AI Interface

The FormFiller system provides a working AI interface that drastically reduces the time to create forms and other structures. This is not a future plan, but a currently available, production feature.

Overview

The AI interface enables:

• Natural language form generation - Working schema from simple text description
• Complex structure creation - Not just forms, but workflows, data models too
• Iterative refinement - Handling "Add to it..." type requests
• Automatic validation rule recognition - Rules inferred from description

Drastic Time Savings

Method	Average Time	Savings
Traditional coding (MVC)	4-8 hours	-
Manual JSON schema writing	30-60 minutes	~85%
AI generation	2-5 minutes	~98%

The AI interface represents up to 98% time savings compared to traditional development!

Operating Principle

flowchart TB
    INPUT["1. USER INPUT<br/>'Create an employee data sheet<br/>with personal data...'"]

    subgraph AI["2. AI PROCESSING"]
        E1["Entity identification"]
        E2["Field type determination"]
        E3["Validation rules"]
        E4["Structure creation"]
    end

    SCHEMA["3. SCHEMA GENERATION<br/>FormFiller-compatible JSON"]
    VALID["4. VALIDATION<br/>Schema and type checking"]
    FORM["5. WORKING FORM<br/>Immediately usable"]

    INPUT --> AI
    AI --> SCHEMA
    SCHEMA --> VALID
    VALID --> FORM

    style INPUT fill:#fff3cd,stroke:#856404
    style AI fill:#e6f3ff,stroke:#0066cc
    style FORM fill:#ccffcc,stroke:#00cc00

Usage Guide

Basic Generation

Simple prompt on the frontend interface:

Prompt: "Create a contact form with name, email, 
        phone number and message fields. Email should be required."

Result:

{
  "type": "form",
  "title": "Contact",
  "items": [
    { 
      "name": "name", 
      "title": "Name", 
      "type": "text",
      "validationRules": [{ "type": "required" }]
    },
    { 
      "name": "email", 
      "title": "Email", 
      "type": "text",
      "validationRules": [
        { "type": "required" },
        { "type": "email" }
      ]
    },
    { 
      "name": "phone", 
      "title": "Phone Number", 
      "type": "text" 
    },
    { 
      "name": "message", 
      "title": "Message", 
      "type": "textarea" 
    }
  ]
}

Complex Structure Generation

The AI can generate complex, nested structures:

Prompt: "Create an order form where customer data 
        (name, email) and shipping address are in separate groups.
        Allow adding multiple products with quantity and price,
        and automatically calculate the total."

The AI recognizes: - Need for grouping - Grid (tabular) input requirement - ComputedRules need for sum calculation

Iterative Refinement

The AI supports iterative development:

1. Prompt: "Create a registration form"
   → Basic registration form generated

2. Prompt: "Add phone number too, and add password confirmation"
   → Existing schema extended

3. Prompt: "Password should be minimum 8 characters"
   → Validation rule added

Supported Generation Types

1. Forms

Prompt types:
• "Create a [type] form..."
• "Create a data entry form..."
• "Generate a registration form..."

2. Workflow Definitions

Prompt: "Create an approval workflow where first the 
        direct supervisor, then the HR manager approves the leave request.
        In case of rejection, send notification to the requester."

The AI generates: - Workflow steps - State transitions - Notification rules

3. Data Structures

Prompt: "Design a CRM data structure with customers, contacts,
        and business opportunities. Projects should belong to customers."

The AI generates: - Entity definitions - Relationships - Basic validations

4. Exam Sheets and Quizzes

Prompt: "Create a 10-question test about JavaScript basics.
        Each question is worth 10 points, and calculate total score.
        Above 60% the exam should be successful."

The AI generates: - Question fields - ComputedRules for scoring - Result evaluation

Non-Form Applications

The AI interface is not limited to forms. FormFiller's declarative schema language is suitable for describing any structured data:

Application	Example Prompt
System configuration	"Define an e-commerce settings structure"
API specification	"Create endpoint definition for product CRUD"
Access control	"Design RBAC structure with admin, manager, user roles"
Content management	"Create blog post schema with categories and tags"
Report definition	"Define monthly sales report structure"

AI Advantages Summary

mindmap
  root((AI Interface<br/>Advantages))
    Speed
      Working configuration in minutes
      98% time savings
    Accuracy
      Validated output
      Schema standard compliant
    Iteration
      Gradual refinement
      Natural language modifications
    Best Practices
      Schema conventions
      Automatic validation
      Optimal structure
    Flexibility
      Not just forms
      Workflow, data model

System Integration

The AI-generated schema is immediately usable:

1. Frontend rendering - Generated JSON can be displayed immediately
2. Backend validation - Validation rules are automatically enforced
3. Database storage - Structure can be stored in MongoDB
4. Workflow integration - Workflow definitions are immediately executable

Technical Architecture

The AI generation feature uses a dedicated backend endpoint:

POST /api/ai/generate-stream

Key technical features:

• Server-Sent Events (SSE): Real-time streaming of generated content
• Two-phase generation: Category detection + config generation for better accuracy
• Token budget validation: Ensures prompts fit within model context limits
• Rate limiting: Prevents API abuse (10 requests/minute by default)
• Retry mechanism: Automatic retry with exponential backoff on failures
• Example-based learning: Uses seed configurations as context for generation

The AI service is implemented in src/routes/ai-generation.ts and uses the AIGenerationService for LLM integration.

Example: Complete Workflow

1. User: "Create a leave request form with dates,
               justification, and approval workflow"

2. AI generates:
   - Form schema (date fields, justification textarea)
   - Validations (date range, required fields)
   - Workflow (submission → approval → notification)

3. Result: Working leave request system in 2 minutes
   (Traditional development: 1-2 days)

Next Steps

The AI interface is continuously evolving. Planned expansions:

• Visual editor integration (AI + Drag & Drop)
• Analysis and optimization of existing systems
• Multi-language prompt support
• Template suggestions based on context