ProcureExtract-T5

About Model

About the Model ProcureExtract-T5 is a domain-specific Natural Language Processing (NLP) model designed to automate the extraction of structured procurement information from unstructured procurement emails. Built by fine-tuning Google's FLAN-T5-Base (250 million parameters), the model converts free-form procurement requests into standardized JSON output that can be directly consumed by enterprise applications, ERP systems, procurement platforms, and business process automation workflows. Organizations often receive procurement requests through emails containing multiple products, technical specifications, brands, quantities, and units of measurement. Since these emails vary widely in writing style, formatting, and terminology, manual extraction is labor-intensive and error-prone. ProcureExtract-T5 addresses this challenge by automatically identifying procurement items and generating structured output without relying on rule-based extraction systems. The model follows a sequence-to-sequence (Seq2Seq) architecture based on the FLAN-T5 encoder-decoder transformer. Unlike traditional Named Entity Recognition (NER) models that perform token classification, ProcureExtract-T5 directly generates a JSON array where each object represents one product. For every identified product, the model extracts five standardized fields: * product_raw_text – Product name as mentioned in the email * product_brand – Brand or manufacturer (if available) * product_specifications – Technical specifications or product attributes * order_qty – Ordered quantity * order_unit_of_measure – Unit of measurement If any field is unavailable in the source email, the model outputs an empty string rather than generating unsupported information, thereby reducing hallucinations and improving data reliability. The model has been fine-tuned using the Hugging Face Transformers framework on 1,304 curated procurement email samples, comprising real-world and augmented procurement data. Training employs supervised sequence-to-sequence learning with the Hugging Face Seq2SeqTrainer. A unique feature of the model is its count-injection prompting strategy, where the expected number of products is included in the instruction prompt during both training and inference. This structural guidance significantly improves the generation of complete multi-product JSON arrays. During inference, a lightweight product-count estimation module first predicts the number of products using heuristic techniques such as bullet-list detection, numbered-list detection, quantity-unit pattern recognition, and conjunction analysis. The estimated count is injected into the prompt before generation, enabling more accurate structured extraction. A multi-stage JSON parsing pipeline further improves robustness by recovering partially malformed outputs before returning the final structured response. Model performance is evaluated using a Hungarian Assignment-based field-level matching algorithm, which aligns predicted products with ground-truth products irrespective of ordering. The model achieves an 86% field-level F1 score on a held-out evaluation dataset and 98.4% product-count accuracy on real procurement emails, demonstrating reliable extraction across diverse procurement formats. ProcureExtract-T5 is suitable for procurement automation, purchase order digitization, ERP data ingestion, supplier communication analysis, intelligent document processing, supply chain analytics, and workflow automation. Since the model runs entirely offline after deployment, it offers a privacy-preserving alternative to cloud-based large language models, making it well suited for organizations with strict data security and compliance requirements. By combining instruction-tuned transformer capabilities with domain-specific fine-tuning, structured prompting, and robust post-processing, ProcureExtract-T5 provides an efficient and production-ready solution for transforming unstructured procurement emails into accurate, machine-readable procurement data.