Retrieval-Augmented Generation: What Changed in 2026
A survey of how RAG architectures evolved as context windows grew and retrieval quality improved.
The starting point
Early RAG systems retrieved a fixed number of document chunks and stuffed them into the prompt, with quality highly dependent on chunking strategy and embedding model choice.
What's improved
Adaptive retrieval — dynamically deciding how much to retrieve and re-ranking results before insertion — combined with much larger context windows has reduced the fragility of naive chunk-based approaches.
Open questions
Balancing retrieval freshness against cost, and handling contradictory information across retrieved sources, remain active problems without a settled best practice.