pip install synth-ai

{
  "tasks": [
    {
      "task_id": "q1",
      "input": {
        "question": "What is the capital of France?",
        "context": "France is a country in Western Europe."
      }
    },
    {
      "task_id": "q2",
      "input": {
        "question": "Who wrote Romeo and Juliet?",
        "context": "Romeo and Juliet is a famous tragedy."
      }
    }
  ],
  "gold_outputs": [
    {
      "task_id": "q1",
      "output": { "answer": "Paris" },
      "score": 1.0
    },
    {
      "task_id": "q2",
      "output": { "answer": "William Shakespeare" },
      "score": 1.0
    }
  ],
  "metadata": {
    "name": "simple_qa",
    "task_description": "Answer questions using the provided context"
  }
}

# config.toml
[graph_optimization]
algorithm = "graph_evolve"
dataset_name = "simple_qa"

# Graph settings
graph_type = "policy"
graph_structure = "dag"
topology_guidance = "First extract relevant information from context, then formulate the answer"

# Models
allowed_policy_models = ["gpt-4o-mini"]
verifier_model = "gpt-4o-mini"
scoring_strategy = "rubric"

# Evolution
[graph_optimization.evolution]
num_generations = 3
children_per_generation = 2

[graph_optimization.proposer]
model = "gpt-4.1"

# Data splits
[graph_optimization.seeds]
train = [0, 1, 2, 3, 4]
validation = [5, 6, 7]

# Budget
[graph_optimization.limits]
max_spend_usd = 5.0
timeout_seconds = 1800

from synth_ai.sdk import GraphOptimizationJob

job = GraphOptimizationJob.from_dataset(
    "dataset.json",
    policy_model="gpt-4o-mini",
    rollout_budget=100,
    proposer_effort="medium",
)
job.submit()
result = job.stream_until_complete()
print(f"Best score: {result.best_score}")

# Using Graph Evolve job
output = job.run_inference({
    "question": "What is the largest planet?",
    "context": "Jupiter is the largest planet in our solar system."
})
print(output)  # {"answer": "Jupiter"}

# Get the optimized graph
graph_export = job.download_prompt()
print(graph_export)

Generation 1: best_score=0.65, candidates=5
Generation 2: best_score=0.72, candidates=5
Generation 3: best_score=0.81, candidates=5
Validation: final_score=0.79

[graph_optimization.seeds]
train = [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14]
validation = [15, 16, 17, 18, 19]

topology_guidance = """
For multi-hop reasoning questions:
1. First identify what information is needed
2. Extract relevant facts from context
3. Combine facts to form the answer
"""

[graph_optimization.evolution]
num_generations = 5        # More iterations
children_per_generation = 2  # Fewer variants per iteration