retrieval progression is BM25 then learned-sparse then dense then hybrid fusion with re-ranking. BM25 is the deterministic robust zero-shot floor and can be built bits-up before any local language model exists. Learned-sparse retrievers like SPLADE uniCOIL and DeepImpact store term-impact weights and can fit inside an inverted-index postings layout as term frequencies. Dense retrieval uses sentence embeddings; ColBERT late-interaction stores per-token vectors for high quality at higher cost. Hybrid search fuses lexical and dense rankings with Reciprocal Rank Fusion or weighted fusion, then a cross-encoder reranker refines the top candidates. Benchmark retrieval quality on BEIR which spans 18 datasets and on MS MARCO using normalized discounted cumulative gain nDCG and mean reciprocal rank MRR. In the 2021 BEIR study BM25 was a robust zero-shot baseline and re-ranking and late-interaction models gave the best zero-shot quality at high computational cost. Post-2021 instruction-tuned dense models such as E5 BGE-M3 and GTE often beat BM25 out-of-distribution by 15 to 25 percent, so treat BM25 as the deterministic first stage and hybrid plus rerank as the quality target. Ranking effectiveness does not scale monotonically with training-data volume, so more data alone does not unlock quality.