workflow_eval.json
1 { 2 "agents": [ 3 "examples/agents/orchestrator_example.yaml", 4 "examples/workflows/level1_simple_sequential.yaml", 5 "examples/workflows/level2_parallel_execution.yaml", 6 "examples/workflows/level3_conditional_branching.yaml", 7 "examples/workflows/level4_map_iteration.yaml", 8 "examples/workflows/level5_loop_polling.yaml", 9 "examples/workflows/level6_workflow_composition.yaml", 10 "examples/workflows/level7_error_handling.yaml" 11 ], 12 "broker": { 13 "SOLACE_BROKER_URL_VAR": "SOLACE_BROKER_URL", 14 "SOLACE_BROKER_USERNAME_VAR": "SOLACE_BROKER_USERNAME", 15 "SOLACE_BROKER_PASSWORD_VAR": "SOLACE_BROKER_PASSWORD", 16 "SOLACE_BROKER_VPN_VAR": "SOLACE_BROKER_VPN" 17 }, 18 "llm_models": [ 19 { 20 "name": "claude-4-sonnet", 21 "env": { 22 "LLM_SERVICE_PLANNING_MODEL_NAME_VAR": "LLM_SERVICE_PLANNING_MODEL_NAME", 23 "LLM_SERVICE_ENDPOINT_VAR": "LLM_SERVICE_ENDPOINT", 24 "LLM_SERVICE_API_KEY_VAR": "LLM_SERVICE_API_KEY" 25 } 26 } 27 ], 28 "results_dir_name": "workflow-eval", 29 "runs": 1, 30 "workers": 2, 31 "test_cases": [ 32 "tests/evaluation/test_cases/workflows/level1_simple_sequential.test.json", 33 "tests/evaluation/test_cases/workflows/level2_parallel_execution.test.json", 34 "tests/evaluation/test_cases/workflows/level3_conditional_small.test.json", 35 "tests/evaluation/test_cases/workflows/level3_conditional_medium.test.json", 36 "tests/evaluation/test_cases/workflows/level3_conditional_large.test.json", 37 "tests/evaluation/test_cases/workflows/level4_map_iteration.test.json", 38 "tests/evaluation/test_cases/workflows/level5_loop_polling.test.json", 39 "tests/evaluation/test_cases/workflows/level6_workflow_composition.test.json", 40 "tests/evaluation/test_cases/workflows/level7_success_path.test.json", 41 "tests/evaluation/test_cases/workflows/level7_failure_path.test.json" 42 ], 43 "evaluation_settings": { 44 "tool_match": { 45 "enabled": true 46 }, 47 "response_match": { 48 "enabled": true 49 }, 50 "llm_evaluator": { 51 "enabled": true, 52 "env": { 53 "LLM_SERVICE_PLANNING_MODEL_NAME_VAR": "LLM_SERVICE_PLANNING_MODEL_NAME", 54 "LLM_SERVICE_ENDPOINT_VAR": "LLM_SERVICE_ENDPOINT", 55 "LLM_SERVICE_API_KEY_VAR": "LLM_SERVICE_API_KEY" 56 } 57 } 58 } 59 }