Skills
Default RAG stack: BGE-M3 + BGE-reranker-v2 with hybrid retrieval for 25–40% precision gain
Most production RAG stacks in 2026 default to BGE-M3 embeddings paired with BGE-reranker-v2, and wrapping naive retrieval with hybrid (dense+sparse) search plus LLM/cross-encoder reranking improves precision 25–40% at modest added latency and cost. Pre-filter with metadata before the vector search to cut noise, and cache both embeddings and rerank calls. The actionable order of operations: metadata filter → hybrid retrieve → rerank, not just 'embed and top-k'.
Source
↳ Follow the thread