docHandler4AI Python SDK

The docHandler4AI Python SDK provides a high-level, asynchronous interface to the docHandler4AI server. It enables advanced PDF processing, multimodal vision analysis, and robust vector indexing for both text and images.

Table of Contents

• Installation
• Quick Start
• PDF Processing
• Vision Analysis
• Document Indexing
  • Searching with Metadata Filters
  • Self-Query Search
• Image Indexing
  • Multimodal Search

Installation

pip install docHandler4AI-sdk

Quick Start

Initialize the main client facade:

import asyncio
from dochandler4ai_sdk import DocHandler4AI

async def main():
    sdk = DocHandler4AI(
        base_url="http://localhost:8000",
        api_key="your_api_key"
    )
    
    # Use the SDK...
    
asyncio.run(main())

PDF Processing

Extract structured content from PDFs using Vision-based AI.

pdf_processor = sdk.get_pdf_process()

# Process PDF to Markdown with chunks
result = await pdf_processor.process(
    pdf_url="https://example.com/document.pdf",
    response_output="markdown",
    return_chunks=True
)

print(result["markdown"])
for chunk in result.get("chunks", []):
    print(chunk["content"])

Vision Analysis

Perform OCR, generate captions, or analyze specific PDF pages.

vision = sdk.get_vision_process()

# Analyze image (OCR, caption, or pdf_page)
analysis = await vision.analyze(
    image_url="https://example.com/image.png",
    type="caption"
)

print(analysis["content"])

Document Indexing

Manage vector collections for text documents with metadata filtering.

The Importance of doc_id

When indexing documents, the doc_id is a unique identifier for a logical document (e.g., a specific PDF file).

• Logical Grouping: A single PDF might be split into 50 chunks. All 50 chunks must share the same doc_id.
• Automatic Updates: If you upsert chunks with a doc_id that already exists in the collection, the server will automatically replace the old chunks with the new ones. This ensures you don't have duplicate content for the same document.
• Deletion: You can delete an entire document and all its associated chunks in one call using its doc_id.

from dochandler4ai_sdk import DocumentChunk

# Get document collection
docs = sdk.get_documents_collection(
    collection_id="my_docs",
    embedding_model="openai/text-embedding-3-small/1536" # Default for this collection
)

# Upsert documents
await docs.upsert(chunks=[
    DocumentChunk(
        doc_id="doc_001",
        page_content="Artificial Intelligence is transforming industries...",
        metadata={"category": "technology", "author": "Alice"}
    )
])

# Count documents
count = await docs.count(where={"category": "technology"})
print(f"Total technology documents: {count['total']}")

# Get sample documents
samples = await docs.get(k=5, where={"author": "Alice"})

Searching with Metadata Filters

The SDK supports complex metadata filtering using MongoDB-like operators ($or, $in, $gt, $lt, etc.).

from dochandler4ai_sdk import DocumentSearchRequest

req = DocumentSearchRequest(
    query="How is AI changing the world?",
    k=3,
    filters={
        "$or": [
            {"category": "technology"},
            {"tags": {"$in": ["AI", "ML"]}}
        ]
    }
)

results = await docs.search(req)

Self-Query Search

Let the LLM automatically translate natural language into structured filters.

req = DocumentSearchRequest(
    query="Show me documents by Alice about technology written after 2023",
    k=5,
    search_type="self_query",
    metadata_field_info=[
        {"name": "author", "description": "The author of the document", "type": "string"},
        {"name": "category", "description": "The document category", "type": "string"},
        {"name": "year", "description": "Year of publication", "type": "integer"}
    ]
)
results = await docs.search(req)

Image Indexing

Multi-modal vector storage for images.

from dochandler4ai_sdk import ImageItem

images = sdk.get_images_collection(
    collection_id="product_catalog",
    embedding_model="nvidia/llama-nemotron-embed-vl-1b-v2/2048"
)

# Upsert images
await images.upsert(images=[
    ImageItem(
        image_id="img_101",
        image_url="https://example.com/headphone.jpg",
        metadata={"category": "headphone", "brand": "Sony"}
    )
])

# Count images
count = await images.count(where={"category": "headphone"})

Multimodal Search

Search images using either a text query (Text-to-Image) or another image (Image-to-Image).

from dochandler4ai_sdk import ImageSearchRequest

# 1. Text-to-Image Search
req = ImageSearchRequest(
    query="blue wireless headphones",
    where={"brand": "Sony"},
    score_threshold=0.7
)
results = await images.search(req)

# 2. Image-to-Image Search
req = ImageSearchRequest(
    image_url="https://example.com/reference_image.jpg",
    k=5
)
results = await images.search(req)