Visual workflow builder and execution platform for SyGra synthetic data pipelines

SyGra is a graph-oriented framework for building synthetic data generation pipelines using LLMs. While powerful, creating and debugging YAML-based workflow configurations requires deep familiarity with the schema and manual iteration.

SyGra Studio provides:

• A visual drag-and-drop interface for designing workflows
• Real-time execution monitoring with node-level progress
• Integrated code editing with syntax highlighting
• Data source preview and transformation testing
• Model management across multiple LLM providers

It replaces the manual YAML editing workflow with an interactive builder while maintaining full compatibility with the SyGra framework.

# From repo root - one command does everything make studio

This automatically builds the frontend (if needed) and starts the server at http://localhost:8000.

# 2. Build the frontend cd studio/frontend npm install npm run build cd ../.. # 3. Start the server uv run python -m studio.server --tasks-dir ./tasks/examples --svelte # 4. Open browser # Navigate to http://localhost:8000

Verification: You should see the SyGra Studio interface with a sidebar listing available workflows from tasks/examples/.

• Production job scheduler
• Multi-tenant platform
• Model training
• Distributed execution

# Clone repository git clone https://github.com/ServiceNow/SyGra.git cd SyGra # Install Python dependencies pip install -e . # Build frontend cd studio/frontend npm install npm run build cd ../..

python -c "from studio import create_server; print('OK')"

Studio uses environment variables for model credentials and settings. Variables are stored in studio/.env and can be managed via the Settings UI.

Model credentials follow the pattern SYGRA_{MODEL_NAME}_{URL|TOKEN}:

# Azure OpenAI SYGRA_GPT-4O_URL=https://your-resource.openai.azure.com/ SYGRA_GPT-4O_TOKEN=your-api-key # vLLM / Self-hosted SYGRA_LLAMA_3_1_8B_URL=http://localhost:8001/v1 SYGRA_LLAMA_3_1_8B_TOKEN=your-token # Ollama (local) SYGRA_QWEN3_URL=http://localhost:11434

Add models to studio/config/custom_models.yaml:

my_custom_model: type: azure_openai model_name: gpt-4o deployment_name: my-deployment api_version: "2024-02-15-preview" parameters: temperature: 0.7 max_tokens: 4096

The project includes a Makefile with convenient commands for Studio. Run from the repository root:

Configuration via environment variables:

# Custom tasks directory and port make studio TASKS_DIR=./my/tasks PORT=9000 # Default values # TASKS_DIR=./tasks/examples # PORT=8000

python -m studio.server [OPTIONS]

# Development with auto-reload python -m studio.server -t ./tasks/examples --reload --svelte # Production python -m studio.server -t /opt/sygra/tasks -p 8080 --svelte # Custom host binding python -m studio.server -H 127.0.0.1 -p 3000 --svelte

from studio import run_server, create_app # Simple: run blocking server run_server(tasks_dir="./tasks", port=8000, use_svelte_ui=True) # Advanced: get FastAPI app for custom middleware app = create_app(tasks_dir="./tasks") # Add custom routes, middleware, etc.

curl -X POST http://localhost:8000/api/workflows/my_workflow/execute \ -H "Content-Type: application/json" \ -d '{  "input_data": [{"question": "What is machine learning?"}],  "num_records": 1,  "batch_size": 25  }'

Response:

{ "execution_id": "exec_abc123", "status": "running", "message": "Execution started" }

curl http://localhost:8000/api/executions/exec_abc123

Response:

{ "id": "exec_abc123", "workflow_id": "my_workflow", "status": "completed", "started_at": "2026-01-19T10:30:00Z", "completed_at": "2026-01-19T10:30:45Z", "duration_ms": 45000, "node_states": { "llm_1": {"status": "completed", "duration_ms": 2500} }, "output_data": [{"response": "Machine learning is..."}] }

1. Click "Create Workflow" in the sidebar
2. Drag nodes from the palette onto the canvas
3. Connect nodes by dragging from output handle to input handle
4. Click each node to configure in the details panel
5. Click "Save" to generate YAML and Python files

1. Drag "LLM" node onto canvas
2. In details panel:
   • Set Summary (display name)
   • Select Model from dropdown
   • Add System and User prompts
   • Set Output Key for state variable
3. Connect to previous/next nodes

1. Select an LLM node
2. Enable "Structured Output"
3. Choose schema mode:
   • Inline: Define fields directly
   • Class Path: Reference a Pydantic model
4. Set fallback strategy (instruction/post_process)

1. Select a Data node
2. Click "Preview" in the data panel
3. Configure source type (CSV, JSON, HuggingFace)
4. Set parameters (file path, repo ID, split)
5. Click "Load Sample" to preview rows

1. Open Workflow
2. Click "Run Workflow" button
3. Configure execution parameters:
   • Number of records
   • Batch size
4. Click "Run Workflow"
5. Monitor progress in the Runs panel

Executions are stored in studio/.executions/:

.executions/ ├── index.json # Metadata index (loaded on startup) └── runs/ ├── exec_abc123.json ├── exec_def456.json └── ... 

Index refresh: If the index becomes stale:

curl -X POST http://localhost:8000/api/executions/storage/refresh