Skip to main content

​
Processing Modes

Fintelite AI offers three primary processing modes to handle different use cases:

Predict

Extract structured data using templates or schemas

Parse

Extract full document text and layout information

Fraud Detection

Analyze documents for authenticity and tampering

​
Predict Mode

Extract specific data fields from documents using AI-powered templates or custom schemas.

​
When to Use

  • Extracting invoice details (line items, totals, dates)
  • Processing ID cards and passports
  • Parsing financial statements
  • Any structured data extraction task

​
How It Works

1

Upload Document

Upload your document or reference an existing file by ID
2

Define Schema

Specify what data to extract using a JSON schema or template
3

AI Processing

The AI analyzes the document and extracts requested fields
4

Get Results

Receive structured JSON with extracted data and confidence scores
For complete request/response examples and API details, see the Predict API Reference. For understanding job status and response formats, see the Jobs Concept Guide.

​
Parse Mode

Extract complete document text with layout information including blocks and markdown.

​
When to Use

  • Full document text extraction
  • Document analysis and search
  • Building document databases
  • RAG applications (document parsed into markdown)
  • Pre-processing for another pipeline

​
Features

  • Block Detection: Identifies headers, footers, tables, figures, and text blocks
  • Markdown Output: Converts document to markdown format
  • Multi-page Support: Handles documents with any number of pages
For complete request/response examples and API details, see the Parse API Reference. For understanding job status and response formats, see the Jobs Concept Guide.

​
Document Formats Supported

PDF

Native PDF parsing and OCR for scanned PDFs

Images

JPG, PNG, HEIC formats with OCR

Multi-page TIFF

Process multi-page TIFF documents

URLs

Process documents from public URLs

​
Synchronous vs Asynchronous

Use async endpoints (/predict-async, /parse-async, /fraud-async) for production. Sync endpoints may timeout on large documents.
  • Sync: Returns results immediately, best for small documents such as ID cards, receipts, and single-page invoices
  • Async: Returns job ID, check status later via /status/{job_id}, best for large documents and batch processing
For job management details, see the Jobs Concept Guide.

​
File Input Methods

Fintelite AI supports three ways to provide document files:

​
1. File Upload

Upload a new file directly in the request: 
curl -X POST https://api-vision.fintelite.ai/predict \
  -H "X-API-Key: YOUR_API_KEY" \
  -F "files=@/path/to/document.pdf" \
  -F 'schema={...}'

​
2. File ID Reference

Reference a previously uploaded file: 
curl -X POST https://api-vision.fintelite.ai/predict \
  -H "X-API-Key: YOUR_API_KEY" \
  -F "files=id://file-uuid-here" \
  -F 'schema={...}'

​
3. URL Reference

Process a document from a public URL: 
curl -X POST https://api-vision.fintelite.ai/predict \
  -H "X-API-Key: YOUR_API_KEY" \
  -F "files=https://example.com/document.pdf" \
  -F 'schema={...}'

​
Configuration Options

Control document processing behavior with parser, chunking, and citation settings.

Parser Configuration

Control parsing behavior and quality modes

Chunking Strategies

Split large documents for better processing

Citations

Track source of extracted data
For all configuration options and examples, see the Configuration Overview.

​
Confidence Scores

Every extraction includes confidence scores on a 0-100 scale, with nested structure matching your extraction data. Confidence structure mirrors extraction: 
{
  "confidences": {
    "average": 87.5,
    "fields": {
      "invoice_number": 92.3,
      "date": 85.0,
      "total": 88.7,
      "items": [
        {
          "description": 90.5,
          "amount": 89.0
        }
      ]
    }
  }
}
Confidence scores are not accuracy metrics. They indicate how confident the AI model is, not how accurate the extraction is. The AI can make mistakes even with high confidence scores. Always implement validation and review processes for critical data.
Fields with confidence below 80 have higher likelihood of errors and should be manually reviewed.

​
Citations

Track which document blocks each extracted value came from for verification and debugging. Key Points:
  • Requires use_parser: true
  • Disabled automatically in parser_mode: LITE
  • Changes output format to include value and citations fields
  • Citation format: b.X = block number, w.X = word number
For complete citation details, format, and usage examples, see the Citations Configuration.

​
Best Practices

  • Use high-resolution scans (300 DPI minimum)
  • Ensure good lighting and contrast
  • Avoid skewed or rotated images
  • Remove artifacts and noise
  • Be specific with field names and descriptions
  • Use appropriate data types (string, number, integer, boolean, array, object)
  • Use enum for fields with predefined values
  • Include nested structures for complex data (max 3 levels)
  • Test with sample documents first
  • Reuse uploaded files via file IDs
  • Use async mode for large documents
  • Implement webhooks instead of polling
  • Cache frequently used templates
  • Check confidence scores
  • Implement retry logic for failures
  • Validate extracted data
  • Handle missing or null fields

​
Next Steps

Templates

Learn about template management

API Reference

Explore document processing endpoints