Quick Start

​

Review Feature Overview

1. Configure Review Rules - Create rule groups and rules in the rule repository to define review standards
2. Create Review Tasks - Select extraction tasks that need to be reviewed and submit review tasks
3. View Review Results - Get review results to see which rules passed, which failed, and the review reasoning

​

Interface Usage Workflow

​

Step 1: Create Rule Repository

• Click the “Create Rule Repository” button
• Enter the repository name (e.g., “Invoice Review Rule Repository”)
• Save to get the repository ID

​

Step 2: Create Rule Group

• Select the created rule repository
• Click “Create Rule Group”
• Enter the group name (e.g., “Invoice Compliance Check”)
• Save to get the group ID

​

Step 3: Create Review Rule

• Rule Name: The identifier name of the rule
• Rule Prompt: A prompt describing the review rule, used to guide AI in making review judgments
• Applicable Categories: Which document categories the rule applies to (e.g., invoice, contract, etc.)
• Risk Level: The risk level of the rule (high risk, medium risk, low risk)
• Referenced Fields: Extraction fields that the rule needs to reference, used to get field values during review

• Rule Name: “Invoice Amount Validation”
• Rule Prompt: “Check if the invoice amount is greater than 0 and less than 1000000”
• Applicable Category: Invoice
• Risk Level: High Risk
• Referenced Field: Invoice Amount

​

Step 4: Create Review Task

• Select extraction tasks that need to be reviewed (extract_task_ids)
• Select the rule repository to use (repo_id)
• Enter the task name
• Submit the review task

​

Step 5: View Review Results

• Task Status: View the overall status of the review task (pending, in progress, successful, failed, etc.)
• Rule Group Results: View the review status of each rule group
• Rule Results: View the review result of each specific rule (passed/failed)
• Review Reasoning: View the review reasoning given by AI to understand why it passed or failed
• Position Anchors: View the position of review reasoning in the original text for easy problem location

​

Review Result Description

• Review Status:
  • 0: Pending
  • 1: Approved
  • 2: Review failed
  • 3: Reviewing
  • 4: Rejected
• Review Reasoning: The review reason given by AI, explaining why this review result was obtained
• Position Anchors: Position coordinates of review reasoning in the original text, which can be used to highlight relevant areas in the document

​

Terminology

​

Rule Repository Management

​

Rule Repository

• Rule repositories are logical groupings of rules, making it easy to manage review rules for different business scenarios
• When creating review tasks, you need to specify the rule repository to use
• Rule repositories can be created, updated, and deleted independently

• “Invoice Review Rule Repository” - Contains all invoice-related review rules
• “Contract Review Rule Repository” - Contains all contract-related review rules

​

Rule Group

• Rule groups are used to categorize rules for easy searching and management
• In review results, rule groups are one dimension of result display
• Rule groups can be created, updated, and deleted independently

• “Invoice Compliance Check” - Checks if invoices meet compliance requirements
• “Invoice Amount Validation” - Checks the reasonableness of invoice amounts
• “Invoice Date Validation” - Checks the validity of invoice dates

​

Rule

1. Rule Name: The identifier name of the rule for easy identification and management
2. Rule Prompt: A prompt describing the review rule, used to guide AI in making review judgments. This is the core of the rule and needs to clearly describe the review standards and logic
3. Applicable Categories (Category IDs): Which document categories the rule applies to. Only extraction tasks that match these categories will apply this rule
4. Risk Level:
   • 10: High risk
   • 20: Medium risk
   • 30: Low risk
5. Referenced Fields: Extraction fields that the rule needs to reference, used to get field values during review

{
  "name": "Invoice Amount Validation",
  "prompt": "Check if the invoice amount is greater than 0 and less than 1000000, if not within range, review fails",
  "category_ids": ["invoice_category_id"],
  "risk_level": 10,
  "referenced_fields": [
    {
      "category_id": "invoice_category_id",
      "category_name": "Invoice",
      "fields": [
        {
          "field_id": "amount_field_id",
          "field_name": "Invoice Amount"
        }
      ]
    }
  ]
}

​

Review Object

• Extract tasks are generated by document upload and extraction workflows
• Each extract task corresponds to one or more files
• Extract tasks contain extraction results, including fields, tables, stamps, and other information
• Only tasks that have completed extraction can be used as review objects

• From the task_id returned by the file upload interface
• From the file query interface to get task IDs corresponding to files that have completed extraction

​

Review Task

• Task Name: The identifier name of the review task
• Rule Repository ID: Specifies the rule repository to use
• Extract Task ID List: List of extraction task IDs that need to be reviewed

1. Rule Matching: The system matches rules in the rule repository with extraction task categories
   • Matching criteria: All categories in the rule’s category_ids must exist in the extraction task’s category list
   • Matching result: Only successfully matched rules will be executed
2. Field Retrieval: For matched rules, if the rule has configured referenced fields, the system retrieves values of these fields from extraction results
   • Referenced fields can be empty; if empty, only the original document content is used
   • If fields are missing, those field values are empty, but review will still execute
3. AI Review: Based on rule prompts and field values (if any), AI makes review judgments
4. Result Generation: Generate review results, including review status, reasoning, position anchors, and other information

​

Matching Method Between Review Rules and Extraction Tasks

​

Matching Rules

• Rules are configured with category_ids (applicable category list)
• Extraction tasks contain one or more document categories
• If all categories in the rule’s category_ids exist in the extraction task’s category list, the rule matches successfully
• If any category in the rule’s category_ids is not in the extraction task’s category list, the rule does not match

​

Matching Example

• Review Rule One: Associated categories ["A", "B"]
• Review Rule Two: Associated category ["C"]
• Review Rule Three: Associated categories ["C", "D"]

• ✅ Review Rule One: Match successful (both A and B are in the extraction task’s category list)
• ✅ Review Rule Two: Match successful (C is in the extraction task’s category list)
• ❌ Review Rule Three: Match failed (D is not in the extraction task’s category list)

​

Role of Referenced Fields

• When a rule matches successfully and the system needs to execute review, if the rule prompt needs to use extraction field values, they are retrieved from referenced fields
• Referenced fields can be empty; if empty, only the original document content is used during review, without using extraction fields
• If a rule references fields but those fields are missing in the extraction task, review will still execute, but those field values will be empty

{
  "name": "Invoice Amount Validation",
  "prompt": "Check if the invoice amount is greater than 0 and less than 1000000",
  "category_ids": ["invoice"],
  "referenced_fields": [
    {
      "category_id": "invoice",
      "fields": [
        {
          "field_id": "amount_field_id",
          "field_name": "Invoice Amount"
        }
      ]
    }
  ]
}

• Rule matching: As long as the extraction task’s categories include invoice, the rule will match
• Referenced fields: If the rule matches successfully, the system will retrieve the “Invoice Amount” field value from extraction results during review for judgment
• If the “Invoice Amount” field is missing, review will still execute, but that field value will be empty

​

Referenced Fields

• Referenced fields can be empty; if empty, only the original document content is used during review
• Referenced fields do not affect rule matching; rule matching is based only on categories (category_ids)
• If referenced fields are missing in extraction results, review will still execute, but those field values will be empty

• Regular Fields (Fields): Key-value pair fields in documents
• Table Fields (Tables): Table fields in documents

• When a rule matches successfully and review is executed, the system retrieves values of referenced fields from extraction results
• Field values are provided to AI as context information, combined with rule prompts for review judgment
• If field values are missing, review will still execute, but may affect review accuracy (depending on whether the rule prompt depends on those fields)

{
  "referenced_fields": [
    {
      "category_id": "invoice_category_id",
      "category_name": "Invoice",
      "fields": [
        {
          "field_id": "invoice_code_id",
          "field_name": "Invoice Code"
        },
        {
          "field_id": "invoice_amount_id",
          "field_name": "Invoice Amount"
        }
      ],
      "tables": [
        {
          "table_id": "invoice_items_table_id",
          "table_name": "Invoice Items",
          "fields": [
            {
              "field_id": "item_name_id",
              "field_name": "Item Name"
            }
          ]
        }
      ]
    }
  ]
}

​

Review Reasoning

• Review reasoning is in text form, describing the review judgment process
• Review reasoning will reference relevant field values or document content
• Review reasoning helps users understand review results

{
  "anchors": [
    {
      "start_pos": 0,      // Starting character position in reasoning
      "end_pos": 10,       // Ending character position in reasoning
      "text": "Original text",   // Original text content
      "vertices": [        // Bounding quadrilateral coordinates of original text
        0, 0, 100, 0, 100, 100, 0, 100
      ],
      "file_id": "file_id" // File ID
    }
  ]
}

• Can highlight the position of review reasoning in the document
• Help users quickly locate problem areas
• Provide visual display of review results

​

Review Result Status

• 0: Pending
• 1: Approved
• 2: Review failed
• 3: Reviewing
• 4: Rejected
• 5: Recognizing
• 6: In Queue
• 7: Recognition Failed

​

Summary

​

Next Steps

• Learn Rule Repository Management via API - Use REST API for rule repository management
• Learn Create Review Tasks via API - Use REST API to create review tasks
• Learn Get and Use Review Results - Use REST API to get and process review results