Extract data from PDF in Python: Text, tables, and key-value pairs

The ICR and VLM-enhanced ICR engines analyze document layout using AI to detect table regions, identify row and column boundaries, and extract individual cell contents. This works on both native PDFs and scanned documents, including tables without visible borders. The output is structured JSON with each cell mapped to its row and column position, making it straightforward to convert to CSV, database records, or any structured format your application needs.

No. Data Extraction analyzes document structure on the fly using AI. It detects tables, key-value pairs, headings, and other elements without predefined templates or training data. This means it works on invoices from any vendor, receipts in any format, and forms with any layout. The AI adapts to each document individually, eliminating the setup and maintenance cost of template-based extraction systems.

Yes. All three engines process both native PDFs and scanned documents. The OCR engine extracts text with bounding boxes from scanned images. The ICR engine goes further, applying AI to understand document structure and extract tables, key-value pairs, and semantic elements from scans. The VLM-enhanced engine adds vision language model processing for maximum accuracy on challenging scanned documents with poor quality or complex layouts.

Data Extraction returns structured JSON containing classified document elements. Each element includes its type (table, paragraph, heading, equation, key-value region, figure), text content, bounding box coordinates, and confidence score. Tables include cell-level data with row and column positions. The JSON also includes the document reading order so you can reconstruct the logical flow of content. You can save output directly to a JSON file or process it in memory.

OCR provides fast, reliable text extraction and is best for simple documents where you need raw text and positions. ICR adds AI-powered structural analysis and significantly improves results on documents with tables, mixed layouts, and form-like content — all while running entirely offline. VLM-enhanced ICR delivers the highest accuracy by combining local AI with a vision language model, particularly for complex financial documents, legal contracts, and pages with irregular table structures.

Yes. The OCR and ICR engines run 100 percent on your infrastructure with no external API calls, and documents never leave your servers. This meets requirements for HIPAA, GDPR, SOC 2, and air-gapped environments. For VLM-enhanced extraction, you can also stay fully on-premises by connecting to a local model server like Ollama, LM Studio, or vLLM instead of cloud providers.

Camelot and pdfplumber are popular open source Python libraries for extracting tables from native PDFs. They parse the underlying text layer and work well on digitally created documents with clear table borders. However, they cannot process scanned documents and struggle with borderless or irregular tables. Nutrient Data Extraction uses AI to detect tables, regardless of borders; works on both scanned and native documents; and also extracts key-value pairs, equations, and full document structure — returning everything as classified JSON with confidence scores.

Data Extraction processes PDFs (both native and scanned) and common image formats, including PNG, JPEG, GIF, BMP, and TIFF. PDF pages are automatically rendered for processing, so you don’t need to handle conversion. All three engines accept the same input formats, making it easy to switch between them based on your accuracy and performance needs.

Yes. The ICR and VLM-enhanced engines are well suited for invoice and financial document processing. They detect tables with line items and totals; identify key-value pairs like vendor names, dates, and amounts; and understand the hierarchical structure of a document. Because the extraction is template-free, it works across different invoice formats and vendors without per-vendor configuration. Use VLM-enhanced mode for the highest accuracy on complex financial layouts.

Install Nutrient Python SDK with pip and follow the getting started guide. For data extraction, open a document, create a vision instance, select your engine (OCR, ICR, or VLM-enhanced), and call the extraction method. The SDK returns structured JSON that you can parse with standard Python tools. The extraction guides include step-by-step examples for each engine and use case.

Yes. Nutrient Extract Studio is a desktop app for Apple silicon Macs that ships with its own Python runtime and the Nutrient Python SDK, so there’s no environment to set up. Run parse, extract, or classify on a sample document, and then view the code in the app to copy Python code for the Data Extraction SDK into your project.