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:, metric:) ⇒ GeneticEngine

Returns a new instance of GeneticEngine.



1338
1339
1340
1341
1342
1343
1344
 
# File 'lib/dspy/teleprompt/gepa.rb', line 1338

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

Instance Attribute Details

#config ⇒ Object (readonly)

Returns the value of attribute config.



1326
1327
1328
 
# File 'lib/dspy/teleprompt/gepa.rb', line 1326

def config
  @config
end

#generation ⇒ Object (readonly)

Returns the value of attribute generation.



1335
1336
1337
 
# File 'lib/dspy/teleprompt/gepa.rb', line 1335

def generation
  @generation
end

#metric ⇒ Object (readonly)

Returns the value of attribute metric.



1329
1330
1331
 
# File 'lib/dspy/teleprompt/gepa.rb', line 1329

def metric
  @metric
end

#population ⇒ Object (readonly)

Returns the value of attribute population.



1332
1333
1334
 
# File 'lib/dspy/teleprompt/gepa.rb', line 1332

def population
  @population
end

Instance Method Details

#evaluate_population(trainset) ⇒ Object 



1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
 
# File 'lib/dspy/teleprompt/gepa.rb', line 1377

def evaluate_population(trainset)
  @fitness_scores = @population.map do |candidate|
    scores = trainset.map do |example|
      prediction = candidate.call(**example.input_values)
      @metric.call(example, prediction).to_f
    rescue => e
      # Handle evaluation errors gracefully
      0.0
    end
    
    scores.sum / scores.size
  end
  
  @fitness_scores
end

#evolve_generation(trainset) ⇒ Object 



1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
1419
1420
 
# File 'lib/dspy/teleprompt/gepa.rb', line 1395

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] }
  survivors = sorted_indices.take(@config.population_size / 2)
  
  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
    variants = generate_instruction_variants(parent.signature_class.description)
    mutated = create_program_with_instruction(parent, variants.first || parent.signature_class.description)
    new_population << mutated
  end
  
  @population = new_population
  @generation += 1
end

#get_best_candidate ⇒ Object 



1461
1462
1463
1464
1465
1466
 
# File 'lib/dspy/teleprompt/gepa.rb', line 1461

def get_best_candidate
  return @population.first if @fitness_scores.empty?
  
  best_index = @fitness_scores.each_with_index.max_by { |score, _| score }[1]
  @population[best_index]
end

#initialize_population(program) ⇒ Object 



1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
 
# File 'lib/dspy/teleprompt/gepa.rb', line 1348

def initialize_population(program)
  @population = []
  
  # Start with original program
  @population << program
  
  # Generate instruction variants to fill population
  original_instruction = program.signature_class.description
  variants = generate_instruction_variants(original_instruction)
  
  # 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
    mutated = create_program_with_instruction(base_program, 
      generate_instruction_variants(base_program.signature_class.description).first)
    @population << mutated
  end
  
  @generation = 0
end

#population_diversity ⇒ Object 



1470
1471
1472
1473
1474
1475
1476
1477
 
# File 'lib/dspy/teleprompt/gepa.rb', line 1470

def population_diversity
  return 0.0 if @population.empty?
  
  instructions = @population.map(&:signature_class).map(&:description)
  unique_instructions = instructions.uniq.size
  
  unique_instructions.to_f / @population.size.to_f
end

#run_evolution(program, trainset) ⇒ Object 



1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
 
# File 'lib/dspy/teleprompt/gepa.rb', line 1424

def run_evolution(program, trainset)
  initialize_population(program)
  
  history = []
  
  # Initial evaluation
  initial_scores = evaluate_population(trainset)
  history << {
    generation: 0,
    best_fitness: initial_scores.max,
    avg_fitness: initial_scores.sum / initial_scores.size,
    diversity: population_diversity
  }
  
  # Evolution loop
  @config.num_generations.times do
    evolve_generation(trainset)
    scores = evaluate_population(trainset)
    
    history << {
      generation: @generation,
      best_fitness: scores.max,
      avg_fitness: scores.sum / scores.size,
      diversity: population_diversity
    }
  end
  
  {
    best_candidate: get_best_candidate,
    best_fitness: @fitness_scores.max,
    generation_history: history,
    final_population: @population.dup
  }
end