A new retrieval architecture for financial document question-answering beats both of the dominant approaches on every measured metric.
Researchers tested three retrieval strategies on FinDER, a benchmark of 1,500 queries over regulatory filings. Chunk-based retrieval (CBR) splits documents into fragments and retrieves by similarity — fast and precise, but prone to mixing up chunks from different filings, producing a 22.5% failure rate. Semantic File Routing (SFR) sidesteps that by using an LLM to route each query to a whole document first, cutting failures to 10.3% and lifting the average score from 6.02 to 6.45 — but it sacrifices the targeted precision that makes CBR produce perfect answers 13.8% of the time (versus SFR's 8.5%). Hybrid Document-Routed Retrieval (HDRR) runs both stages in sequence: SFR narrows the search to the right document, then chunk retrieval operates within that scope.
The combined approach scores 7.54 on average — 25.2% above CBR and 16.9% above SFR — with a 6.4% failure rate and 67.7% correctness, a gain of 18.7 percentage points over CBR alone. That matters because financial RAG is a domain where a wrong answer to a regulatory query isn't just an inconvenience. More notable is the cost profile: HDRR keeps per-query token use in the 5K-15K range, the same compact budget as CBR and an order of magnitude below SFR's 50K-200K, while avoiding the one-time roughly $100 indexing cost of contextual indexing approaches.
The paper is an academic preprint, so peer review is pending — but the benchmark numbers make a straightforward case that the robustness-versus-precision trade-off in financial RAG is an architecture problem, not an inherent one.