Default RAG chunks to 512 tokens with 10-20% overlap, then scale by query type
RAG Chunking Strategies & Embeddings Optimization: The 2026 Benchmark Guide·medium signal
A 50-document real-world benchmark in early 2026 found 512-token chunks with 50-100 token (10-20%) overlap scored the highest composite retrieval accuracy as a default. But tune by query shape: factoid lookups do well at 256-512 tokens, while analytical and multi-hop questions benefit from 512-1,024 token chunks that preserve more context per retrieval. Combined with hybrid retrieval and reranking — which lifts precision 25-40% over naive vector search — chunk sizing is the cheapest knob with the largest quality payoff.