dspy 0.28.1 → 0.29.0

data/lib/dspy/propose/dataset_summary_generator.rb

@@ -0,0 +1,210 @@
+# frozen_string_literal: true
+
+require 'sorbet-runtime'
+require 'json'
+require_relative '../signature'
+require_relative '../predict'
+require_relative '../type_serializer'
+require_relative '../few_shot_example'
+
+module DSPy
+  module Propose
+    # Dataset Summary Generator for creating concise dataset descriptions
+    # Used by GroundedProposer for data-aware instruction generation
+    module DatasetSummaryGenerator
+      extend T::Sig
+
+      # Signature for summarizing observations into a brief summary
+      class ObservationSummarizer < DSPy::Signature
+        description "Given a series of observations I have made about my dataset, please summarize them into a brief 2-3 sentence summary which highlights only the most important details."
+
+        input do
+          const :observations, String, description: "Observations I have made about my dataset"
+        end
+
+        output do
+          const :summary, String, description: "Two to Three sentence summary of only the most significant highlights of my observations"
+        end
+      end
+
+      # Signature for generating initial dataset observations
+      class DatasetDescriptor < DSPy::Signature
+        description "Given several examples from a dataset please write observations about trends that hold for most or all of the samples. " \
+                   "Some areas you may consider in your observations: topics, content, syntax, conciceness, etc. " \
+                   "It will be useful to make an educated guess as to the nature of the task this dataset will enable. Don't be afraid to be creative"
+
+        input do
+          const :examples, String, description: "Sample data points from the dataset"
+        end
+
+        output do
+          const :observations, String, description: "Somethings that holds true for most or all of the data you observed"
+        end
+      end
+
+      # Signature for refining observations with prior context
+      class DatasetDescriptorWithPriorObservations < DSPy::Signature
+        description "Given several examples from a dataset please write observations about trends that hold for most or all of the samples. " \
+                   "I will also provide you with a few observations I have already made.  Please add your own observations or if you feel the observations are comprehensive say 'COMPLETE' " \
+                   "Some areas you may consider in your observations: topics, content, syntax, conciceness, etc. " \
+                   "It will be useful to make an educated guess as to the nature of the task this dataset will enable. Don't be afraid to be creative"
+
+        input do
+          const :examples, String, description: "Sample data points from the dataset"
+          const :prior_observations, String, description: "Some prior observations I made about the data"
+        end
+
+        output do
+          const :observations, String, description: "Somethings that holds true for most or all of the data you observed or COMPLETE if you have nothing to add"
+        end
+      end
+
+      # Helper function to ensure consistent ordering of input keys in string representations
+      # This helps with caching and consistent LLM prompts
+      sig { params(unordered_repr: String).returns(String) }
+      def self.order_input_keys_in_string(unordered_repr)
+        # Regex pattern to match the input keys structure
+        pattern = /input_keys=\{([^}]+)\}/
+
+        # Function to reorder keys
+        unordered_repr.gsub(pattern) do |match|
+          keys_str = Regexp.last_match(1)
+          # Split the keys, strip extra spaces, and sort them
+          keys = keys_str.split(',').map(&:strip).sort
+          # Format the sorted keys back into the expected structure
+          "input_keys={#{keys.join(', ')}}"
+        end
+      end
+
+      # Strip common prefixes from LLM outputs (e.g., "Answer:", "Output:")
+      sig { params(text: String).returns(String) }
+      def self.strip_prefix(text)
+        # Pattern matches up to 4 words followed by a colon
+        pattern = /^[\*\s]*(([\w'\-]+\s+){0,4}[\w'\-]+):\s*/
+        modified_text = text.gsub(pattern, '')
+        modified_text.strip.gsub(/^["']|["']$/, '')
+      end
+
+      # Generate a concise 2-3 sentence summary of a training dataset
+      # Used for data-aware instruction proposal in MIPROv2
+      #
+      # @param trainset [Array<DSPy::Example>] Training examples to summarize
+      # @param view_data_batch_size [Integer] Number of examples to process per batch
+      # @param prompt_model [DSPy::LM, nil] Language model to use (defaults to DSPy.lm)
+      # @param verbose [Boolean] Whether to print progress information
+      # @return [String] 2-3 sentence summary of the dataset characteristics
+      #
+      # @example Basic usage
+      #   summary = DatasetSummaryGenerator.create_dataset_summary(
+      #     trainset,
+      #     view_data_batch_size: 10,
+      #     prompt_model: DSPy::LM.new('gpt-4o-mini')
+      #   )
+      #
+      sig do
+        params(
+          trainset: T::Array[DSPy::Example],
+          view_data_batch_size: Integer,
+          prompt_model: T.nilable(DSPy::LM),
+          verbose: T::Boolean
+        ).returns(String)
+      end
+      def self.create_dataset_summary(trainset, view_data_batch_size, prompt_model, verbose: false)
+        if verbose
+          puts "\nBootstrapping dataset summary (this will be used to generate instructions)..."
+        end
+
+        # Use provided model or fall back to global LM
+        lm = prompt_model || DSPy.lm
+        raise ArgumentError, "No language model configured. Set prompt_model or DSPy.lm" unless lm
+
+        # Use provided LM in a block context
+        DSPy.with_lm(lm) do
+          # Initial observation from first batch
+          upper_lim = [trainset.length, view_data_batch_size].min
+          batch_examples = trainset[0...upper_lim]
+          predictor = DSPy::Predict.new(DatasetDescriptor)
+          examples_repr = format_examples_for_prompt(batch_examples)
+
+          observation = predictor.call(examples: examples_repr)
+          observations = observation.observations
+
+          # Iteratively refine observations with additional batches
+          skips = 0
+          max_calls = 10
+          calls = 0
+
+          begin
+            (view_data_batch_size...trainset.length).step(view_data_batch_size) do |b|
+              calls += 1
+              break if calls >= max_calls
+
+              puts "Processing batch starting at index #{b}" if verbose
+
+              upper_lim = [trainset.length, b + view_data_batch_size].min
+
+              predictor = DSPy::Predict.new(DatasetDescriptorWithPriorObservations)
+              batch_examples = trainset[b...upper_lim]
+              examples_repr = format_examples_for_prompt(batch_examples)
+
+              output = predictor.call(
+                prior_observations: observations,
+                examples: examples_repr
+              )
+
+              # Check if LLM indicates observations are complete
+              if output.observations.length >= 8 && output.observations[0...8].upcase == "COMPLETE"
+                skips += 1
+                break if skips >= 5
+                next
+              end
+
+              observations += output.observations
+            end
+          rescue => e
+            if verbose
+              puts "Error during observation refinement: #{e.message}. Using observations from past round for summary."
+            end
+          end
+
+          # Generate final summary from accumulated observations
+          predictor = DSPy::Predict.new(ObservationSummarizer)
+          summary = predictor.call(observations: observations)
+
+          if verbose
+            puts "\nGenerated summary: #{strip_prefix(summary.summary)}\n"
+          end
+
+          strip_prefix(summary.summary)
+        end
+      end
+
+      sig { params(examples: T::Array[T.untyped]).returns(String) }
+      def self.format_examples_for_prompt(examples)
+        serialized_examples = examples.map do |example|
+          case example
+          when DSPy::Example
+            {
+              signature: example.signature_class.name,
+              input: DSPy::TypeSerializer.serialize(example.input),
+              expected: DSPy::TypeSerializer.serialize(example.expected)
+            }
+          when DSPy::FewShotExample
+            base = {
+              input: example.input,
+              output: example.output
+            }
+            base[:reasoning] = example.reasoning if example.reasoning
+            base
+          when Hash
+            example
+          else
+            example.respond_to?(:to_h) ? example.to_h : { value: example }
+          end
+        end
+
+        JSON.pretty_generate(serialized_examples)
+      end
+    end
+  end
+end