History — 2026 week 39 (Sep 21 – Sep 27, 2026)
Newest first.
2026-09-26 — Query routing, from web verticals to RAG (18 new, 23 updated)
Routing a query means choosing where it goes — a source, a retriever, an embedding model, or whether to retrieve at all — and web search met an early version in 2009, picking which verticals a query should trigger (Sources of Evidence for Vertical Selection). Query Routing now traces that line into RAG: complexity routing, expert encoders, ranked retrievers and per-query hybrid weights, plus the practice: a cheapest-first cascade from regex to LLM, and Qdrant’s classifier sending each query to sparse, dense or Reciprocal Rank Fusion. One paper’s complexity router, only 55% accurate, still matched always-multi-step answers at under half the time. Paper figures are the authors’ own; the tutorials and Qdrant’s talk report no measured gains.
Articles — Sources of Evidence for Vertical Selection · Adaptive-RAG - Learning to Adapt Retrieval-Augmented LLMs through Question Complexity · RouterRetriever - Routing over a Mixture of Expert Embedding Models · LTRR - Learning To Rank Retrievers for LLMs · Efficient Federated Search for RAG using Lightweight Routing · Lightweight Query Routing for Adaptive RAG - A Baseline Study on RAGRouter-Bench · DAT - Dynamic Alpha Tuning for Hybrid Retrieval in RAG · Build an Advanced RAG App - Query Routing · Query Routing - Direct Queries to the Right Source Videos — Andrei Cristea - Qdrant Vector Search and Hybrid Routing Concepts — Adaptive Retrieval (whether and how much to retrieve) · Vertical Selection People — Fernando Diaz · Jaime Arguello · Jamie Callan · Roger Oriol · Andrei Cristea Companies — Ailog (hosted RAG vendor) Updated — Query Routing (hub rewrite: routes, labels, evaluation, IR roots) · Query Classification · Federated Search · Federated vs Unified Search · Linear Score Combination · Reciprocal Rank Fusion · Hybrid Search · Search Governance · RRF is Not Enough · RAG · Agentic Search · Learning to Rank · LLM as Judge · LoRA · FLAN-T5 · BEIR · MS MARCO · Natural Questions · LlamaIndex · Sentence Transformers · Sami Maameri · Qdrant · Haystack EU (Plain Schwarz as co-organizer)
2026-09-21 — Jev’s calibration claim, from the evaluation side (2 new, 5 updated)
A model whose selling point is a calibrated probability invites one question: calibrated to what? Jev and the Return of AI-ML Engineering puts that question to Jev and takes apart the four properties its launch leads with, arguing three of them — structured outputs, low latency, low cost — already come off the shelf, since for small prefills the cost and latency gap between autoregressive and non-autoregressive models is negligible. That leaves calibration, where Han-chung Lee reports high calibration error over a coin toss, two dice and three UCI datasets, and suspects the working demonstrations owe more to the base model than to RLCD. The experiments are described, not published; his own figures, he notes, establish nothing about predictive value.
Articles — Jev and the Return of AI-ML Engineering People — Han-chung Lee Updated — Reception of Jev · Jev · System One Model · Expected Calibration Error · Reinforcement Learning for Calibrated Decisions