Efficient Semantic Retrieval via Multilingual Embeddings and Reranking

Retrieving semantically relevant information from large-scale multilingual corpora remains a fundamental challenge in information retrieval. While dense retrieval with multilingual sentence encoders enables scalable matching, it often lacks the granularity required for accurate top-k ranking, particularly in domain-specific or noisy query contexts. This paper presents a compact, cascaded retrieval framework that integrates dense multilingual encoders with a lightweight reranking module, all operating under a sub-7B parameter constraint. Evaluations are conducted on two representative English subsets of the BEIR benchmark, FIQA (financial-domain, long-form queries) and Quora (short, general-domain paraphrases). Results demonstrate that reranking provides substantial improvements for mid-tier models in challenging settings (e.g., +13% MAP@10 for e5-base on FIQA), and modest but consistent gains even in cleaner retrieval scenarios. Through ablation on reranking components, we show their critical role in refining semantic alignment, particularly for models with weaker initial rankings. Overall, the proposed framework achieves robust, high-accuracy retrieval with minimal computational overhead, offering a practical solution for scalable deployment in real-world, resource-constrained environments.