Evidence-Grounded RAG for Tokenized Trade Receivable Disclosure QA under U.S. Capital Market Standards

Abstract

Tokenized trade receivables and other real-world-asset products expose investors to short-dated credit, dilution, servicing, transfer, and disclosure risks. A question-answering system for these instruments must therefore return not only fluent text, but also evidence-grounded answers that map to issuer reports, accounting line items, and disclosure controls. This paper implements and evaluates an evidence-grounded retrieval-augmented generation pipeline on the assigned 2023 Data Set A, FinanceBench, a 150-question open-book financial QA sample. The evaluation uses every question in the CSV file and splits multi-evidence examples into 189 passage units. We compare BM25, dense latent semantic retrieval, hybrid retrieval, and a hybrid retriever followed by a deterministic financial reranker and citation-constrained extractive answerer. The full experiment is reproduced by the included scripts, which compute answer support, citation accuracy, faithfulness, and hallucination-risk metrics from the dataset fields. Hybrid+Reranker obtains the strongest evidence retrieval result, with Recall@3 of 0.460, MRR of 0.389, and citation accuracy@3 of 0.460, compared with BM25 Recall@3 of 0.313 and MRR of 0.273. The best pipeline also reaches answer accuracy@3 of 0.393 under the paper's support-based criterion, while the accepted-answer error rate remains high for calculation-heavy and receivable-adjacent questions. The results show that citation grounding improves investor-facing disclosure QA, but also show that retrieval alone does not solve numerical reasoning or multi-passage receivable due diligence.