Skip to content

Workflows

Workflows are the primary way to get value from G6. Each workflow takes your data and produces a measurable, auditable output artifact.

The pattern

Every G6 workflow follows the same self-training loop:

graph LR
    A[Your Data<br/>examples] --> B[Build<br/>Harness]
    B --> C[Evaluate<br/>held-out]
    C --> D{Quality<br/>Target Met?}
    D -->|No| E[Diagnose<br/>Failures]
    E --> F[Improve<br/>Harness]
    F --> C
    D -->|Yes| G[Report<br/>evidence]
  1. Load your examples — labeled data for the task you want to automate
  2. Build a harness — G6 creates a task-specific classifier/processor
  3. Evaluate on held-out data — measure quality with data the harness hasn't seen
  4. Diagnose failures — classify errors into actionable root causes
  5. Improve the harness — fix identified issues based on diagnosis
  6. Generate a report — evidence of quality, cost, safety, and recommendations

Available workflows

  • Business Document Classifier

    Turn labeled support tickets, queries, or documents into a tested classification harness. Demonstrates the full improvement loop with failure taxonomy and quality reporting.

    Business Classifier

Surface guides

Run workflows through your preferred interface:

Surface Guide Best for
MCP (Claude Code) MCP First Workflow Interactive exploration with AI assistant
REST API REST Workflow Guide Programmatic integration and automation
Web Portal Web Onboarding Browser-based setup and billing

Understanding G6 behavior

Topic Guide
Cold start vs cached Cold Start Explained
When things degrade Degradation Warnings

What workflows produce

Every workflow generates a report containing:

  • Quality metrics — accuracy, precision, recall, F1 on held-out data
  • Failure taxonomy — root causes grouped into fixable categories
  • Before/after comparison — proof of improvement (or honest admission of stagnation)
  • Cost summary — time, API calls, estimated spend
  • Safety status — guardrail compliance, data containment, resource bounds
  • Recommendations — what to do next (deploy, iterate, or pivot)

Bring your own data

All workflows accept user-provided data in CSV format. Use the customer_template.csv in each workflow directory as a starting point. Minimum 20 labeled examples recommended.