Class: DSPy::Teleprompt::GEPA::GeneticEngine

Inherits:
Object
  • Object
show all
Extended by:
T::Sig
Defined in:
lib/dspy/teleprompt/gepa.rb

Overview

GeneticEngine orchestrates the genetic algorithm for prompt evolution Manages population, selection, and evolution across generations

Instance Attribute Summary collapse

Instance Method Summary collapse

Constructor Details

#initialize(config:, fitness_evaluator:) ⇒ GeneticEngine

Returns a new instance of GeneticEngine.



1035
1036
1037
1038
1039
1040
1041
 
# File 'lib/dspy/teleprompt/gepa.rb', line 1035

def initialize(config:, fitness_evaluator:)
  @config = config
  @fitness_evaluator = fitness_evaluator
  @population = T.let([], T::Array[T.untyped])
  @generation = 0
  @fitness_scores = T.let([], T::Array[FitnessScore])
end

Instance Attribute Details

#config ⇒ Object (readonly)

Returns the value of attribute config.



1023
1024
1025
 
# File 'lib/dspy/teleprompt/gepa.rb', line 1023

def config
  @config
end

#fitness_evaluator ⇒ Object (readonly)

Returns the value of attribute fitness_evaluator.



1026
1027
1028
 
# File 'lib/dspy/teleprompt/gepa.rb', line 1026

def fitness_evaluator
  @fitness_evaluator
end

#generation ⇒ Object (readonly)

Returns the value of attribute generation.



1032
1033
1034
 
# File 'lib/dspy/teleprompt/gepa.rb', line 1032

def generation
  @generation
end

#population ⇒ Object (readonly)

Returns the value of attribute population.



1029
1030
1031
 
# File 'lib/dspy/teleprompt/gepa.rb', line 1029

def population
  @population
end

Instance Method Details

#evaluate_population(trainset) ⇒ Object 



1092
1093
1094
1095
1096
1097
1098
 
# File 'lib/dspy/teleprompt/gepa.rb', line 1092

def evaluate_population(trainset)
  @fitness_scores = @population.map do |candidate|
    @fitness_evaluator.evaluate_candidate(candidate, trainset)
  end

  @fitness_scores
end

#evolve_generation(trainset) ⇒ Object 



1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
 
# File 'lib/dspy/teleprompt/gepa.rb', line 1102

def evolve_generation(trainset)
  current_scores = evaluate_population(trainset)

  # Simple selection: keep top 50% and mutate them
  sorted_indices = (0...@population.size).sort_by { |i| -current_scores[i].overall_score }
  survivors = sorted_indices.take([@config.population_size / 2, 1].max)

  new_population = []

  # Keep best performers
  survivors.each { |i| new_population << @population[i] }

  # Fill rest with mutations of survivors
  while new_population.size < @config.population_size
    parent_index = survivors.sample
    parent = @population[parent_index]

    # Generate mutation if parent has signature_class
    if parent.respond_to?(:signature_class) && parent.signature_class.respond_to?(:description)
      variants = generate_instruction_variants(parent.signature_class.description)
      mutated = create_program_with_instruction(parent, variants.first || parent.signature_class.description)
      new_population << mutated
    else
      # If no signature_class, just duplicate the parent
      new_population << parent
    end
  end

  @population = new_population
  @generation += 1
end

#get_best_candidate ⇒ Object 



1184
1185
1186
1187
1188
1189
 
# File 'lib/dspy/teleprompt/gepa.rb', line 1184

def get_best_candidate
  return @population.first if @fitness_scores.empty?

  best_index = @fitness_scores.each_with_index.max_by { |score, _| score.overall_score }[1]
  @population[best_index]
end

#initialize_population(program) ⇒ Object 



1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
 
# File 'lib/dspy/teleprompt/gepa.rb', line 1045

def initialize_population(program)
  @population = []

  # Start with original program
  @population << program

  # Generate instruction variants to fill population if program has signature_class
  if program.respond_to?(:signature_class) && program.signature_class.respond_to?(:description)
    original_instruction = program.signature_class.description
    if original_instruction && !original_instruction.empty?
      variants = generate_instruction_variants(original_instruction)
    else
      variants = []
    end
  else
    variants = []
  end

  # Create program copies with different instructions
  variants.take(@config.population_size - 1).each do |variant|
    variant_program = create_program_with_instruction(program, variant)
    @population << variant_program
  end

  # If we need more candidates, duplicate and mutate
  while @population.size < @config.population_size
    base_program = @population.sample
    if base_program.respond_to?(:signature_class) && base_program.signature_class.respond_to?(:description)
      instruction_variants = generate_instruction_variants(base_program.signature_class.description)
      if instruction_variants.any?
        mutated = create_program_with_instruction(base_program, instruction_variants.first)
        @population << mutated
      else
        # If no variants available, just duplicate the base program
        @population << base_program
      end
    else
      # If no signature_class available, just duplicate the base program
      @population << base_program
    end
  end

  @generation = 0
end

#population_diversity ⇒ Object 



1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
 
# File 'lib/dspy/teleprompt/gepa.rb', line 1193

def population_diversity
  return 0.0 if @population.empty?

  # Only calculate diversity for programs that have signature_class
  instructions = @population.filter_map do |program|
    if program.respond_to?(:signature_class) && program.signature_class.respond_to?(:description)
      program.signature_class.description
    else
      nil
    end
  end

  return 0.0 if instructions.empty?

  unique_instructions = instructions.uniq.size
  unique_instructions.to_f / instructions.size.to_f
end

#run_evolution(program, trainset) ⇒ Object 



1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
 
# File 'lib/dspy/teleprompt/gepa.rb', line 1136

def run_evolution(program, trainset)
  initialize_population(program)

  history = []

  # Initial evaluation
  initial_scores = evaluate_population(trainset)
  best_initial = initial_scores.max_by(&:overall_score)
  avg_initial = initial_scores.map(&:overall_score).sum / initial_scores.size
  history << {
    generation: 0,
    best_fitness: best_initial.overall_score,
    avg_fitness: avg_initial,
    diversity: population_diversity
  }

  # Evolution loop
  @config.num_generations.times do
    evolve_generation(trainset)
    scores = evaluate_population(trainset)
    best_score = scores.max_by(&:overall_score)
    avg_score = scores.map(&:overall_score).sum / scores.size

    history << {
      generation: @generation,
      best_fitness: best_score.overall_score,
      avg_fitness: avg_score,
      diversity: population_diversity
    }
  end

  best_fitness_score = @fitness_scores.max_by(&:overall_score)
  {
    best_candidate: get_best_candidate,
    best_fitness: best_fitness_score || FitnessScore.new(
      primary_score: 0.0,
      secondary_scores: {},
      overall_score: 0.0,
      metadata: {}
    ),
    generation_history: history,
    generation_count: @generation,
    final_population: @population.dup
  }
end