Hybrid Retrieval
A retrieval pattern that combines **dense vector similarity** (semantic search via embeddings) with **sparse lexical search** (BM25 over an inverted index), merges the two ranked result sets using **Reciprocal Rank Fusion (RRF)**, and passes the fused candidates through a **cross-encoder reranker** for a high-precision final pass. The output: a small, deeply-relevant context set for the LLM, anchored both semantically and lexically.
Definition
A retrieval pattern that combines **dense vector similarity** (semantic search via embeddings) with **sparse lexical search** (BM25 over an inverted index), merges the two ranked result sets using **Reciprocal Rank Fusion (RRF)**, and passes the fused candidates through a **cross-encoder reranker** for a high-precision final pass. The output: a small, deeply-relevant context set for the LLM, anchored both semantically and lexically.
Pure vector search excels at concept matching but fails at exact-match scenarios — product SKUs, legal clause numbers, specific API names, regulatory references. Pure BM25 is the inverse: rock-solid on exact-match, blind to paraphrase. 2026 production retrieval systems run both in parallel and fuse them because either signal alone leaks recall in the cases the other handles best.
Enterprise RAG over regulated corpora (financial filings, legal contracts, medical records), code-aware retrieval where identifier-level precision matters, agentic memory systems requiring verifiable provenance against retrieved chunks, search over technical documentation with high jargon density.
Recent developments
- BGE-Reranker-v2-m3 is the current best open-weight reranker (English + multilingual). Based on MiniLM architecture (278M params) — small enough for CPU under 100-pair batches, fast on a single GPU for larger workloads. Per Local AI Master — Reranking & Cross-Encoders for RAG 2026.
- 2026 reranker landscape: BGE / Jina v2 / mxbai-rerank / bge-reranker-v2-gemma + Cohere + Voyage. Open + hosted options across English + multilingual. Two-stage pipeline (bi-encoder retrieval → cross-encoder reranker) is now the canonical 2026 production pattern. Per Markaicode — Build BGE Reranker: Cross-Encoder Reranking 2026.
- Quality lift from rerank: +5 to +15 NDCG@10 points on MTEB and BEIR. Often the difference between a usable RAG system and one that actually answers questions correctly. Reranking is not optional for production-grade retrieval. Per Local AI Master — Reranking Guide 2026.
- Canonical 2026 production pipeline: Dense + Sparse → RRF Fusion → Top 50-200 → Cohere/Voyage rerank → Top 5-10 → LLM. Each stage adds precision while keeping latency bounded. Cross-encoder takes the top 100 candidates from vector search, jointly attends over each query-doc pair, returns precision-tuned top 10. Per Medium — Why Re-Rankers Decide RAG Quality.
- Multilingual reranking expanding — ViRanker for Vietnamese as an example. 2025 arXiv paper covers ViRanker (BGE-M3 + Blockwise Parallel Transformer cross-encoder) for Vietnamese reranking — illustrates the broader pattern of language-specific reranker derivatives proliferating off the BGE-M3 base. Per arXiv 2509.09131 — ViRanker.
- HYRR: Hybrid Infused Reranking research direction. Academic work formalizing how hybrid retrieval signals (dense + sparse) inform reranking — the next architectural step beyond plain two-stage pipelines. Per arXiv 2212.10528 — HYRR: Hybrid Infused Reranking for Passage Retrieval.
Connections 8
Outbound 7
depends_on1solves1constrained_by1Inbound 1
enables1