RAG paper finds stronger retrieval beats bigger generators, proposes DRAG query-adaptive routing
_reachsumit · x · 2026-09-17
Researchers systematically analyze how retriever and generator complexity interact in RAG and propose DRAG, a query-adaptive configuration framework.
Key findings
- Fixed retriever+generator setups ignore query complexity differences, wasting compute
- Stronger retrieval generally yields larger gains than extra generator reasoning, but both show diminishing, non-monotonic returns across queries
- Validated on factoid and multi-hop QA, including bridge and composition reasoning
Methods
- DRAGQPP: training-free routing using Query Performance Prediction signals for retriever selection and perplexity-based measures over retrieved context for generator selection
- DRAGSFT: supervised routing that fine-tunes an LLM to jointly predict retriever-generator configurations
Evaluated across three LLM families and four QA benchmarks.
More from coding & agent
- GitHub flags LobeHub org, hiding 82K-star open-source project from public view — dotey · 2026-09-17
- Don't use an agent where a sequence works: the biggest AI-building mistake, per Hugo Bowne — hugobowne · 2026-09-17
- Plasma AI launches Radio: a shared chat room for Claude, Codex and Grok agents — rohanpaul_ai · 2026-09-17
- UIUC Launches Center for Orchestrated Agentic Biology to Study Brain Aging with Multi-Agent AI — dotey · 2026-09-17
- Dev uses LLMs to extract structured data from corporate financing documents — mayfer · 2026-09-17
- Open-source Claude skill builds MCP servers end-to-end via mcpfy SDK — ZealousidealTax42 · 2026-09-17