History — 2026 week 43 (Oct 19 – Oct 25, 2026)
Newest first.
2026-10-22 — BlitzRank: cheaper LLM reranking via tournament graphs (6 new, 9 updated)
LLM reranking gets expensive when results are compared a pair at a time. BlitzRank, from Contextual AI, asks the LLM to order a small window and reads that answer as every pairwise preference inside it, infers more pairs by transitivity, and stops once the top ten are provably settled. Averaged over 14 TREC DL and BEIR datasets and two LLMs, it scored 56.9 nDCG@10 on 42k tokens per query, against 57.0 and 324k for a pairwise reranker. Doug Turnbull has pitched it as a search-evaluation judge, but the paper tests it only as a reranker and assumes noise-free judgments. Sourced from the paper.
Articles — BlitzRank - Principled Zero-shot Ranking Agents with Tournament Graphs (the arXiv paper) Concepts — BlitzRank People — Sheshansh Agrawal · Thien Hang Nguyen · Douwe Kiela Companies — Contextual AI Updated — Reranking · Reasoning Reranking · RankGPT · Pairwise Relevance Evaluation · Listwise Relevance Evaluation · LLM as Judge · TREC Deep Learning Track · BEIR · Doug Turnbull (LinkedIn pitch for BlitzRank)
2026-10-20 — MUVERA: multi-vector retrieval through one vector (17 new, 12 updated)
Multi-vector models like ColBERT score with MaxSim, which ANN indexes can’t search directly. MUVERA folds each document’s token vectors into one fixed-dimensional encoding whose inner product approximates MaxSim, retrieves on that, then reranks with the real vectors. MUVERA now covers the Google paper’s construction and its comparison with PLAID: Recall@k 10% higher and latency 90% lower on average over six BEIR datasets, with no win on MS MARCO. It also covers implementations in Weaviate Vector DB, Qdrant Vector DB via FastEmbed, and an OpenSearch plugin. Its benchmarks on three small corpora keep 87–97% of exact quality after rerank; Elastic’s approach, average vectors, keeps 28–52%. All figures are the authors’ own.
Articles — MUVERA - Multi-Vector Retrieval via Fixed Dimensional Encodings (the paper) · More Efficient Multi-Vector Embeddings with MUVERA · MUVERA in Qdrant - Multi-Vector Search Course · MUVERA - Transforming Multi-Vector Information Retrieval Through Fixed Dimensional Encodings (paper walkthrough) · Scaling Late Interaction Retrieval in OpenSearch with MUVERA · Muvera - Making Multi-Vector Retrieval as Fast as Single-Vector Search (HN Discussion) Concepts — PLAID (ColBERTv2’s multi-vector engine) People — Laxman Dhulipala · Majid Hadian · Rajesh Jayaram · Jason Lee · Vahab Mirrokni · Roberto Esposito · Keyur Ramoliya Companies — Google Tools — FastEmbed · opensearch-muvera-plugin Updated — MUVERA (construction, benchmarks, engine support, vs average vectors) · Late Interaction · Late Interaction in Elasticsearch · Late Interaction in OpenSearch · Late Interaction in Qdrant · Weaviate Vector DB · Qdrant Vector DB · Weaviate · Qdrant · Amazon Web Services · Joon-Pil (JP) Hwang · LoTTE
2026-10-20 — Click residuals for finding underperforming queries (2 new, 6 updated)
When a query gets plenty of searches but few clicks, click residual puts a number on the shortfall: actual clicks minus attempts × the site’s average CTR. Ranking queries by residual brings high-impact failures to the top, where a plain low-CTR list would bury them in the long tail. Clicks Residual now covers Walter Underwood’s original definition with his Netflix cases. It also covers Charlie Hull’s e-commerce extensions: per-category CTR baselines, exit rate and business value for prioritising, and query clustering to catch whole failing families such as misspellings. The metric shows that a query fails, not why. Sourced from Underwood’s 2022 blog post and Click Residuals - A Clever Way to Find Underperforming Queries.
Corrections
- Click Residual - A Query Success Metric credited the metric to “Ted Underwood” and linked a dead Medium URL. It is by Walter Underwood, published on his own blog in 2022.
- That note’s position-discounted formula and “running shoes” example are not in the source. It now gives Underwood’s exp = att × ctr, res = clk − exp.
- Search Observability called click residual “position-bias corrected”. It uses a single average CTR baseline.
Articles — Click Residuals - A Clever Way to Find Underperforming Queries People — Walter Underwood Updated — Clicks Residual · Click Residual - A Query Success Metric · Netflix (Underwood’s batman/cj7 examples) · Charlie Hull · The Search Juggler · Search Observability