ruby_llm 1.9.2 → 1.10.0

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

@@ -0,0 +1,154 @@
+# frozen_string_literal: true
+
+module RubyLLM
+  module Providers
+    class OpenRouter
+      # Chat methods of the OpenRouter API integration
+      module Chat
+        module_function
+
+        def render_payload(messages, tools:, temperature:, model:, stream: false, schema: nil, thinking: nil) # rubocop:disable Metrics/ParameterLists
+          payload = {
+            model: model.id,
+            messages: format_messages(messages),
+            stream: stream
+          }
+
+          payload[:temperature] = temperature unless temperature.nil?
+          payload[:tools] = tools.map { |_, tool| OpenAI::Tools.tool_for(tool) } if tools.any?
+
+          if schema
+            strict = schema[:strict] != false
+            payload[:response_format] = {
+              type: 'json_schema',
+              json_schema: {
+                name: 'response',
+                schema: schema,
+                strict: strict
+              }
+            }
+          end
+
+          reasoning = build_reasoning(thinking)
+          payload[:reasoning] = reasoning if reasoning
+
+          payload[:stream_options] = { include_usage: true } if stream
+          payload
+        end
+
+        def parse_completion_response(response)
+          data = response.body
+          return if data.empty?
+
+          raise Error.new(response, data.dig('error', 'message')) if data.dig('error', 'message')
+
+          message_data = data.dig('choices', 0, 'message')
+          return unless message_data
+
+          usage = data['usage'] || {}
+          cached_tokens = usage.dig('prompt_tokens_details', 'cached_tokens')
+          thinking_tokens = usage.dig('completion_tokens_details', 'reasoning_tokens')
+          thinking_text = extract_thinking_text(message_data)
+          thinking_signature = extract_thinking_signature(message_data)
+
+          Message.new(
+            role: :assistant,
+            content: message_data['content'],
+            thinking: Thinking.build(text: thinking_text, signature: thinking_signature),
+            tool_calls: OpenAI::Tools.parse_tool_calls(message_data['tool_calls']),
+            input_tokens: usage['prompt_tokens'],
+            output_tokens: usage['completion_tokens'],
+            cached_tokens: cached_tokens,
+            cache_creation_tokens: 0,
+            thinking_tokens: thinking_tokens,
+            model_id: data['model'],
+            raw: response
+          )
+        end
+
+        def format_messages(messages)
+          messages.map do |msg|
+            {
+              role: format_role(msg.role),
+              content: OpenAI::Media.format_content(msg.content),
+              tool_calls: OpenAI::Tools.format_tool_calls(msg.tool_calls),
+              tool_call_id: msg.tool_call_id
+            }.compact.merge(format_thinking(msg))
+          end
+        end
+
+        def format_role(role)
+          case role
+          when :system
+            @config.openai_use_system_role ? 'system' : 'developer'
+          else
+            role.to_s
+          end
+        end
+
+        def build_reasoning(thinking)
+          return nil unless thinking&.enabled?
+
+          reasoning = {}
+          reasoning[:effort] = thinking.effort if thinking.respond_to?(:effort) && thinking.effort
+          reasoning[:max_tokens] = thinking.budget if thinking.respond_to?(:budget) && thinking.budget
+          reasoning[:enabled] = true if reasoning.empty?
+          reasoning
+        end
+
+        def format_thinking(msg)
+          thinking = msg.thinking
+          return {} unless thinking && msg.role == :assistant
+
+          details = []
+          if thinking.text
+            details << {
+              type: 'reasoning.text',
+              text: thinking.text,
+              signature: thinking.signature
+            }.compact
+          elsif thinking.signature
+            details << {
+              type: 'reasoning.encrypted',
+              data: thinking.signature
+            }
+          end
+
+          details.empty? ? {} : { reasoning_details: details }
+        end
+
+        def extract_thinking_text(message_data)
+          candidate = message_data['reasoning']
+          return candidate if candidate.is_a?(String)
+
+          details = message_data['reasoning_details']
+          return nil unless details.is_a?(Array)
+
+          text = details.filter_map do |detail|
+            case detail['type']
+            when 'reasoning.text'
+              detail['text']
+            when 'reasoning.summary'
+              detail['summary']
+            end
+          end.join
+
+          text.empty? ? nil : text
+        end
+
+        def extract_thinking_signature(message_data)
+          details = message_data['reasoning_details']
+          return nil unless details.is_a?(Array)
+
+          signature = details.filter_map do |detail|
+            detail['signature'] if detail['signature'].is_a?(String)
+          end.first
+          return signature if signature
+
+          encrypted = details.find { |detail| detail['type'] == 'reasoning.encrypted' && detail['data'].is_a?(String) }
+          encrypted&.dig('data')
+        end
+      end
+    end
+  end
+end

data/lib/ruby_llm/providers/openrouter/streaming.rb

@@ -0,0 +1,74 @@
+# frozen_string_literal: true
+
+module RubyLLM
+  module Providers
+    class OpenRouter
+      # Streaming methods of the OpenRouter API integration
+      module Streaming
+        module_function
+
+        def stream_url
+          completion_url
+        end
+
+        def build_chunk(data)
+          usage = data['usage'] || {}
+          cached_tokens = usage.dig('prompt_tokens_details', 'cached_tokens')
+          delta = data.dig('choices', 0, 'delta') || {}
+
+          Chunk.new(
+            role: :assistant,
+            model_id: data['model'],
+            content: delta['content'],
+            thinking: Thinking.build(
+              text: extract_thinking_text(delta),
+              signature: extract_thinking_signature(delta)
+            ),
+            tool_calls: OpenAI::Tools.parse_tool_calls(delta['tool_calls'], parse_arguments: false),
+            input_tokens: usage['prompt_tokens'],
+            output_tokens: usage['completion_tokens'],
+            cached_tokens: cached_tokens,
+            cache_creation_tokens: 0,
+            thinking_tokens: usage.dig('completion_tokens_details', 'reasoning_tokens')
+          )
+        end
+
+        def parse_streaming_error(data)
+          OpenAI::Streaming.parse_streaming_error(data)
+        end
+
+        def extract_thinking_text(delta)
+          candidate = delta['reasoning']
+          return candidate if candidate.is_a?(String)
+
+          details = delta['reasoning_details']
+          return nil unless details.is_a?(Array)
+
+          text = details.filter_map do |detail|
+            case detail['type']
+            when 'reasoning.text'
+              detail['text']
+            when 'reasoning.summary'
+              detail['summary']
+            end
+          end.join
+
+          text.empty? ? nil : text
+        end
+
+        def extract_thinking_signature(delta)
+          details = delta['reasoning_details']
+          return nil unless details.is_a?(Array)
+
+          signature = details.filter_map do |detail|
+            detail['signature'] if detail['signature'].is_a?(String)
+          end.first
+          return signature if signature
+
+          encrypted = details.find { |detail| detail['type'] == 'reasoning.encrypted' && detail['data'].is_a?(String) }
+          encrypted&.dig('data')
+        end
+      end
+    end
+  end
+end

data/lib/ruby_llm/providers/openrouter.rb

@@ -4,7 +4,9 @@ module RubyLLM
   module Providers
     # OpenRouter API integration.
     class OpenRouter < OpenAI
+      include OpenRouter::Chat
       include OpenRouter::Models
+      include OpenRouter::Streaming
 
       def api_base
         'https://openrouter.ai/api/v1'

data/lib/ruby_llm/providers/vertexai.rb

@@ -16,7 +16,11 @@ module RubyLLM
       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

data/lib/ruby_llm/stream_accumulator.rb

@@ -7,11 +7,16 @@ 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
 
@@ -19,12 +24,8 @@ module RubyLLM
       RubyLLM.logger.debug chunk.inspect if RubyLLM.config.log_stream_debug
       @model_id ||= chunk.model_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
     end
@@ -33,12 +34,19 @@ module RubyLLM
       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,
         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
@@ -58,12 +66,13 @@ 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)
+    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
@@ -72,12 +81,18 @@ module RubyLLM
           @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
+            existing.arguments << tool_call.arguments
+            if tool_call.thought_signature && existing.thought_signature.nil?
+              existing.thought_signature = tool_call.thought_signature
+            end
+          end
         end
       end
     end
@@ -96,6 +111,88 @@ 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

@@ -29,7 +29,7 @@ module RubyLLM
     end
 
     def handle_stream(&block)
-      to_json_stream do |data|
+      build_on_data_handler do |data|
         block.call(build_chunk(data)) if data
       end
     end
@@ -40,19 +40,15 @@ 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, &)
@@ -67,22 +63,6 @@ module RubyLLM
       end
     end
 
-    def legacy_stream_processor(parser, &block)
-      proc do |chunk, _size|
-        process_stream_chunk(chunk, parser, nil, &block)
-      end
-    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
-    end
-
     def error_chunk?(chunk)
       chunk.start_with?('event: error')
     end
@@ -92,30 +72,18 @@ module RubyLLM
     end
 
     def handle_json_error_chunk(chunk, env)
-      parsed_data = JSON.parse(chunk)
-      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)
-    rescue JSON::ParserError => e
-      RubyLLM.logger.debug "Failed to parse JSON error chunk: #{e.message}"
+      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: ')
-      parsed_data = JSON.parse(error_data)
-      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)
-    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)
-      status, _message = parse_streaming_error(error_data.to_json)
-      error_response = env.merge(body: error_data, status: status || env.status)
-      ErrorMiddleware.parse_error(provider: self, response: error_response)
+      handle_parsed_error(error_data, env)
     rescue JSON::ParserError
       RubyLLM.logger.debug "Accumulating error chunk: #{chunk}"
     end
@@ -135,20 +103,13 @@ module RubyLLM
       parsed = JSON.parse(data)
       return parsed unless parsed.is_a?(Hash) && parsed.key?('error')
 
-      status, _message = parse_streaming_error(parsed.to_json)
-      error_response = build_stream_error_response(parsed, env, status)
-      ErrorMiddleware.parse_error(provider: self, response: error_response)
+      handle_parsed_error(parsed, env)
     rescue JSON::ParserError => e
       RubyLLM.logger.debug "Failed to parse data chunk: #{e.message}"
     end
 
     def handle_error_event(data, env)
-      parsed_data = JSON.parse(data)
-      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)
-    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)
@@ -159,6 +120,19 @@ module RubyLLM
       [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
 
@@ -168,5 +142,34 @@ module RubyLLM
         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