Route RAG queries by complexity — Adaptive RAG is the 2026 practitioner consensus

Atlan·low signalXBlueskyLinkedInCopy link

The emerging production default is a query classifier that routes each request to a different pipeline based on complexity, rather than running one retrieval path for everything. Two supporting numbers reframe where to spend effort: when RAG fails, the failure is in retrieval 73% of the time rather than generation, and 60% of new 2026 deployments build systematic evaluation in from day one (up from under 30% in early 2025). Nugget-based evaluation is becoming the de facto scoring strategy. Instrument retrieval first — generation-side prompt tuning is chasing the smaller share of failures.