​
Overview

Document parsing in our Vision API is achieved through intelligent chunking strategies that analyze document structure using advanced AI and vision language models. The parsing functionality identifies different document elements like text blocks, tables, images, headers, and footers while maintaining proper reading order and extracting content with high accuracy.
Parsing capabilities are accessed through the /chunking endpoint, which combines structure detection with content extraction and intelligent segmentation.

​
Key Features

Element Detection

Identify and classify document elements using advanced AI models

Content Extraction

Extract text, tables, and images with appropriate processing methods

Reading Order

Maintain proper document flow and reading sequence

Multi-Modal Processing

Handle text, images, tables, and formulas with specialized extractors

​
Document Element Types

The parsing system can identify and process the following element types through chunking strategies:

​
Text Elements

  • Text: Regular paragraph text
  • Title: Document and section titles
  • Section-header: Section headings
  • Page-header: Header content
  • Page-footer: Footer content
  • Caption: Image and table captions
  • Footnote: Footnote references and content
  • List-item: Bulleted and numbered lists

​
Visual Elements

  • Table: Structured tabular data
  • Picture: Images, charts, and diagrams
  • Formula: Mathematical equations and expressions

​
Parsing Through Chunking Strategies

​
Semantic Parsing with Chunking

import requests

# Parse and chunk document semantically
url = "https://prod.visionapi.unsiloed.ai/chunking"
headers = {
    "accept": "application/json",
    "X-API-Key": "your-api-key"
}

files = {
    "document_file": ("document.pdf", open("document.pdf", "rb"), "application/pdf")
}

data = {
    "strategy": "semantic",  # Uses AI for intelligent parsing
    "chunk_size": 1000,
    "overlap": 100
}

response = requests.post(url, headers=headers, files=files, data=data)
job_info = response.json()

print(f"Parsing job started: {job_info['job_id']}")

# Check job status
status_url = f"https://prod.visionapi.unsiloed.ai/jobs/{job_info['job_id']}"
status_response = requests.get(status_url, headers=headers)
print(f"Status: {status_response.json()['status']}")

​
Parsing Strategies

Uses advanced AI to intelligently identify and group document elements. Best for maintaining semantic context and understanding document structure.

​
Hybrid Parsing

Multi-modal processing that combines table extraction, image analysis, and enhanced metadata extraction. Most comprehensive parsing option.

​
Heading-Based Parsing

Analyzes document structure based on detected headings and section hierarchy. Ideal for structured documents with clear heading patterns.

​
Page-Based Parsing

Processes documents page by page, maintaining page boundaries. Useful for documents where page structure is important.

​
Response Format

​
Chunked Parsing Results

When using chunking strategies (/chunking), you get parsed content organized into meaningful chunks: 
{
  "job_id": "b2094b38-e432-44b6-a5d0-67bed07d5de1",
  "status": "completed",
  "result": {
    "chunks": [
      {
        "content": "Annual Financial Report 2023\n\nExecutive Summary\n\nThis report presents our financial performance for the fiscal year 2023...",
        "metadata": {
          "chunk_index": 0,
          "page_numbers": [1],
          "element_types": ["Title", "Section-header", "Text"],
          "reading_order": [0, 1, 2],
          "confidence_scores": [0.95, 0.92, 0.88]
        }
      },
      {
        "content": "Financial Performance\n\n| Quarter | Revenue | Profit |\n|---------|---------|--------|\n| Q1 | $1.2M | $200K |\n| Q2 | $1.4M | $280K |",
        "metadata": {
          "chunk_index": 1,
          "page_numbers": [1, 2],
          "element_types": ["Section-header", "Table"],
          "reading_order": [3, 4],
          "confidence_scores": [0.91, 0.92]
        }
      }
    ],
    "total_chunks": 2,
    "processing_metadata": {
      "strategy_used": "semantic",
      "total_pages": 2,
      "total_elements_detected": 8,
      "processing_time": 4.2
    }
  }
}

​
Advanced Parsing Features

​
Multi-Format Support

The parsing system supports multiple document formats: 
# Parse different document types
document_types = [
    ("document.pdf", "application/pdf"),
    ("presentation.pptx", "application/vnd.openxmlformats-officedocument.presentationml.presentation"),
    ("report.docx", "application/vnd.openxmlformats-officedocument.wordprocessingml.document")
]
    
for filename, content_type in document_types:
    files = {"document_file": (filename, open(filename, "rb"), content_type)}
    response = requests.post(chunking_url, headers=headers, files=files, data=data)
    print(f"Parsed {filename}: {response.json()['job_id']}")

​
Element-Specific Processing

Different element types are processed with specialized methods:
  • Text Elements: OCR with post-processing for clarity
  • Tables: Structured extraction maintaining column/row relationships
  • Images: Vision analysis for content description
  • Formulas: Mathematical expression recognition
  • Headers/Footers: Context-aware text extraction