Long-Document RAG for Contractual and Insurance Clause Analysis in Receivables RWA Structures

Abstract

Receivables real-world-asset (RWA) structures require repeated verification of master receivables purchase agreements, insurance policies, repurchase language, default triggers, and dilution-risk provisions. These clauses are difficult for ordinary question answering systems because the relevant evidence is dispersed across long documents, tables, definitions, and numerical schedules. This paper presents a reproducible long-document retrieval-augmented generation (RAG) evaluation for clause analysis using a FinLongDocQA-RWA evaluation layer that preserves the public FinLongDocQA schema, row count, company coverage, fiscal-year range, and question-type distribution. The benchmark contains 7,527 financial long-document QA instances over 489 companies and fiscal years 2022-2024, with 5,951 mixed, 1,319 table, and 257 text instances. We compare a long-context baseline, BM25 RAG, TF-IDF/SVD RAG, hybrid lexical-semantic RAG, hybrid RAG with clause filtering, and the proposed RWA-ClauseLongRAG pipeline. RWA-ClauseLongRAG combines adaptive chunking, hybrid retrieval, clause-level extraction, page-evidence aggregation, and deterministic answer reconciliation. Across all 7,527 instances, it obtains answer F1 of 0.821, evidence recall of 0.843, faithfulness of 0.470, exact match of 0.654, and mean latency of 347.6 ms. Relative to Hybrid-RAG+Clause, the proposed pipeline raises answer F1 by 0.091 and evidence recall by 0.096 while keeping latency under 350 ms. Ablations show that 512-token chunks and top-k evidence between 8 and 12 give the best accuracy-latency trade-off. The results support clause-aware RAG as a practical RegTech design for auditable receivables RWA analysis.