Retrieval design influences which evidence a retrieval-augmented generation (RAG) system can access and how long that evidence takes to obtain. This study compares TF-IDF, BM25, a dynamically quantized MiniLM dense retriever, and reciprocal rank fusion (RRF) of BM25 and MiniLM on the complete BEIR FiQA test split. All 648 test questions were evaluated against 57,638 documents using the supplied relevance judgments. Outcomes were nDCG@10, the proportion of questions with a relevant document in the first five results (Hit@5), recall, reciprocal rank, and local CPU retrieval latency. Accuracy uncertainty was estimated with 10,000 paired query bootstrap samples; latency was measured over three randomized passes. Mean nDCG@10 was 0.179 for TF-IDF, 0.235 for BM25, 0.334 for MiniLM, and 0.337 for RRF. RRF minus MiniLM was +0.003 (95% confidence interval -0.015 to +0.022). Hit@5 ranged from 30.2% to 54.0%; median retrieval latency ranged from 5.74 to 23.92 ms. The results quantify implementation-specific effectiveness and latency trade-offs and show why fusion should be evaluated against its strongest constituent. FiQA is a public financial forum benchmark, and the evaluated model may have encountered overlapping training material. These findings concern retrieval relevance under the supplied judgments, not generated-answer accuracy or production enterprise performance.
