Dharma Field Extractor Model Released for Structured Data Extraction from Documents

Introduction of Dharma Field Extractor

A new model named Dharma Field Extractor has been introduced, designed for the specific task of extracting structured data from various document types. This 2-billion-parameter model processes document images and returns requested fields in a structured JSON format, including typed and nested data where applicable.

Addressing Data Extraction Challenges

Traditional Optical Character Recognition (OCR) provides raw text from documents, which often requires additional coding to map specific values to database fields. While general-purpose Large Language Models (LLMs) can perform this mapping with careful prompting, their API pricing and latency make them expensive for high-volume business needs. The Dharma Field Extractor was developed to provide a more efficient and cost-effective solution for this well-defined problem.

Performance and Cost Efficiency

The Dharma Field Extractor reportedly outperforms frontier LLMs in data extraction tasks. It operates at a fraction of the cost, making it suitable for businesses that need to process large batches of documents, such as invoices, medical requisitions, or contracts, to feed into their operational systems. The model handles diverse document types without requiring per-type fine-tuning.

Functionality and Use Cases

Users can provide the Field Extractor with a document image and either request all discoverable fields or specify an exact schema to fill. For example, an invoice can return line items as an ordered list with quantities and values, alongside totals. An ID can return fields mapped to existing system column names. This functionality supports various business processes, including automating data entry for supplier invoices, scheduling medical appointments, or populating compliance systems.