History — 2026 week 33 (Aug 10 – Aug 16, 2026)
Newest first.
2026-08-12 — Query classification: the label set is the hard part (6 new, 21 updated)
Classifying a query — into an intent, a catalog category, a brand — fails less often at the model than at the seam between what it can emit and what the system will accept. Query Classification works that tension from both ends. Don’t Classify, Hallucinate has Doug Turnbull drop the taxonomy from the prompt entirely, letting a cheap model invent a fake category path that an embedding lookup snaps onto a real one. foodpanda - Classifying 300K Noisy Search Terms Across 16 Markets instead finds 37% of 300,000 terms sitting in an Others bucket — Turkish offal dishes and Lao script, not noise — repairs the label set, and ships deterministic rules with no model in the serving path. Figures are author-reported.
Corrections
- Ecommerce Search UX - 8 Query Types was written as twelve query types in three invented groups; Baymard Institute defines eight, each carrying the share of sites that fail it.
- Query Types attributed Comparative, Thematic and Branded to Baymard; those are not theirs, and the eight now carry their real names and failure percentages.
Articles — Don’t Classify, Hallucinate · Ecommerce Search UX - 8 Query Types Topics — Query Classification People — Hailey Cheong Companies — Delivery Hero Case Studies — foodpanda - Classifying 300K Noisy Search Terms Across 16 Markets
Updated — Query Types · Query Understanding · Query Understanding in Practice · Search Scopes · Hypothetical Document Embeddings (the taxonomy-shaped variant) · Semantic Search Without Embeddings · Classic ML to Cope with Dumb LLM Judges · Query Understanding - Query Scoping · Semantic Equivalence of e-Commerce Queries · You Say Search I Say Recs - Spotify Agentic Query Understanding · Metadata - The 3rd Kind of Retrieval · Targeting Broad Queries in Search · Facets of Faceted Search · Etsy - Search Quality and Query Understanding · Uber Eats - Scaling Search for Food Delivery · Search Problem Archetypes · E-commerce Search · Search UX · Baymard Institute · Doug Turnbull · Deconstructing E-Commerce Search - The 12 Query Types (now a redirect to the corrected note)
2026-08-10 — ASH: spend your bit budget on fewer dimensions (2 new, 4 updated)
Given a fixed number of bits per vector, is it better to keep every dimension at one bit or half the dimensions at two? ASH (Tepper & Willke, arXiv:2606.07870) measures the second winning — better recall and lower latency at identical on-disk size — provided the reduction is learned, since a random projection at the same budget loses by a margin that widens the more dimensions you drop. It is PCA plus a rotation tuned for the quantizer that follows, over data re-centered on cluster centroids because real embeddings sit off-centre rather than evenly on the sphere, which is also why the bounds carried by TurboQuant and EDEN do not transfer to them. ITQ, its 2011 ancestor, and RaBitQ turn out to be ASH with pieces switched off.
Corrections
- PCA claimed that documents and queries must share one fitted transform, full stop. That still holds, but two edges are now qualified: data-agnostic random rotations (RaBitQ, TurboQuant) are still shared artifacts that break the index if regenerated, they just never need refitting; and “asymmetric” methods are not exceptions — the query goes through the same projection, it only skips the lossy quantization.
- PCA’s fitting-sample section previously offered flat vector counts. ASH adds a third measurement that scales with dimensionality instead: ten times the embedding dimension, independent of corpus size — a more transferable rule than any fixed number.
Concepts — ASH · ITQ Updated — PCA (asymmetry and sampling qualifications) · RaBitQ (special-case-of-ASH section) · LSH (hash functions as a versioned artifact) · Dimensionality Reduction vs Quantization (the stack, fused)
2026-08-10 — A refitted PCA is an index schema change (0 new, 1 updated)
The rule that documents and queries must share one fitted transform is usually stated and left there. PCA now follows it to its operational conclusion: refit the projection and start applying the new one to incoming queries, and every cross-boundary similarity is silently wrong — nothing raises, recall just sags. A newly fitted projection is therefore a schema change, carrying the obligations of one: version the artifact with the index it built, treat a refit as a full reprojection and reindex, cut queries over atomically, never mix vintages inside one index. Matryoshka Embeddings truncation sidesteps all of it, being a fixed slice with no fitted parameters to drift.
Updated — PCA
2026-08-10 — Turbo4 drops the originals, and rescoring with them (2 new, 3 updated)
Quantization normally keeps the original vectors around: the compressed copy filters fast, the float32 originals stay on disk, and top candidates are rescored against them to recover the recall compression cost. Qdrant 1.19’s turbo4 datatype drops the originals entirely, storing only the 4-bit TurboQuant code — nine times less disk, and no rescoring stage, so the error is permanent. Qdrant 1.19 - Turbo4 Datatype and Memory Tiers covers that trade alongside three other changes, including BM25 statistics that can now be scoped per tenant inside a shared collection, which Multi-Tenancy in Search previously reached only through index-per-tenant. Written by Mohamed Arbi Nsibi; the recall figures come from random gaussian vectors, a deliberate worst case.
Articles — Qdrant 1.19 - Turbo4 Datatype and Memory Tiers People — Mohamed Arbi Nsibi Updated — TurboQuant (compression vs. storage format) · Qdrant Vector DB · Multi-Tenancy in Search
2026-08-10 — Driving Quepid against engines it does not support (1 new, 1 updated)
Quepid ships native drivers for a handful of engines; everything else goes through one escape hatch — a Custom Search API endpoint running user-supplied JavaScript over the response. Quepid Beyond Supported Engines collects what that costs. Charlie Hull hit an auth wall driving Vespa Cloud, which defaults to mTLS certificates Quepid cannot present, and switched the application to read-only token authentication. Andrew Kornilov hit a different one against Qdrant: a CLIP embedding is neither readable nor short enough for the query field, so the vector rides in a query option, leaving readable text for the raters. The official support list also disagrees with itself — Algolia and Vectara read as first-class on the marketing site, experimental in the library underneath.
New — Quepid Beyond Supported Engines Updated — Quepid
2026-08-10 — PCA fitting: memory, sampling, and the query transform (0 new, 1 updated)
PCA does not require the corpus in RAM — the fit consumes a d × d covariance matrix,
8 MB at 1,024 dimensions, whether the corpus holds ten thousand vectors or ten million.
PCA now covers the three ways to fit one — a random sample, IncrementalPCA across
two bounded-memory passes, and randomized SVD — and how much the sample decides. Size is
the easy part: Doug Turnbull fit on 100K vectors, Dylan Castillo on 1,000, and
the components stop moving well before the corpus runs out. Composition is the hard part,
and it is more than “sample randomly” — a uniform draw is dominated by the largest
categories, so the projection it learns is worse for everything rare. Sample along the
axes that actually vary — category, language, market — because stratifying 100K matters
more than going from 100K to 1M. It also states the fitted artifact — projection matrix
plus mean vector — and why documents and queries must share one transform, which
refitting invalidates.
Updated — PCA
2026-08-10 — Matryoshka truncation vs PCA on BEIR (2 new, 6 updated)
Shrinking an embedding has two routes — truncate an MRL-trained model, or fit a PCA projection afterwards — and the assumption has been that truncation wins where the model supports it. Honey, I Shrunk the Embeddings - Matryoshka vs PCA measures both across eight BEIR subsets and reports the reverse: PCA holds 82% of nDCG@10 at 64 of 1,536 dimensions against truncation’s 71%, and works on text-embedding-ada-002, which was never trained to be truncated. Fitting is cheap too: 1,000 documents matched the full corpus, and an MS MARCO fit transferred to unrelated datasets. One practitioner’s benchmark by Dylan Castillo, with exact search and no ANN indexing; Dimensionality Reduction vs Quantization records it as counter-evidence without changing its recommendation.
Corrections
- Matryoshka Embeddings gave “~98% of performance at 64 dims” as a flat property of MRL; it now marks that figure as measurement-dependent, against 71% nDCG@10 retention measured independently at the same dimension.
Articles — Honey, I Shrunk the Embeddings - Matryoshka vs PCA People — Dylan Castillo Updated — PCA · Matryoshka Embeddings · Dimensionality Reduction vs Quantization (dissenting-benchmark section) · Principal Component Analysis - an embedding shrink-ray · BEIR · Embedding Models Compared