Workflow

Multi-step AI task orchestration — triggers, execution engine, variable management

The workflow engine is the execution backbone of ZGI agents. It turns visual node graphs into reliable, production-grade AI pipelines.

Execution Engine

Workflow execution uses SSE (Server-Sent Events) streaming, delivering each node’s results in real-time. The frontend can display step-by-step progress:

Node 1 [Start] ✓ → Node 2 [Knowledge Retrieval] ✓ → Node 3 [LLM Analysis] ⏳ → Node 4 [Output]

Each node execution event is pushed to the client, including:

• Node ID and name
• Execution status (running / success / failed)
• Execution duration
• Output variables and values

Production-Grade Reliability

ZGI workflows are designed for production, not just demos:

• Assertions — Define validation rules per node (e.g., “retrieval must return citations”, “output must match JSON schema”)
• Retry — Failed nodes auto-retry with configurable limits and backoff
• Fallback — When retries exhaust, degrade gracefully (e.g., output a “pending review” list instead of failing silently)

Workflow Lifecycle

1. Draft — Edit and save on the visual canvas. Auto-saved every 30 seconds.
2. Debug — Test with sample inputs. SSE streams node-by-node execution in the debug panel.
3. Publish — One-click publish generates a version number and makes the workflow available via API or chat.
4. Monitor — View run history, execution logs, token consumption, and error rates.

Example: Contract Review Workflow

[Start: Upload Contract]
    → [OCR Parse Document]
    → [Knowledge Retrieval: Legal Clauses]
    → [LLM: Risk Analysis]
    → [Conditional Branch: Risk Level]
        ├── High Risk → [Notify Legal Team via Webhook]
        └── Low Risk → [Generate Report]
    → [End: Return Structured Report with Citations]