History — 2026 week 35 (Aug 24 – Aug 30, 2026)
Newest first.
2026-08-28 — RelativeDB’s Relational Transformer Reranks on Structured Facts (6 new, 4 updated)
A reranker that scores candidates from typed database fields instead of serialized text: Relational Transformer (RT), an 85M-parameter model from RelativeDB, represents each fact as a Cell — name, type, value — connected to linked rows through schema relationships like also_buy. RelativeDB’s introduction reports RT reaching 0.633 NDCG@10 on MTEB DeepPlanning’s product-search slice (up from 0.286 for MiniLM retrieval alone), and on STaRK shows the also_buy edge alone moving the correct product from rank #2 to #1. The design follows RelBench’s finding that relational structure beats tuned LightGBM baselines on 11 of 15 tasks. An LLM query-parsing step that converts free text into Cells reportedly doubled accuracy.
Articles — Relational Reranking - Scoring Search Results with Structured Facts Concepts — Relational Transformer People — Daniel Henneberger Companies — RelativeDB Datasets — RelBench · STaRK Updated — Reranking · RankLLaMA · ColBERT · LightGBM
2026-08-28 — How Vector Databases Handle Updates and Deletes (2 new, 6 updated)
Vector search benchmarks that only index once and query a static corpus miss what happens as records constantly turn over. Updating a Vector Database Is No Simple Thing (Doug Turnbull) argues an update is usually a delete-then-insert upsert, and outright deleting a node from an HNSW graph would leave navigational gaps — so engines tombstone it instead, keeping edges intact but excluding it from results. Vector Index Updates compares vendor strategies: Vespa updates a single graph concurrently in place, Weaviate tombstones without parallel segments, and Qdrant and Milvus keep a mutable insert structure separate from the read-optimized graph. Turnbull’s conclusion: never trust a benchmark that just inserts once and replays queries — shadow-test against real traffic instead.
Articles — Updating a Vector Database Is No Simple Thing Concepts — Vector Index Updates Updated — HNSW · Vespa · Weaviate · Qdrant · Milvus Vector DB · Doug Turnbull
2026-08-28 — Kendall Rank Correlation as a Reranking Diagnostic (1 new, 4 updated)
Two rankings of the same candidate set can be compared without relevance labels: Kendall Rank Correlation counts concordant versus discordant pairs between an input and output order, ranging from +1 (identical) through 0 (unrelated) to -1 (reversed). The note contrasts this with NDCG and MAP, which score a ranking against Judgment Lists — tau measures how much an ordering changed, not whether the change helped. Metarank logs it as krr on every rerank request, a server-side diagnostic never returned to the caller; a low value confirms the Learning to Rank model is reshuffling candidates rather than echoing input order. Doug Turnbull’s judgment-list FAQ separately uses it to check agreement between judgment systems.
New — Kendall Rank Correlation Updated — Metarank · Reranking · Learning to Rank · NDCG