You need to validate that adding table-aware retrieval...

The choice of benchmark must align with the production task. For financial filings, you need datasets that include long, multi-document contexts and complex table structures. Popular benchmarks like MRAG-Bench are often unsuitable for this specific use case because they lack the necessary complexity to test multi-document table retrieval. Additionally, use ablation studies to confirm that performance gains are specifically due to table-aware retrieval rather than just increased context length.