History — 2026 week 40 (Sep 28 – Oct 4, 2026)
Newest first.
2026-10-04 — autofaiss: FAISS index choice under budgets (5 new, 3 updated)
Choosing a FAISS index means picking among hundreds of combinations, each with up to six build hyperparameters, and its authors say most users end up with sub-optimal ones. autofaiss, from Criteo, turns that into a constrained search: give it a memory cap and a query-time cap, and it returns the FAISS index and parameters with the highest recall inside them, leaning on product quantization. On a 1 TB corpus of 200 million image embeddings, its authors report a 10 GB index built in 3 hours and answering in about 10 ms, with no recall figure given. Vector Search Tradeoffs now points to it for automated index choice. From the launch post by Victor Paltz and Romain Beaumont.
Articles — Introducing Autofaiss - An Automatic K-Nearest-Neighbor Indexing Library At Scale People — Victor Paltz · Romain Beaumont Companies — Criteo Tools — autofaiss (automatic FAISS index selection) Updated — FAISS · Vector Quantization (product quantization detail) · Vector Search Tradeoffs
2026-10-03 — Elasticsearch API served from Qdrant (1 new, 2 updated)
An Elasticsearch application can keep its client code and change the engine underneath it. qdrant-es-gateway is a Rust gateway that does this for Qdrant: it serves a subset of the Elasticsearch REST API, scores text with Qdrant’s server-side BM25 sparse model, and rejects unsupported features such as scripts and scroll with explicit errors. On its author’s own run over 500,000 generated products at 50 clients, it searched faster than Elasticsearch, used about 43× the storage, and did not reproduce Elasticsearch’s top-ranked results. Migration between Search Engines now covers API-compatibility gateways as an option alongside rewriting clients. Built by David Stockton.
Corrections
- Migration between Search Engines said a pure vector DB has no lexical BM25 tier → it now notes that Qdrant ships server-side BM25 over sparse vectors, with non-Lucene scoring.
Tools — qdrant-es-gateway (Elasticsearch-compatible gateway over Qdrant) Updated — Migration between Search Engines · Qdrant Vector DB
2026-10-03 — Diversity built into inner-product search (4 new, 8 updated)
Top-k inner-product search returns the most relevant vectors and, left alone, the most alike ones: in one MovieLens example, a user with ten rated genres got recommendations from five. Diversity-Aware k-Maximum Inner Product Search Revisited covers DkMIPS, which moves the MMR trade-off inside Maximum Inner Product Search — a λ dial between relevance and average or maximum pairwise similarity, greedy algorithms with a 1/4 approximation guarantee for the average form, and tree-based pruning, since standard ANN indexes don’t fit the pairwise term. DiversiNews, its news demo over the NewsSpectrum corpus, measures diversity as a spread of outlet bias ratings. By Qiang Huang, Yanhao Wang, Yiqun Sun and Anthony K.H. Tung; all performance figures are the authors’ own.
Articles — Diversity-Aware k-Maximum Inner Product Search Revisited (DkMIPS, with DiversiNews) Concepts — Maximum Inner Product Search People — Qiang Huang Datasets — NewsSpectrum (news balanced across political bias ratings) Updated — MMR (approximation guarantees) · Search Result Diversity · APD · Approximate Nearest Neighbor Search · Vector Similarity Metrics · Yiqun Sun · Anthony K.H. Tung · NEWSCOPE
2026-10-01 — ESCI as a label scheme (1 new, 5 updated)
A product search result can be the thing asked for, a stand-in for it, an accessory to it, or unrelated — and binary relevance cannot tell those apart. ESCI Label Scheme covers the Exact / Substitute / Complement / Irrelevant grading introduced with the Amazon ESCI Dataset and how it has travelled without the data: Allegro’s judgment guidelines were inspired by it, Etsy plans the same substitute/complement split, and sQuIrRel’s relevance model predicts its labels. It also covers where the grades are hard: in Allegro’s tests, an LLM judge reached F1 around 0.94 on exact matches but only 0.33–0.51 when separating degrees of substitutability.
Concepts — ESCI Label Scheme Updated — Amazon ESCI Dataset · E-commerce Search Evaluation Datasets · Semantic Relevance · Judgment Lists · sQuIrRel - Large-Scale Evaluation of E-commerce Query Classification Models
2026-10-01 — sQuIrRel: query-classifier labels from relevance (1 new, 3 updated)
Evaluating a query-to-product-type classifier needs labels for thousands of product types, and neither clicks nor annotators supply them well. sQuIrRel, from Amazon, derives them from relevance instead: a high-precision model keeps each query’s exact-match results, and the query inherits their dominant catalogue product type. A manual check of 400 queries found 87% of labels correct, with complements as the main error. The paper publishes only the procedure, built on internal Amazon data. Its limits: it skips multi-intent queries and can only label queries where the engine already returns the right products.
Articles — sQuIrRel - Large-Scale Evaluation of E-commerce Query Classification Models Updated — Query Classification · E-commerce Search Evaluation Datasets · Bag-of-Documents Model (as a source of evaluation labels)
2026-10-01 — E-commerce search evaluation datasets compared (1 new, 6 updated)
Public product-search datasets differ less in size than in what a relevance label means. ESCI separates substitutes from complements, WANDS folds both into one Partial grade, and Home Depot averages annotator scores on a continuous scale. E-commerce Search Evaluation Datasets compares them alongside ESCI-S and its decaying image links, and matches each to the catalogs and query styles it suits. It also collects the cross-catalog evidence: in one fine-tuning series, an ESCI-tuned SPLADE model gained 27.5% nDCG@10 over BM25 on ESCI itself, but only 8–10% on WANDS and Home Depot.
Topics — E-commerce Search Evaluation Datasets Updated — E-commerce Search · Retrieval Benchmarks and Leaderboards · Amazon ESCI Dataset · ESCI-S Dataset · WANDS Dataset · Home Depot Product Search Relevance