dspy 0.2.0 → 0.3.1

data/lib/dspy/chain_of_thought.rb

@@ -1,22 +1,162 @@
+# typed: strict
 # frozen_string_literal: true
 
+require 'sorbet-runtime'
+require_relative 'predict'
+require_relative 'signature'
+require_relative 'instrumentation'
+
 module DSPy
   # Enhances prediction by encouraging step-by-step reasoning
-  # before providing a final answer.
+  # before providing a final answer using Sorbet signatures.
   class ChainOfThought < Predict
+    extend T::Sig
+
+    FieldDescriptor = DSPy::Signature::FieldDescriptor
 
+    sig { params(signature_class: T.class_of(DSPy::Signature)).void }
     def initialize(signature_class)
-      @signature_class = signature_class
-      chain_of_thought_schema = Dry::Schema.JSON do
-        required(:reasoning).
-          value(:string).
-          meta(description: "Reasoning: Let's think step by step in order to #{signature_class.description}")
+      @original_signature = signature_class
+
+      # Create enhanced output struct with reasoning
+      enhanced_output_struct = create_enhanced_output_struct(signature_class)
+
+      # Create enhanced signature class
+      enhanced_signature = Class.new(DSPy::Signature) do
+        # Set the description
+        description "#{signature_class.description} Think step by step."
+
+        # Use the same input struct and copy field descriptors
+        @input_struct_class = signature_class.input_struct_class
+        @input_field_descriptors = signature_class.instance_variable_get(:@input_field_descriptors) || {}
+
+        # Use the enhanced output struct and create field descriptors for it
+        @output_struct_class = enhanced_output_struct
+
+        # Create field descriptors for the enhanced output struct
+        @output_field_descriptors = {}
+
+        # Copy original output field descriptors
+        original_output_descriptors = signature_class.instance_variable_get(:@output_field_descriptors) || {}
+        @output_field_descriptors.merge!(original_output_descriptors)
+
+        # Add reasoning field descriptor (ChainOfThought always provides this)
+        @output_field_descriptors[:reasoning] = FieldDescriptor.new(String, "Step by step reasoning process")
+
+        class << self
+          attr_reader :input_struct_class, :output_struct_class
+        end
+      end
+
+      # Call parent constructor with enhanced signature
+      super(enhanced_signature)
+      @signature_class = enhanced_signature
+    end
+
+    # Override forward_untyped to add ChainOfThought-specific instrumentation
+    sig { override.params(input_values: T.untyped).returns(T.untyped) }
+    def forward_untyped(**input_values)
+      # Prepare instrumentation payload
+      input_fields = input_values.keys.map(&:to_s)
+      
+      # Instrument ChainOfThought lifecycle
+      result = Instrumentation.instrument('dspy.chain_of_thought', {
+        signature_class: @original_signature.name,
+        model: lm.model,
+        provider: lm.provider,
+        input_fields: input_fields
+      }) do
+        # Call parent prediction logic
+        prediction_result = super(**input_values)
+        
+        # Analyze reasoning if present
+        if prediction_result.respond_to?(:reasoning) && prediction_result.reasoning
+          reasoning_content = prediction_result.reasoning.to_s
+          reasoning_length = reasoning_content.length
+          reasoning_steps = count_reasoning_steps(reasoning_content)
+          
+          # Emit reasoning analysis event
+          Instrumentation.emit('dspy.chain_of_thought.reasoning_complete', {
+            signature_class: @original_signature.name,
+            reasoning_steps: reasoning_steps,
+            reasoning_length: reasoning_length,
+            has_reasoning: !reasoning_content.empty?
+          })
+        end
+        
+        prediction_result
+      end
+      
+      result
+    end
+
+    private
+
+    # Count reasoning steps by looking for step indicators
+    def count_reasoning_steps(reasoning_text)
+      return 0 if reasoning_text.nil? || reasoning_text.empty?
+      
+      # Look for common step patterns
+      step_patterns = [
+        /step \d+/i,
+        /\d+\./,
+        /first|second|third|then|next|finally/i,
+        /\n\s*-/
+      ]
+      
+      max_count = 0
+      step_patterns.each do |pattern|
+        count = reasoning_text.scan(pattern).length
+        max_count = [max_count, count].max
+      end
+      
+      # Fallback: count sentences if no clear steps
+      max_count > 0 ? max_count : reasoning_text.split(/[.!?]+/).reject(&:empty?).length
+    end
+
+    sig { params(signature_class: T.class_of(DSPy::Signature)).returns(T.class_of(T::Struct)) }
+    def create_enhanced_output_struct(signature_class)
+      # Get original output props
+      original_props = signature_class.output_struct_class.props
+
+      # Create new struct class with reasoning added
+      Class.new(T::Struct) do
+        # Add all original fields
+        original_props.each do |name, prop|
+          # Extract the type and other options
+          type = prop[:type]
+          options = prop.except(:type, :type_object, :accessor_key, :sensitivity, :redaction)
+
+          # Handle default values
+          if options[:default]
+            const name, type, default: options[:default]
+          elsif options[:factory]
+            const name, type, factory: options[:factory]
+          else
+            const name, type
+          end
+        end
+
+        # Add reasoning field (ChainOfThought always provides this)
+        const :reasoning, String
+
+        # Add to_h method to serialize the struct to a hash
+        define_method :to_h do
+          hash = {}
+
+          # Start with input values if available
+          if self.instance_variable_defined?(:@input_values)
+            hash.merge!(self.instance_variable_get(:@input_values))
+          end
+
+          # Then add output properties
+          self.class.props.keys.each do |key|
+            hash[key] = self.send(key)
+          end
+
+          hash
+        end
       end
-      @signature_class.output_schema = Dry::Schema.JSON(parent:
-                                                          [
-                                                            @signature_class.output_schema,
-                                                            chain_of_thought_schema
-                                                          ])
     end
   end
 end

data/lib/dspy/instrumentation/token_tracker.rb

@@ -0,0 +1,54 @@
+# frozen_string_literal: true
+
+module DSPy
+  module Instrumentation
+    # Utility for extracting token usage from different LM adapters
+    # Uses actual token counts from API responses for accuracy
+    module TokenTracker
+      extend self
+
+      # Extract actual token usage from API responses
+      def extract_token_usage(response, provider)
+        case provider.to_s.downcase
+        when 'openai'
+          extract_openai_tokens(response)
+        when 'anthropic'
+          extract_anthropic_tokens(response)
+        else
+          {} # No token information for other providers
+        end
+      end
+
+      private
+
+      def extract_openai_tokens(response)
+        return {} unless response&.usage
+
+        usage = response.usage
+        return {} unless usage.is_a?(Hash)
+        
+        {
+          tokens_input: usage[:prompt_tokens] || usage['prompt_tokens'],
+          tokens_output: usage[:completion_tokens] || usage['completion_tokens'],
+          tokens_total: usage[:total_tokens] || usage['total_tokens']
+        }
+      end
+
+      def extract_anthropic_tokens(response)
+        return {} unless response&.usage
+
+        usage = response.usage
+        return {} unless usage.is_a?(Hash)
+        
+        input_tokens = usage[:input_tokens] || usage['input_tokens'] || 0
+        output_tokens = usage[:output_tokens] || usage['output_tokens'] || 0
+        
+        {
+          tokens_input: input_tokens,
+          tokens_output: output_tokens,
+          tokens_total: input_tokens + output_tokens
+        }
+      end
+    end
+  end
+end

data/lib/dspy/instrumentation.rb

@@ -0,0 +1,113 @@
+# frozen_string_literal: true
+
+require 'dry-monitor'
+require 'dry-configurable'
+
+module DSPy
+  # Core instrumentation module using dry-monitor for event emission
+  # Provides extension points for logging, Langfuse, New Relic, and custom monitoring
+  module Instrumentation
+    # Get the current logger subscriber instance (lazy initialization)
+    def self.logger_subscriber
+      @logger_subscriber ||= begin
+        require_relative 'subscribers/logger_subscriber'
+        DSPy::Subscribers::LoggerSubscriber.new
+      end
+    end
+
+    def self.notifications
+      @notifications ||= Dry::Monitor::Notifications.new(:dspy).tap do |n|
+        # Register all DSPy events
+        n.register_event('dspy.lm.request')
+        n.register_event('dspy.lm.tokens')
+        n.register_event('dspy.lm.response.parsed')
+        n.register_event('dspy.predict')
+        n.register_event('dspy.predict.validation_error')
+        n.register_event('dspy.chain_of_thought')
+        n.register_event('dspy.chain_of_thought.reasoning_step')
+        n.register_event('dspy.react')
+        n.register_event('dspy.react.tool_call')
+        n.register_event('dspy.react.iteration_complete')
+        n.register_event('dspy.react.max_iterations')
+      end
+    end
+
+    # High-precision timing for performance tracking
+    def self.instrument(event_name, payload = {}, &block)
+      # If no block is given, return early
+      return unless block_given?
+
+      start_time = Process.clock_gettime(Process::CLOCK_MONOTONIC)
+      start_cpu = Process.clock_gettime(Process::CLOCK_PROCESS_CPUTIME_ID)
+
+      begin
+        result = yield
+
+        end_time = Process.clock_gettime(Process::CLOCK_MONOTONIC)
+        end_cpu = Process.clock_gettime(Process::CLOCK_PROCESS_CPUTIME_ID)
+
+        enhanced_payload = payload.merge(
+          duration_ms: ((end_time - start_time) * 1000).round(2),
+          cpu_time_ms: ((end_cpu - start_cpu) * 1000).round(2),
+          status: 'success',
+          timestamp: Time.now.iso8601
+        )
+
+        self.emit_event(event_name, enhanced_payload)
+        result
+      rescue => error
+        end_time = Process.clock_gettime(Process::CLOCK_MONOTONIC)
+        end_cpu = Process.clock_gettime(Process::CLOCK_PROCESS_CPUTIME_ID)
+
+        error_payload = payload.merge(
+          duration_ms: ((end_time - start_time) * 1000).round(2),
+          cpu_time_ms: ((end_cpu - start_cpu) * 1000).round(2),
+          status: 'error',
+          error_type: error.class.name,
+          error_message: error.message,
+          timestamp: Time.now.iso8601
+        )
+
+        self.emit_event(event_name, error_payload)
+        raise
+      end
+    end
+
+    # Emit event without timing (for discrete events)
+    def self.emit(event_name, payload = {})
+      enhanced_payload = payload.merge(
+        timestamp: Time.now.iso8601,
+        status: payload[:status] || 'success'
+      )
+
+      self.emit_event(event_name, enhanced_payload)
+    end
+
+    # Register additional events dynamically (useful for testing)
+    def self.register_event(event_name)
+      notifications.register_event(event_name)
+    end
+
+    # Subscribe to DSPy instrumentation events
+    def self.subscribe(event_pattern = nil, &block)
+      if event_pattern
+        notifications.subscribe(event_pattern, &block)
+      else
+        # Subscribe to all DSPy events
+        %w[dspy.lm.request dspy.lm.tokens dspy.lm.response.parsed dspy.predict dspy.predict.validation_error dspy.chain_of_thought dspy.chain_of_thought.reasoning_step dspy.react dspy.react.tool_call dspy.react.iteration_complete dspy.react.max_iterations].each do |event_name|
+          notifications.subscribe(event_name, &block)
+        end
+      end
+    end
+
+    def self.emit_event(event_name, payload)
+      # Ensure logger subscriber is initialized
+      logger_subscriber
+      notifications.instrument(event_name, payload)
+    end
+
+    def self.setup_subscribers
+      # Lazy initialization - will be created when first accessed
+    end
+  end
+end

data/lib/dspy/lm/adapter.rb

@@ -0,0 +1,41 @@
+# frozen_string_literal: true
+
+module DSPy
+  class LM
+    # Base adapter interface for all LM providers
+    class Adapter
+      attr_reader :model, :api_key
+
+      def initialize(model:, api_key:)
+        @model = model
+        @api_key = api_key
+        validate_configuration!
+      end
+
+      # Chat interface that all adapters must implement
+      # @param messages [Array<Hash>] Array of message hashes with :role and :content
+      # @param block [Proc] Optional streaming block
+      # @return [DSPy::LM::Response] Normalized response
+      def chat(messages:, &block)
+        raise NotImplementedError, "Subclasses must implement #chat method"
+      end
+
+      private
+
+      def validate_configuration!
+        raise ConfigurationError, "Model is required" if model.nil? || model.empty?
+        raise ConfigurationError, "API key is required" if api_key.nil? || api_key.empty?
+      end
+
+      # Helper method to normalize message format
+      def normalize_messages(messages)
+        messages.map do |msg|
+          {
+            role: msg[:role].to_s,
+            content: msg[:content].to_s
+          }
+        end
+      end
+    end
+  end
+end

data/lib/dspy/lm/adapter_factory.rb

@@ -0,0 +1,59 @@
+# frozen_string_literal: true
+
+module DSPy
+  class LM
+    # Factory for creating appropriate adapters based on model_id
+    class AdapterFactory
+      # Maps provider prefixes to adapter classes
+      ADAPTER_MAP = {
+        'openai' => 'OpenAIAdapter',
+        'anthropic' => 'AnthropicAdapter',
+        'ruby_llm' => 'RubyLLMAdapter'
+      }.freeze
+
+      class << self
+        # Creates an adapter instance based on model_id
+        # @param model_id [String] Full model identifier (e.g., "openai/gpt-4")
+        # @param api_key [String] API key for the provider
+        # @return [DSPy::LM::Adapter] Appropriate adapter instance
+        def create(model_id, api_key:)
+          provider, model = parse_model_id(model_id)
+          adapter_class = get_adapter_class(provider)
+          
+          adapter_class.new(model: model, api_key: api_key)
+        end
+
+        private
+
+        # Parse model_id to determine provider and model
+        def parse_model_id(model_id)
+          if model_id.include?('/')
+            provider, model = model_id.split('/', 2)
+            [provider, model]
+          else
+            # Legacy format: assume ruby_llm for backward compatibility
+            ['ruby_llm', model_id]
+          end
+        end
+
+        def get_adapter_class(provider)
+          adapter_class_name = ADAPTER_MAP[provider]
+          
+          unless adapter_class_name
+            available_providers = ADAPTER_MAP.keys.join(', ')
+            raise UnsupportedProviderError, 
+                  "Unsupported provider: #{provider}. Available: #{available_providers}"
+          end
+
+          begin
+            Object.const_get("DSPy::LM::#{adapter_class_name}")
+          rescue NameError
+            raise UnsupportedProviderError, 
+                  "Adapter not found: DSPy::LM::#{adapter_class_name}. " \
+                  "Make sure the corresponding gem is installed."
+          end
+        end
+      end
+    end
+  end
+end

data/lib/dspy/lm/adapters/anthropic_adapter.rb

@@ -0,0 +1,96 @@
+# frozen_string_literal: true
+
+require 'anthropic'
+
+module DSPy
+  class LM
+    class AnthropicAdapter < Adapter
+      def initialize(model:, api_key:)
+        super
+        @client = Anthropic::Client.new(api_key: api_key)
+      end
+
+      def chat(messages:, &block)
+        # Anthropic requires system message to be separate from messages
+        system_message, user_messages = extract_system_message(normalize_messages(messages))
+        
+        request_params = {
+          model: model,
+          messages: user_messages,
+          max_tokens: 4096, # Required for Anthropic
+          temperature: 0.0 # DSPy default for deterministic responses
+        }
+
+        # Add system message if present
+        request_params[:system] = system_message if system_message
+
+        # Add streaming if block provided
+        if block_given?
+          request_params[:stream] = true
+        end
+
+        begin
+          if block_given?
+            content = ""
+            @client.messages.stream(**request_params) do |chunk|
+              if chunk.respond_to?(:delta) && chunk.delta.respond_to?(:text)
+                chunk_text = chunk.delta.text
+                content += chunk_text
+                block.call(chunk)
+              end
+            end
+            
+            Response.new(
+              content: content,
+              usage: nil, # Usage not available in streaming
+              metadata: {
+                provider: 'anthropic',
+                model: model,
+                streaming: true
+              }
+            )
+          else
+            response = @client.messages.create(**request_params)
+            
+            if response.respond_to?(:error) && response.error
+              raise AdapterError, "Anthropic API error: #{response.error}"
+            end
+
+            content = response.content.first.text if response.content.is_a?(Array) && response.content.first
+            usage = response.usage
+
+            Response.new(
+              content: content,
+              usage: usage.respond_to?(:to_h) ? usage.to_h : usage,
+              metadata: {
+                provider: 'anthropic',
+                model: model,
+                response_id: response.id,
+                role: response.role
+              }
+            )
+          end
+        rescue => e
+          raise AdapterError, "Anthropic adapter error: #{e.message}"
+        end
+      end
+
+      private
+
+      def extract_system_message(messages)
+        system_message = nil
+        user_messages = []
+
+        messages.each do |msg|
+          if msg[:role] == 'system'
+            system_message = msg[:content]
+          else
+            user_messages << msg
+          end
+        end
+
+        [system_message, user_messages]
+      end
+    end
+  end
+end

data/lib/dspy/lm/adapters/openai_adapter.rb

@@ -0,0 +1,53 @@
+# frozen_string_literal: true
+
+require 'openai'
+
+module DSPy
+  class LM
+    class OpenAIAdapter < Adapter
+      def initialize(model:, api_key:)
+        super
+        @client = OpenAI::Client.new(api_key: api_key)
+      end
+
+      def chat(messages:, &block)
+        request_params = {
+          model: model,
+          messages: normalize_messages(messages),
+          temperature: 0.0 # DSPy default for deterministic responses
+        }
+
+        # Add streaming if block provided
+        if block_given?
+          request_params[:stream] = proc do |chunk, _bytesize|
+            block.call(chunk) if chunk.dig("choices", 0, "delta", "content")
+          end
+        end
+
+        begin
+          response = @client.chat.completions.create(**request_params)
+          
+          if response.respond_to?(:error) && response.error
+            raise AdapterError, "OpenAI API error: #{response.error}"
+          end
+
+          content = response.choices.first.message.content
+          usage = response.usage
+
+          Response.new(
+            content: content,
+            usage: usage.respond_to?(:to_h) ? usage.to_h : usage,
+            metadata: {
+              provider: 'openai',
+              model: model,
+              response_id: response.id,
+              created: response.created
+            }
+          )
+        rescue => e
+          raise AdapterError, "OpenAI adapter error: #{e.message}"
+        end
+      end
+    end
+  end
+end

data/lib/dspy/lm/adapters/ruby_llm_adapter.rb

@@ -0,0 +1,81 @@
+# frozen_string_literal: true
+
+begin
+  require 'ruby_llm'
+rescue LoadError
+  # ruby_llm is optional for backward compatibility
+end
+
+module DSPy
+  class LM
+    class RubyLLMAdapter < Adapter
+      def initialize(model:, api_key:)
+        super
+        
+        unless defined?(RubyLLM)
+          raise ConfigurationError, 
+                "ruby_llm gem is required for RubyLLMAdapter. " \
+                "Add 'gem \"ruby_llm\"' to your Gemfile."
+        end
+
+        configure_ruby_llm
+      end
+
+      def chat(messages:, &block)
+        begin
+          chat = RubyLLM.chat(model: model)
+          
+          # Add messages to chat
+          messages.each do |msg|
+            chat.add_message(role: msg[:role].to_sym, content: msg[:content])
+          end
+
+          # Get the last user message for ask method
+          last_user_message = messages.reverse.find { |msg| msg[:role] == 'user' }
+          
+          if last_user_message
+            # Remove the last user message since ask() will add it
+            chat.messages.pop if chat.messages.last&.content == last_user_message[:content]
+            chat.ask(last_user_message[:content], &block)
+          else
+            raise AdapterError, "No user message found in conversation"
+          end
+
+          content = chat.messages.last&.content || ""
+
+          Response.new(
+            content: content,
+            usage: nil, # ruby_llm doesn't provide usage info
+            metadata: {
+              provider: 'ruby_llm',
+              model: model,
+              message_count: chat.messages.length
+            }
+          )
+        rescue => e
+          raise AdapterError, "RubyLLM adapter error: #{e.message}"
+        end
+      end
+
+      private
+
+      def configure_ruby_llm
+        # Determine provider from model for configuration
+        if model.include?('gpt') || model.include?('openai')
+          RubyLLM.configure do |config|
+            config.openai_api_key = api_key
+          end
+        elsif model.include?('claude') || model.include?('anthropic')
+          RubyLLM.configure do |config|
+            config.anthropic_api_key = api_key
+          end
+        else
+          # Default to OpenAI configuration
+          RubyLLM.configure do |config|
+            config.openai_api_key = api_key
+          end
+        end
+      end
+    end
+  end
+end

data/lib/dspy/lm/errors.rb

@@ -0,0 +1,10 @@
+# frozen_string_literal: true
+
+module DSPy
+  class LM
+    class Error < StandardError; end
+    class AdapterError < Error; end
+    class UnsupportedProviderError < Error; end
+    class ConfigurationError < Error; end
+  end
+end