PRISM: A Training-Free Financial Retrieval Framework Improving Recall and Cost Efficiency

01 ABSTRACT

PRISM is a training-free financial retrieval framework composed of three deployable modules: prompt engineering, in-context learning, and candidate refinement. On FiQA-2018, stacked reranking and k-reciprocal encoding raise Recall@100 from 0.7783 to 0.8326; on FinAgentBench, a Cohere-refined top-50 pool yields a higher observed end-to-end NDCG@5 that is below the evaluation's minimum detectable effect, while cutting per-query cost by about 35%. The authors also release an expert-annotated FinAgentBench evaluation set and a label-robustness sweep over 16 alternative ground-truth constructions, and note that added complexity does not always pay off: k-reciprocal encoding stops helping when the upstream stage is already high-recall, multi-agent coordination is not a default for fine-grained chunk ranking, and matched pairs show multi-agent coordination loses 0.0702 and 0.0681 combined NDCG@5 at the chunk stage.

02 KEY FINDINGS

1. PRISM requires no fine-tuning and raises Recall@100 on FiQA-2018 from 0.7783 to 0.8326 via stacked reranking and k-reciprocal encoding.
2. On FinAgentBench, a Cohere-refined top-50 pool increases end-to-end NDCG@5 but below the minimum detectable effect, while cutting per-query cost by about 35%.
3. Multi-agent coordination is not default-better for fine-grained chunk ranking; matched pairs show combined NDCG@5 losses at the chunk stage.