dspy 0.28.1 → 0.29.0

data/lib/dspy/propose/grounded_proposer.rb

@@ -11,40 +11,80 @@ module DSPy
     class GroundedProposer
       extend T::Sig
 
-      # Configuration for instruction proposal
+      MAX_HISTORY_INSTRUCTIONS = 5
+
+      # Python-compatible TIPS dictionary for instruction generation
+      TIPS = {
+        "none" => "",
+        "creative" => "Don't be afraid to be creative when creating the new instruction!",
+        "simple" => "Keep the instruction clear and concise.",
+        "description" => "Make sure your instruction is very informative and descriptive.",
+        "high_stakes" => "The instruction should include a high stakes scenario in which the LM must solve the task!",
+        "persona" => 'Include a persona that is relevant to the task in the instruction (ie. "You are a ...")'
+      }.freeze
+
+      # Configuration for instruction proposal (Python-compatible)
       class Config
         extend T::Sig
 
+        # Core parameters
         sig { returns(Integer) }
         attr_accessor :num_instruction_candidates
 
+        # Python-compatible awareness flags (match Python defaults exactly)
+        sig { returns(T::Boolean) }
+        attr_accessor :program_aware
+
+        sig { returns(T::Boolean) }
+        attr_accessor :use_dataset_summary
+
+        sig { returns(T::Boolean) }
+        attr_accessor :use_task_demos
+
+        sig { returns(T::Boolean) }
+        attr_accessor :use_tip
+
+        sig { returns(T::Boolean) }
+        attr_accessor :use_instruct_history
+
+        # Additional parameters
         sig { returns(Integer) }
-        attr_accessor :max_examples_for_analysis
+        attr_accessor :view_data_batch_size
 
         sig { returns(Integer) }
-        attr_accessor :max_instruction_length
+        attr_accessor :num_demos_in_context
 
         sig { returns(T::Boolean) }
-        attr_accessor :use_task_description
+        attr_accessor :set_tip_randomly
 
         sig { returns(T::Boolean) }
-        attr_accessor :use_input_output_analysis
+        attr_accessor :set_history_randomly
 
-        sig { returns(T::Boolean) }
-        attr_accessor :use_few_shot_examples
+        sig { returns(Float) }
+        attr_accessor :init_temperature
 
-        sig { returns(String) }
-        attr_accessor :proposal_model
+        sig { returns(T::Boolean) }
+        attr_accessor :verbose
 
         sig { void }
         def initialize
+          # Core parameters
           @num_instruction_candidates = 5
-          @max_examples_for_analysis = 10
-          @max_instruction_length = 200
-          @use_task_description = true
-          @use_input_output_analysis = true
-          @use_few_shot_examples = true
-          @proposal_model = "gpt-4o-mini"
+
+          # Python-compatible awareness flags (match Python defaults)
+          @program_aware = true
+          @use_dataset_summary = true
+          @use_task_demos = true
+          @use_tip = true
+          @use_instruct_history = true
+
+          # Additional parameters
+          @view_data_batch_size = 10
+          @num_demos_in_context = 3
+          @set_tip_randomly = true
+          @set_history_randomly = true
+          @init_temperature = 1.0
+          @verbose = false
         end
       end
 
@@ -55,6 +95,9 @@ module DSPy
         sig { returns(T::Array[String]) }
         attr_reader :candidate_instructions
 
+        sig { returns(T::Hash[Integer, T::Array[String]]) }
+        attr_reader :predictor_instructions
+
         sig { returns(T::Hash[Symbol, T.untyped]) }
         attr_reader :analysis
 
@@ -65,11 +108,16 @@ module DSPy
           params(
             candidate_instructions: T::Array[String],
             analysis: T::Hash[Symbol, T.untyped],
-            metadata: T::Hash[Symbol, T.untyped]
+            metadata: T::Hash[Symbol, T.untyped],
+            predictor_instructions: T.nilable(T::Hash[Integer, T::Array[String]])
           ).void
         end
-        def initialize(candidate_instructions:, analysis:, metadata:)
+        def initialize(candidate_instructions:, analysis:, metadata:, predictor_instructions: nil)
           @candidate_instructions = candidate_instructions.freeze
+        normalized_predictor_instructions = (predictor_instructions || {}).each_with_object({}) do |(index, instructions), memo|
+            memo[index] = instructions.dup.freeze
+          end
+          @predictor_instructions = normalized_predictor_instructions.freeze
           @analysis = analysis.freeze
           @metadata = metadata.freeze
         end
@@ -88,21 +136,77 @@ module DSPy
       sig { returns(Config) }
       attr_reader :config
 
-      sig { params(config: T.nilable(Config)).void }
-      def initialize(config: nil)
+      sig do
+        params(
+          config: T.nilable(Config),
+          program: T.nilable(T.untyped),
+          trainset: T.nilable(T::Array[DSPy::Example])
+        ).void
+      end
+      def initialize(config: nil, program: nil, trainset: nil)
         @config = config || Config.new
+        @program = program
+        @trainset = trainset
+        @dataset_summary = nil
+        @program_code_string = nil
+
+        # Generate dataset summary if data-aware mode enabled (Python: use_dataset_summary)
+        if @config.use_dataset_summary && trainset && !trainset.empty?
+          begin
+            require_relative 'dataset_summary_generator'
+            @dataset_summary = DatasetSummaryGenerator.create_dataset_summary(
+              trainset,
+              @config.view_data_batch_size,
+              DSPy.current_lm,
+              verbose: @config.verbose
+            )
+          rescue => e
+            DSPy.logger.warn("Failed to generate dataset summary: #{e.message}")
+            @dataset_summary = nil
+          end
+        end
+
+        # Extract program source code if program-aware mode enabled
+        if @config.program_aware && program
+          @program_code_string = extract_program_source(program)
+        end
+      end
+
+      private
+
+      # Extract source code from program for program-aware mode
+      sig { params(program: T.untyped).returns(T.nilable(String)) }
+      def extract_program_source(program)
+        # Get the program's class
+        klass = program.is_a?(Class) ? program : program.class
+
+        # Try to get source location
+        source_location = klass.instance_method(:forward).source_location rescue nil
+        return nil unless source_location
+
+        file, line = source_location
+        # Read the source file and extract the class definition
+        # This is a simplified version - could be enhanced with method_source gem
+        code = "Program: #{klass.name}\nSource: #{file}:#{line}"
+        code
+      rescue => e
+        DSPy.logger.warn("Could not extract program source: #{e.message}")
+        nil
       end
 
+      public
+
       # Generate instruction candidates for a signature and training examples
       sig do
         params(
           signature_class: T.class_of(DSPy::Signature),
           examples: T::Array[T.untyped],
           few_shot_examples: T.nilable(T::Array[T.untyped]),
-          current_instruction: T.nilable(String)
+          current_instruction: T.nilable(String),
+          trial_logs: T.nilable(T::Hash[Integer, T::Hash[Symbol, T.untyped]])
         ).returns(ProposalResult)
       end
-      def propose_instructions(signature_class, examples, few_shot_examples: nil, current_instruction: nil)
+      def propose_instructions(signature_class, examples, few_shot_examples: nil, current_instruction: nil, trial_logs: nil)
         DSPy::Context.with_span(
           operation: 'optimization.instruction_proposal',
           'dspy.module' => 'GroundedProposer',
@@ -116,9 +220,11 @@ module DSPy
           
           # Generate instruction candidates
           candidates = generate_instruction_candidates(
-            signature_class, 
-            analysis, 
-            current_instruction
+            signature_class,
+            analysis,
+            current_instruction,
+            few_shot_examples: few_shot_examples,
+            trial_logs: trial_logs
           )
 
           # Filter and rank candidates
@@ -126,8 +232,8 @@ module DSPy
 
           metadata = {
             generation_timestamp: Time.now.iso8601,
-            model_used: @config.proposal_model,
-            num_examples_analyzed: [examples.size, @config.max_examples_for_analysis].min,
+            model_used: DSPy.current_lm.model,
+            num_examples_analyzed: [examples.size, @config.view_data_batch_size].min,
             original_instruction: current_instruction
           }
 
@@ -142,6 +248,50 @@ module DSPy
         end
       end
 
+      sig do
+        params(
+          trainset: T::Array[T.untyped],
+          program: T.untyped,
+          demo_candidates: T::Hash[Integer, T::Array[T::Array[DSPy::FewShotExample]]],
+          trial_logs: T.nilable(T::Hash[Integer, T::Hash[Symbol, T.untyped]]),
+          num_instruction_candidates: T.nilable(Integer)
+        ).returns(ProposalResult)
+      end
+      def propose_instructions_for_program(trainset:, program:, demo_candidates:, trial_logs: nil, num_instruction_candidates: nil)
+        num_candidates = num_instruction_candidates || @config.num_instruction_candidates
+
+        current_instruction = if program.respond_to?(:prompt) && program.prompt.respond_to?(:instruction)
+          program.prompt.instruction
+        else
+          nil
+        end
+
+        few_shot_examples = demo_candidates[0]&.flatten&.take(@config.num_demos_in_context) || []
+
+        signature_class = if program.respond_to?(:signature_class)
+          program.signature_class
+        else
+          raise ArgumentError, "Program must expose signature_class for instruction proposal"
+        end
+
+        base_result = propose_instructions(
+          signature_class,
+          trainset,
+          few_shot_examples: few_shot_examples,
+          current_instruction: current_instruction,
+          trial_logs: trial_logs
+        )
+
+        predictor_instructions = { 0 => base_result.candidate_instructions.take(num_candidates) }
+
+        ProposalResult.new(
+          candidate_instructions: base_result.candidate_instructions,
+          analysis: base_result.analysis,
+          metadata: base_result.metadata,
+          predictor_instructions: predictor_instructions
+        )
+      end
+
       private
 
       # Analyze the task based on signature and training examples
@@ -204,7 +354,7 @@ module DSPy
       # Analyze patterns in training examples
       sig { params(examples: T::Array[T.untyped]).returns(T::Hash[Symbol, T.untyped]) }
       def analyze_example_patterns(examples)
-        analysis_examples = examples.take(@config.max_examples_for_analysis)
+        analysis_examples = examples.take(@config.view_data_batch_size)
         
         {
           total_examples: examples.size,
@@ -323,12 +473,20 @@ module DSPy
         params(
           signature_class: T.class_of(DSPy::Signature),
           analysis: T::Hash[Symbol, T.untyped],
-          current_instruction: T.nilable(String)
+          current_instruction: T.nilable(String),
+          few_shot_examples: T.nilable(T::Array[T.untyped]),
+          trial_logs: T.nilable(T::Hash[Integer, T::Hash[Symbol, T.untyped]])
         ).returns(T::Array[String])
       end
-      def generate_instruction_candidates(signature_class, analysis, current_instruction)
+      def generate_instruction_candidates(signature_class, analysis, current_instruction, few_shot_examples: nil, trial_logs: nil)
         # Build context for instruction generation
-        context = build_generation_context(signature_class, analysis, current_instruction)
+        context = build_generation_context(
+          signature_class,
+          analysis,
+          current_instruction,
+          few_shot_examples: few_shot_examples,
+          trial_logs: trial_logs
+        )
         
         # Create instruction generation signature
         instruction_signature = create_instruction_generation_signature
@@ -346,16 +504,7 @@ module DSPy
             )
             
             instruction = result.instruction.strip
-            
-            # Truncate if too long
-            if instruction.length > @config.max_instruction_length
-              instruction = instruction[0, @config.max_instruction_length].strip
-              # Try to end at a word boundary
-              if instruction.include?(' ')
-                instruction = instruction.rpartition(' ').first + '.'
-              end
-            end
-            
+
             candidates << instruction if instruction.length > 0
           rescue => error
             DSPy.logger.warn("Failed to generate instruction candidate #{i + 1}: #{error.message}")
@@ -375,32 +524,64 @@ module DSPy
         params(
           signature_class: T.class_of(DSPy::Signature),
           analysis: T::Hash[Symbol, T.untyped],
-          current_instruction: T.nilable(String)
+          current_instruction: T.nilable(String),
+          few_shot_examples: T.nilable(T::Array[T.untyped]),
+          trial_logs: T.nilable(T::Hash[Integer, T::Hash[Symbol, T.untyped]])
         ).returns(String)
       end
-      def build_generation_context(signature_class, analysis, current_instruction)
+      def build_generation_context(signature_class, analysis, current_instruction, few_shot_examples: nil, trial_logs: nil)
         context_parts = []
-        
-        context_parts << "Task: #{signature_class.description}" if @config.use_task_description
-        
-        if @config.use_input_output_analysis
-          # Build detailed field descriptions including enum values
-          input_descriptions = analysis[:input_fields].map { |f| format_field_description(f) }
-          output_descriptions = analysis[:output_fields].map { |f| format_field_description(f) }
-          
-          context_parts << "Input fields: #{input_descriptions.join(', ')}"
-          context_parts << "Output fields: #{output_descriptions.join(', ')}"
+
+        # Include dataset summary if enabled and available
+        if @config.use_dataset_summary && @dataset_summary
+          context_parts << "Dataset Summary: #{@dataset_summary}"
         end
-        
+
+        # Include program code if enabled and available
+        if @config.program_aware && @program_code_string
+          context_parts << "Program Code:\n#{@program_code_string}"
+        end
+
+        # Always include task description (fundamental to understanding the task)
+        context_parts << "Task: #{signature_class.description}"
+
+        # Always include field analysis (fundamental to understanding inputs/outputs)
+        input_descriptions = analysis[:input_fields].map { |f| format_field_description(f) }
+        output_descriptions = analysis[:output_fields].map { |f| format_field_description(f) }
+
+        context_parts << "Input fields: #{input_descriptions.join(', ')}"
+        context_parts << "Output fields: #{output_descriptions.join(', ')}"
+
+        # Include task demos if enabled and available
+        if @config.use_task_demos && few_shot_examples && !few_shot_examples.empty?
+          demo_strings = few_shot_examples.take(@config.num_demos_in_context).map do |example|
+            format_example_as_demo(example)
+          end
+          context_parts << "Task Demos:\n#{demo_strings.join("\n\n")}"
+        end
+
         if analysis[:common_themes] && analysis[:common_themes].any?
           context_parts << "Task themes: #{analysis[:common_themes].join(', ')}"
         end
-        
+
         if current_instruction
           context_parts << "Current instruction: \"#{current_instruction}\""
         end
-        
-        context_parts.join("\n")
+
+        # Include tip if enabled
+        if @config.use_tip
+          tip = select_tip
+          context_parts << "Tip: #{tip}" if tip && !tip.empty?
+        end
+
+        if @config.use_instruct_history
+          history_summary = build_instruction_history_summary(trial_logs, predictor_index: 0, top_n: MAX_HISTORY_INSTRUCTIONS)
+          unless history_summary.empty?
+            context_parts << "Previous instructions:\n#{history_summary}"
+          end
+        end
+
+        context_parts.join("\n\n")
       end
 
       # Format field description with enum values if applicable
@@ -414,6 +595,83 @@ module DSPy
         end
       end
 
+      # Format an example as a demo for context
+      sig { params(example: T.untyped).returns(String) }
+      def format_example_as_demo(example)
+        return example.to_s unless example.respond_to?(:inputs) && example.respond_to?(:expected)
+
+        parts = []
+
+        # Format inputs
+        if example.inputs && !example.inputs.empty?
+          input_strs = example.inputs.map { |k, v| "#{k}: #{v.inspect}" }
+          parts << "Inputs: #{input_strs.join(', ')}"
+        end
+
+        # Format expected outputs
+        if example.expected && !example.expected.empty?
+          output_strs = example.expected.map { |k, v| "#{k}: #{v.inspect}" }
+          parts << "Expected: #{output_strs.join(', ')}"
+        end
+
+        parts.join(" | ")
+      end
+
+      # Select a tip based on configuration
+      sig { returns(T.nilable(String)) }
+      def select_tip
+        if @config.set_tip_randomly
+          # Randomly select a tip (excluding "none")
+          tip_keys = TIPS.keys.reject { |k| k == "none" }
+          selected_key = tip_keys.sample
+          TIPS[selected_key]
+        else
+          # Return empty string when not using random tips
+          ""
+        end
+      end
+
+      sig do
+        params(
+          trial_logs: T.nilable(T::Hash[Integer, T::Hash[Symbol, T.untyped]]),
+          predictor_index: Integer,
+          top_n: Integer
+        ).returns(String)
+      end
+      def build_instruction_history_summary(trial_logs, predictor_index:, top_n:)
+        return "" unless @config.use_instruct_history
+
+        logs = trial_logs || {}
+        aggregate = Hash.new { |hash, key| hash[key] = { total: 0.0, count: 0 } }
+
+        logs.each_value do |entry|
+          score = entry[:score]
+          next unless score.respond_to?(:to_f)
+
+          instructions = entry[:instructions]
+          instruction = nil
+          if instructions.respond_to?(:[])
+            instruction = instructions[predictor_index] || instructions[:default]
+          end
+          instruction ||= entry[:instruction]
+
+          next unless instruction.is_a?(String) && !instruction.empty?
+
+          aggregate[instruction][:total] += score.to_f
+          aggregate[instruction][:count] += 1
+        end
+
+        return "" if aggregate.empty?
+
+        ranked = aggregate.map do |instruction, stats|
+          average = stats[:total] / stats[:count]
+          [instruction, average]
+        end
+
+        top_entries = ranked.sort_by { |(_, avg)| -avg }.take(top_n).reverse
+        top_entries.map { |instruction, avg| format("%s | Score: %.4f", instruction, avg) }.join("\n\n")
+      end
+
       # Build requirements text for instruction generation
       sig { params(analysis: T::Hash[Symbol, T.untyped]).returns(String) }
       def build_requirements_text(analysis)
@@ -478,25 +736,21 @@ module DSPy
         # Filter out duplicates and empty candidates
         filtered = candidates.uniq.reject(&:empty?)
         
-        # Simple ranking based on length and content quality
+        # Simple ranking based on content quality (Python-compatible: no length scoring)
         filtered.sort_by do |instruction|
           score = 0
-          
-          # Prefer moderate length instructions
-          length_score = [instruction.length, @config.max_instruction_length].min / @config.max_instruction_length.to_f
-          score += length_score * 0.3
-          
+
           # Prefer instructions with action words
           action_words = %w[analyze classify generate explain solve determine identify]
           action_score = action_words.count { |word| instruction.downcase.include?(word) }
           score += action_score * 0.4
-          
+
           # Prefer instructions that mention reasoning for complex tasks
           if analysis[:complexity_indicators][:requires_reasoning]
             reasoning_score = instruction.downcase.match?(/\b(step|think|reason|explain)\b/) ? 1 : 0
             score += reasoning_score * 0.3
           end
-          
+
           -score # Negative for descending sort
         end
       end
@@ -588,9 +842,9 @@ module DSPy
           'proposal.num_candidates' => result.num_candidates,
           'proposal.best_instruction_length' => result.best_instruction.length,
           'proposal.analysis_themes' => result.analysis[:common_themes] || [],
-          'proposal.model_used' => @config.proposal_model
+          'proposal.model_used' => DSPy.current_lm.model
         })
       end
     end
   end
-end
+end

data/lib/dspy/re_act.rb

@@ -144,6 +144,19 @@ module DSPy
       super(enhanced_signature)
     end
 
+    sig { override.returns(T::Array[[String, DSPy::Module]]) }
+    def named_predictors
+      pairs = T.let([], T::Array[[String, DSPy::Module]])
+      pairs << ["thought_generator", @thought_generator]
+      pairs << ["observation_processor", @observation_processor]
+      pairs
+    end
+
+    sig { override.returns(T::Array[DSPy::Module]) }
+    def predictors
+      named_predictors.map { |(_, predictor)| predictor }
+    end
+
     sig { params(kwargs: T.untyped).returns(T.untyped).override }
     def forward(**kwargs)
       # Validate input

data/lib/dspy/reflection_lm.rb

@@ -0,0 +1,36 @@
+# frozen_string_literal: true
+
+require 'sorbet-runtime'
+
+module DSPy
+  # Lightweight wrapper for running reflection prompts without structured outputs.
+  class ReflectionLM
+    extend T::Sig
+
+    sig do
+      params(
+        model_id: String,
+        api_key: T.nilable(String),
+        options: T.untyped
+      ).void
+    end
+    def initialize(model_id, api_key: nil, **options)
+      opts = options.each_with_object({}) do |(key, value), memo|
+        memo[key.to_sym] = value
+      end
+      opts[:api_key] = api_key if api_key
+      @lm = DSPy::LM.new(model_id, structured_outputs: false, schema_format: :json, **opts)
+    end
+
+    sig { params(prompt: String).returns(String) }
+    def call(prompt)
+      response = @lm.raw_chat([{ role: 'user', content: prompt }])
+      response.respond_to?(:content) ? response.content : response.to_s
+    end
+
+    sig { params(messages: T.nilable(T::Array[T::Hash[Symbol, String]]), block: T.nilable(T.proc.params(arg0: T.untyped).void)).returns(T.untyped) }
+    def raw_chat(messages = nil, &block)
+      @lm.raw_chat(messages, &block)
+    end
+  end
+end

data/lib/dspy/teleprompt/bootstrap_strategy.rb

@@ -0,0 +1,26 @@
+# frozen_string_literal: true
+
+require 'sorbet-runtime'
+
+module DSPy
+  module Teleprompt
+    # Bootstrap strategy enum for create_n_fewshot_demo_sets
+    # Provides type-safe alternatives to Python's magic number seeds
+    class BootstrapStrategy < T::Enum
+      enums do
+        # No demonstrations - zero-shot learning (Python seed = -3)
+        ZeroShot = new
+
+        # Labeled examples only - no bootstrap generation (Python seed = -2)
+        LabeledOnly = new
+
+        # Bootstrapped demonstrations without shuffling (Python seed = -1)
+        Unshuffled = new
+
+        # Bootstrapped demonstrations with shuffling and random size (Python seed >= 0)
+        # Requires separate seed parameter for reproducibility
+        Shuffled = new
+      end
+    end
+  end
+end