SIRA

SIRA is a multi-stage retrieval pipeline that uses LLMs to enrich both documents and queries, improving BM25 retrieval quality without training. The pipeline consists of five stages: data preparation, BM25 indexing, corpus enrichment (LLM-generated indexing phrases for documents), query expansion (LLM-generated search terms), and LLM-based pointwise reranking. SIRA achieves state-of-the-art results on BEIR benchmarks using only inference-time compute.