HIRA Boosts Document Classification in Regulated Industries.

Document classification in highly regulated sectors faces unique challenges, including strict data residency rules, limited initial labels, scarce human review capacity, and complex model governance. Traditional machine learning often requires costly retraining and extensive labeled data, which is impractical in these environments. HIRA (Human-in-the-Loop Retrieval-Augmented Cascade) offers a novel solution. It's a training-free, on-premises system that integrates BM25 over OCR text, dense text embeddings, and image-level representations, fused using validation-calibrated reciprocal-rank weighting. Confident documents are classified directly by retrieval. Uncertain or visually ambiguous documents are routed to a locally hosted LLM verifier, which receives relevant context like OCR text, retrieved examples, and label descriptions. If the LLM remains uncertain, the document is sent for human review. Crucially, every human correction is stored as a margin-weighted retrieval exemplar, and a Dirichlet-smoothed confusion graph is updated, allowing the system to continuously improve without needing to retrain core model weights. This approach significantly boosts accuracy while drastically reducing human intervention and LLM calls.