty?
+
+      return nil if text.nil? && signature.nil?
+
+      new(text: text, signature: signature)
+    end
+
+    def pretty_print(printer)
+      printer.object_group(self) do
+        printer.breakable
+        printer.text 'text='
+        printer.pp text
+        printer.comma_breakable
+        printer.text 'signature='
+        printer.pp(signature ? '[REDACTED]' : nil)
+      end
+    end
+  end
+
+  class Thinking
+    # Normalized config for thinking across providers.
+    class Config
+      attr_reader :effort, :budget
+
+      def initialize(effort: nil, budget: nil)
+        @effort = effort.is_a?(Symbol) ? effort.to_s : effort
+        @budget = budget
+      end
+
+      def enabled?
+        !effort.nil? || !budget.nil?
+      end
+    end
+  end
+end

data/lib/ruby_llm/tokens.rb

@@ -0,0 +1,47 @@
+# frozen_string_literal: true
+
+module RubyLLM
+  # Represents token usage for a response.
+  class Tokens
+    attr_reader :input, :output, :cached, :cache_creation, :thinking
+
+    # rubocop:disable Metrics/ParameterLists
+    def initialize(input: nil, output: nil, cached: nil, cache_creation: nil, thinking: nil, reasoning: nil)
+      @input = input
+      @output = output
+      @cached = cached
+      @cache_creation = cache_creation
+      @thinking = thinking || reasoning
+    end
+    # rubocop:enable Metrics/ParameterLists
+
+    # rubocop:disable Metrics/ParameterLists
+    def self.build(input: nil, output: nil, cached: nil, cache_creation: nil, thinking: nil, reasoning: nil)
+      return nil if [input, output, cached, cache_creation, thinking, reasoning].all?(&:nil?)
+
+      new(
+        input: input,
+        output: output,
+        cached: cached,
+        cache_creation: cache_creation,
+        thinking: thinking,
+        reasoning: reasoning
+      )
+    end
+    # rubocop:enable Metrics/ParameterLists
+
+    def to_h
+      {
+        input_tokens: input,
+        output_tokens: output,
+        cached_tokens: cached,
+        cache_creation_tokens: cache_creation,
+        thinking_tokens: thinking
+      }.compact
+    end
+
+    def reasoning
+      thinking
+    end
+  end
+end