Domain-Adapted RAG for Financial Regulatory Compliance with Compact Models

A three-stage RAG pipeline using LegalBERT-based retrieval and RAFT fine-tuning enables 2B-12B parameter models to perform financial compliance QA. The approach uses entailment tuning and contrastive learning to reduce hallucinations in quantized, on-premise deployments.