HITL Correction Node

Pause workflow execution for human data correction and editing.

Overview

The HITL Correction Node pauses workflows and presents data to human reviewers for editing or correction. Unlike the Approval Node (which only allows approve/reject), the Correction Node enables reviewers to:

Edit extracted fields - Fix OCR or LLM extraction errors
Add missing data - Fill in fields the automation couldn’t extract
Validate and refine - Verify and improve automated outputs
Annotate - Add notes or metadata to processed data

Corrected data replaces the original in the workflow context and execution continues.

When to Use

Use a HITL Correction Node when you need:

Fix extraction errors - OCR misreads or LLM hallucinations
Handle edge cases - Documents with unusual layouts or missing fields
Training data collection - Gather human corrections to improve models
Incremental automation - Start with human review, automate over time
Quality improvement - Refine outputs before exposing to downstream systems

For simple yes/no approval, use HITL Approval Node instead.

Configuration

Correction Task

Define what reviewers are correcting:

Field	Description	Example
Title	Short summary	”Verify extracted invoice fields”
Description	Instructions for reviewer	”Review and correct any extraction errors”
Input source	Data to display for correction	`$.nodes.extract_invoice.output`
Editable fields	Which fields can be edited	See schema configuration below

Field Schema

Define editable fields with JSON schema:


{
  "fields": [
    {
      "name": "invoice_number",
      "label": "Invoice Number",
      "type": "string",
      "required": true,
      "validation": {
        "pattern": "^INV-\\d{6}$",
        "error_message": "Format: INV-123456"
      }
    },
    {
      "name": "date",
      "label": "Invoice Date",
      "type": "date",
      "required": true
    },
    {
      "name": "total_amount",
      "label": "Total Amount",
      "type": "number",
      "required": true,
      "validation": {
        "min": 0,
        "error_message": "Amount must be positive"
      }
    },
    {
      "name": "line_items",
      "label": "Line Items",
      "type": "array",
      "items": {
        "type": "object",
        "properties": {
          "description": { "type": "string" },
          "quantity": { "type": "number" },
          "unit_price": { "type": "number" }
        }
      }
    },
    {
      "name": "notes",
      "label": "Reviewer Notes",
      "type": "textarea",
      "required": false
    }
  ]
}

Supported field types:

string - Single-line text input
textarea - Multi-line text input
number - Numeric input with optional min/max
date - Date picker
select - Dropdown with predefined options
checkbox - Boolean value
array - Repeating sub-fields (for line items, etc.)
object - Nested field group

Reviewer Assignment

Specify who can make corrections:


{
  "reviewers": ["role:data_entry", "team:processing"],
  "assignment_type": "any",
  "priority": "high"
}

Assignment types:

any: Any reviewer can correct
round_robin: Distribute tasks evenly across reviewers
load_balanced: Assign to reviewer with fewest pending tasks

Reference Data

Display supporting information to help reviewers:


{
  "reference_data": {
    "original_document": {
      "type": "image",
      "url": "$.data.document_url",
      "label": "Original Invoice"
    },
    "ocr_confidence": {
      "type": "metric",
      "value": "$.nodes.ocr.output.confidence",
      "label": "OCR Confidence",
      "format": "percentage"
    }
  }
}

Reference data appears in a side panel while reviewers edit fields.

Validation Rules

Enforce data quality with validation:


{
  "validation": {
    "required_fields": ["invoice_number", "date", "total_amount"],
    "cross_field_checks": [
      {
        "rule": "line_items_total == total_amount",
        "error_message": "Line items must sum to total amount"
      }
    ],
    "custom_validator": {
      "endpoint": "validator_executor://validate_invoice",
      "function": "check_invoice_fields"
    }
  }
}

Validation runs when reviewer submits corrections. Errors prevent submission until fixed.

Review Interface

Reviewers access correction tasks via:

In-app queue - OPERATE → Reviews → Corrections tab
Email notification - Link to specific task
API integration - Embed correction UI in external tools

Correction UI displays:

Original values - What automation extracted (grayed out for reference)
Editable fields - Form inputs matching configured schema
Reference documents - Original PDFs, images, etc. in side panel
Confidence indicators - Visual cues for low-confidence fields
Validation feedback - Real-time error messages for invalid inputs
Action buttons - Submit, Skip, Escalate

Field-Level Confidence

Highlight low-confidence extractions for reviewer attention:


{
  "fields": [
    {
      "name": "invoice_number",
      "value": "INV-123456",
      "confidence": 0.95,
      "highlight_threshold": 0.8
    },
    {
      "name": "date",
      "value": "2026-03-19",
      "confidence": 0.65,
      "highlight_threshold": 0.8,
      "needs_review": true
    }
  ]
}

Fields with confidence < highlight_threshold are marked for review (yellow border, warning icon).

Output

The HITL Correction Node produces corrected data accessible in downstream nodes:


{
  "correction_result": {
    "corrected_data": {
      "invoice_number": "INV-654321",
      "date": "2026-03-20",
      "total_amount": 1234.56,
      "line_items": [
        { "description": "Item A", "quantity": 2, "unit_price": 500.00 },
        { "description": "Item B", "quantity": 1, "unit_price": 234.56 }
      ],
      "notes": "Corrected date and line items"
    },
    "changes": [
      {
        "field": "invoice_number",
        "original": "INV-123456",
        "corrected": "INV-654321",
        "confidence_original": 0.95
      },
      {
        "field": "date",
        "original": "2026-03-19",
        "corrected": "2026-03-20",
        "confidence_original": 0.65
      }
    ],
    "reviewer_id": "user_xyz",
    "timestamp": "2026-03-19T15:45:30Z",
    "selected_path_id": "corrected"
  }
}

Access corrected data via JSONPath:

$.nodes.correction_node.output.correction_result.corrected_data - Full corrected object
$.nodes.correction_node.output.correction_result.changes - List of changes made
$.nodes.correction_node.output.correction_result.reviewer_id - Who made corrections

Routing Paths

Configure paths based on correction outcome:


{
  "paths": [
    {
      "path_id": "corrected",
      "target_node_ids": ["validate_corrected", "process_invoice"]
    },
    {
      "path_id": "skipped",
      "target_node_ids": ["needs_investigation"]
    },
    {
      "path_id": "escalated",
      "target_node_ids": ["expert_review"]
    }
  ]
}

Paths:

corrected: Reviewer submitted corrections
skipped: Reviewer marked task as unable to correct
escalated: Reviewer escalated to expert

Example Configurations

Invoice Field Correction


{
  "node_id": "correct_invoice",
  "node_type": "HITL_CORRECTION",
  "definition": {
    "title": "Verify Invoice Extraction",
    "description": "Review and correct extracted invoice fields",
    "input_source": "$.nodes.extract_invoice.output",
    "reviewers": ["role:data_entry"],
    "assignment_type": "load_balanced",
    "fields": [
      {
        "name": "invoice_number",
        "label": "Invoice Number",
        "type": "string",
        "required": true,
        "value": "$.nodes.extract_invoice.output.invoice_number",
        "confidence": "$.nodes.extract_invoice.output.confidence.invoice_number"
      },
      {
        "name": "date",
        "label": "Invoice Date",
        "type": "date",
        "required": true,
        "value": "$.nodes.extract_invoice.output.date",
        "confidence": "$.nodes.extract_invoice.output.confidence.date"
      },
      {
        "name": "total_amount",
        "label": "Total Amount",
        "type": "number",
        "required": true,
        "value": "$.nodes.extract_invoice.output.total_amount",
        "confidence": "$.nodes.extract_invoice.output.confidence.total_amount"
      }
    ],
    "reference_data": {
      "original_invoice": {
        "type": "pdf",
        "url": "$.data.document_url"
      }
    },
    "paths": [
      {
        "path_id": "corrected",
        "target_node_ids": ["process_invoice"]
      },
      {
        "path_id": "skipped",
        "target_node_ids": ["flag_for_review"]
      }
    ]
  }
}

Multi-Page Document Correction

For documents with repeating structures (line items, signatures):


{
  "fields": [
    {
      "name": "header",
      "label": "Invoice Header",
      "type": "object",
      "properties": {
        "invoice_number": { "type": "string" },
        "date": { "type": "date" },
        "vendor": { "type": "string" }
      }
    },
    {
      "name": "line_items",
      "label": "Line Items",
      "type": "array",
      "min_items": 1,
      "items": {
        "type": "object",
        "properties": {
          "description": { "type": "string", "required": true },
          "quantity": { "type": "number", "required": true },
          "unit_price": { "type": "number", "required": true },
          "subtotal": { "type": "number", "computed": "quantity * unit_price" }
        }
      },
      "actions": ["add_row", "delete_row", "reorder"]
    }
  ]
}

Reviewers can add/remove line items and reorder them.

Training Data Collection

Collect correction data to improve automation models:


{
  "training_data": {
    "enabled": true,
    "fields_to_log": ["invoice_number", "date", "total_amount"],
    "include_original": true,
    "include_corrected": true,
    "include_document": true,
    "destination": "s3://training-data/invoice-corrections/"
  }
}

Each correction is logged as:


{
  "document_id": "doc_abc123",
  "original_extraction": { "invoice_number": "INV-123456", ... },
  "corrected_extraction": { "invoice_number": "INV-654321", ... },
  "document_url": "s3://...",
  "reviewer_id": "user_xyz",
  "timestamp": "2026-03-19T15:45:30Z"
}

Use this data to retrain extraction models and reduce correction rate over time.

Metrics and Analytics

Track correction performance:

Metric	Description	Use Case
Correction rate	% of documents requiring correction	Measure automation quality
Fields corrected	Which fields are corrected most often	Identify extraction weaknesses
Time per correction	Average reviewer time	Estimate capacity needs
Reviewer accuracy	% of corrections later deemed incorrect	Quality control for reviewers

Access metrics via OPERATE → Analytics → Corrections.

Best Practices

Prioritize by confidence - Route low-confidence extractions to correction queue first
Show original document - Always display source PDF/image for reference
Pre-fill with automation - Start with automated extraction, let reviewers fix
Validate aggressively - Catch errors before reviewers submit
Minimize fields - Only show fields that need correction, hide confident extractions
Track changes - Log all corrections for training data and audit trail
Set deadlines - Configure timeouts to prevent tasks from aging indefinitely
Monitor trends - If correction rate increases, investigate automation quality

Start with 100% correction (all documents reviewed) and gradually reduce as automation improves. Monitor metrics to safely decrease correction rate.

Integration with Quality Loops

Use corrections to improve automation:

This creates a feedback loop where human corrections improve the model over time.

API Access

Programmatically manage correction tasks:


// Get pending corrections
const pending = await trpc.corrections.getPending.query();
 
// Submit corrections
await trpc.corrections.submit.mutate({
  task_id: 'task_abc123',
  corrected_data: {
    invoice_number: 'INV-654321',
    date: '2026-03-20',
    total_amount: 1234.56
  }
});
 
// Skip a task
await trpc.corrections.skip.mutate({
  task_id: 'task_abc123',
  reason: 'Document is illegible'
});

HITL Approval Node - Simple approve/reject
HITL Router Node - Route to specific reviewers
Guardrail Node - Automated validation