DocIntelligence

DocIntelligence is a powerful document processing engine that transforms PDFs into structured, searchable knowledge bases. It combines advanced layout analysis with intelligent content extraction to make your documents truly AI-ready for RAG (Retrieval-Augmented Generation) applications.

Background

In today's AI-driven world, RAG (Retrieval-Augmented Generation) has become essential for creating AI applications that can access and reason about specific document collections. While many excellent open-source PDF processors rely solely on OCR (Optical Character Recognition), they often struggle with:

• Complex document layouts
• Mixed content types (text, tables, images, formulas)
• Maintaining document structure and hierarchy
• Generating high-quality embeddings for semantic search

DocIntelligence addresses these challenges by combining advanced layout analysis with intelligent content extraction, making your documents truly AI-ready.

Key Features

Feature	Description	Benefit
🔍 Smart Layout Analysis	Automatically detects and preserves document structure	Maintains context and relationships between elements
📊 Multi-Modal Extraction	Handles text, tables, images, and formulas	Complete document understanding
🧠 Semantic Embeddings	Generates vector embeddings for all content	Enables powerful similarity search
☁️ Flexible Storage	Store locally or in Google Cloud (BigQuery/Storage)	Scales with your needs
🔗 Knowledge Base Ready	Structured output perfect for RAG applications	Build smarter AI applications

Installation

1. Clone the repository:

git clone https://github.com/Zereo0317/DocIntelligence.git
cd DocIntelligence

2. Install dependencies:

pip install -e .  # Install DocIntelligence in development mode

Google Cloud Setup

DocIntelligence uses Google Cloud Platform (GCP) for OCR and optionally for enhanced storage features. You can set up your environment using tools like python-dotenv or direnv.

1. Set up GCP Authentication:

   • Create a Google Cloud Project
   • Enable the Cloud Vision API
   • Create a Service Account with these roles:
     • Cloud Vision API User
     • Storage Object Viewer (if using Cloud Storage)
     • BigQuery Data Editor (if using BigQuery)
   • Download the JSON key file

2. Create a .env file in your project root:

# Required: Google Cloud Vision API for OCR
GCP_PROJECT_ID=""              # Your GCP Project ID
GCP_LOCATION=""               # Service location (defaults to us-central1)

# Optional: Cloud Storage for visual & tabular content
GCP_BUCKET_NAME=""            # Cloud Storage bucket name
GOOGLE_APPLICATION_CREDENTIALS=""   # Path to service account JSON file

# Optional: BigQuery for embeddings & search
GCP_DATASET_ID=""            # BigQuery dataset name
GCP_CONNECTION_ID=""         # BigQuery connection ID
GCP_BIGQUERY_INSERT_BATCH=500      # Batch size for insertions

Usage

DocIntelligence offers two main processing modes:

Local Processing

Perfect for smaller projects or testing. Stores results locally:

from dotenv import load_dotenv
load_dotenv()  # Load variables from .env file
from DocIntelligence import DocIntelligence

engine = DocIntelligence()

# Process documents and get results directly
elements, documents, embeddings = engine.process_documents(
    input_dir="./Documents/",
    output_dir="./Output/",
    store_to_db=False,    # False => Store locally, True => Store in BigQuery
    cloud_storage=False   # False => Store locally, True => Store in Google Cloud Storage
)

Cloud Processing

Ideal for large-scale applications. Stores results in GCP:

engine = DocIntelligence()

# Process and store in GCP
engine.process_documents(
    input_dir="./Documents/",
    output_dir="./Output/",
    store_to_db=True,     # Store in BigQuery
    cloud_storage=True    # Use Google Cloud Storage
)

Understanding Storage Options

• store_to_db=False: Results are stored in the local output directory

  • Extracted text stored as JSON
  • Images and tables saved as files
  • Embeddings returned as Python objects

• store_to_db=True: Results are stored in BigQuery

  • Structured data stored in tables
  • Enables fast querying and search
  • Perfect for production deployments

• cloud_storage=False: Visual elements stored locally

• cloud_storage=True: Visual elements stored in Google Cloud Storage

  • Better for sharing and scalability
  • Enables cloud-based processing

Return Values Structure

When store_to_db=False, the function returns three lists of dictionaries:

elements

List of detected elements from the document:

{
    'doc_id': str,              # Document ID (Required)
    'page_num': int,            # Page number (Required)
    'element_type': str,        # Element type (Text, Table, Picture, etc.)
    'element_id': str,          # Unique identifier
    'storage_path': str,        # Storage path for the element
    'embedding_id': str,        # Embedding identifier
    'content': str,             # Element content
    'metadata': dict,           # Element metadata in JSON format
    'mapped_to_element_id': str,# Related element ID if applicable
    'store_in_bigquery': bool,  # Whether to store in BigQuery
    'section': str,             # Section information
    'title': str               # Title information
}

documents

List of document metadata:

{
    'doc_id': str,          # Document ID (Required)
    'title': str,           # Document title
    'total_pages': int,     # Total number of pages
    'storage_path': str     # Storage path
}

embeddings

List of vector embeddings:

{
    'embedding_id': str,           # Embedding identifier
    'vector': "" | List[float]     # Empty string when store_to_db=False, 
                                  # List of floats when store_to_db=True
    'original_text': str,          # Original text content
    'content_type': str,           # Content type
    'doc_id': str,                 # Document ID
    'page_num': int,               # Page number
    'element_id': str,             # Element identifier
    'mapped_to_element_id': str,   # Related element ID
    'coordinates': dict            # Position coordinates in JSON
}

Output Structure

Output/
└── {document_name}/
    ├── images/         # Page images
    ├── labeled/        # Extracted elements
    │   ├── text/      # Text blocks
    │   ├── table/     # Table images
    │   ├── formula/   # Mathematical formulas
    │   └── picture/   # Figures and diagrams
    └── ocr/           # Processed content
        ├── text/      # OCR results
        ├── table/     # Table data
        └── caption/   # Image captions

Contributing

We welcome contributions! Here's how you can help:

1. Check our Issues for tasks
2. Fork the repository and create a new branch
3. Submit a Pull Request with your changes
4. Join our community discussions

Need technical support or want to connect? Feel free to email Zereo at zereo@zereo-ai.com or connect on LinkedIn. We're always happy to help!

Future Roadmap

We're actively developing new features to make DocIntelligence even more powerful:

1. Knowledge Graph Integration

   • Automatically build knowledge graphs from documents
   • Discover relationships between concepts
   • Enable graph-based querying

2. AgentBasis Integration

   • Seamless connection to our upcoming AgentBasis framework
   • Build intelligent agents that can reason over your documents
   • Create powerful RAG applications with minimal code

Requirements

• Python ≥ 3.10
• Google Cloud Vision API (for OCR)
• Optional: Google Cloud Storage and BigQuery (for cloud features)

License

DocIntelligence is released under the MIT License. You are free to use, modify, and distribute the code for both commercial and non-commercial purposes.