ORDER: query-conditioned routing for RAG beats SOTA on heterogeneous archives
_reachsumit · x · 2026-09-16
A new arXiv paper introduces ORDER (Optimal Routing for Dynamic Evidence Retrieval), a query-conditioned RAG framework:
- Problem: typical RAG pipelines fix indexing/retrieval at preprocessing time, but heterogeneous expert queries need different chunking granularities, metadata constraints, and source selection.
- Method: discover semantic clusters over questions, learn per-cluster chunking plus metadata filtering/reranking configs; route queries to pre-built indices via nearest-centroid assignment. A supervised query router (QRe) predicts likely relevant collections, with a Uniform Multi-source Sampler (UMS) spreading retrieval budget evenly.
- Results: on large-scale heterogeneous historical archives, conditioning both indexing and retrieval on the query consistently outperforms naive baselines and strong SOTA RAG systems.
Paper and code linked in the post.
More from coding & agent
- Robot arm Podrick learns to load utensils by distilling lessons from its own failed runs — ycombinator · 2026-10-23
- Agents that don't close connections on retry waste ~148 tokens per task, vLLM tests show — No-Peanut-6988 · 2026-10-23
- Agent sandbox test: auto-approve let exfiltration to new hosts through 12 of 12 times — No-Peanut-6988 · 2026-10-23
- For writing JSON schemas, Claude is the best pick and Gemini the worst, dev says — gaganghotra_ · 2026-10-23
- OSI Hackathon winner Consilith: an open multiplayer studio where researchers and their AI agents collaborate on science — BenBlaiszik · 2026-10-23
- Claims Engine: a graph-based system connecting evidence to claims — anselm · 2026-10-23