History — 2026 week 26 (Jun 22 – Jun 28, 2026)
Newest first.
2026-06-28 — Quepid Practical Use Cases: Vector & Image Search Evaluation (8 notes)
Added a cluster of practical, “what actually breaks” use cases for Quepid beyond the canonical lexical workflow — collaborative team judging, and the hacks needed to evaluate vector and image search. Anchored by a new Vector Search Evaluation concept that catalogs why judgment-list tooling built for text queries struggles with embeddings (query-length limits, JSON-validity catch-22, non-human-readable queries, non-text results) and the workarounds.
Concepts — Vector Search Evaluation (new; ties together cross-modal, image-to-image, hybrid evaluation and the Quepid limitations/workarounds).
Articles — Creating Judgement Lists with Quepid (Daniel Wrigley, Elastic Search Labs; collaborative judging, information needs, books of judgements, v8 AI judgements); Why Setting Up Quepid for Vector Search Evaluation Went Wrong, Oops, I Did It Again, How to Evaluate Image Search in Qdrant Using Quepid Part 1, How to Evaluate Image Search in Qdrant Using Quepid Part 2 (all Andrew Kornilov; the vector/image series — dimension reduction to fit limits, query-option vector injection, Qdrant-as-endpoint, scorer hacks + upstream PR #1683).
People — Andrew Kornilov (vault author; hands-on Quepid/vector series; unofficial Quepid API), Daniel Wrigley (Quepid judgement-lists guide).
Updated — Quepid (new Practical Use Cases section + expanded Related Articles), global_toc, Index / All about Information Retrieval & Search (Evaluation concepts + Key People + stats), MOC - Search Quality Assurance and Query Understanding (Judgment & Annotation + Key People).
2026-06-27 — Search UX & Discovery, Click Models & UX Research Orgs (13 notes)
A browseability + UX-research build-out, applying README-gap analysis and surfacing the Search UX & Discovery and Lexical Query Operations families across the indices.
Concepts — Neural Click Models (NCM, CACM, GraphCM, two-tower; deep-learning successors to Click Models); Search Scopes (pre-query narrowing vs. post-query Faceted Search; the sticky-scope failure mode).
Topics — Search UX Research (new hub for independent search-UX research organizations).
Companies — Baymard Institute, Nielsen Norman Group, Enterprise Knowledge LLC (search-UX research orgs).
People — Heather Hedden (The Accidental Taxonomist); Elzbieta Jakubowska (filled a Women of Search placeholder / dangling link).
Articles — four Nielsen Norman Group studies: The Pinball Pattern - Complex Search-Results Pages Change Search Behavior, Scoped Search - Dangerous but Sometimes Useful, Good Abandonment on Search Results Pages, Search-Log Analysis - The Most Overlooked Opportunity in UX Research.
MOCs — MOC - Case Studies (new curated map; resolved a long-standing dangling link); MOC - Search UX and Discovery rewritten to surface the full family.
Index restructuring — split visible Search UX & Discovery and Lexical Query Operations families in global_toc and the home indices; moved Knowledge Graph Search and Results Boosting to coherent homes; relocated the Search UX topic out of Strategy & Meta. Fixed an inverted alias link in Neural Click Models.
2026-06-27 — Duality in Measuring Search Topic (1 note)
Created a Topic hub for James Rubinstein’s recurring “qual + quant” thesis — that knowing whether search works needs two complementary signals and “it’s not one-or-the-other, it’s yes, and!” The hub organizes the duality under two lenses: improving search (statistical / metrics-driven vs. human-centered) and measuring search (online / log-based vs. offline / human-rated), reconciled at the launch review.
Topics — Duality in Measuring Search (new hub — two side-by-side comparison tables for the improvement and measurement lenses; why neither half suffices; how it fits the broader Search Evaluation / Relevance Program Setup / A-B Testing for Search stack; alias “Two Approaches to Measuring Search”).
Updated — Measuring Search - A Human Approach and Statistical and Human-Centered Approaches to Search Improvement (both normalized and marked as paywalled — access: paywalled, paywalled tag, published dates; the second’s body left faithful to source; fabricated content in the former replaced with a faithful summary and its accidental duplicate removed). James Rubinstein (article list extended with the two posts + The Launch Review).
2026-06-25 — Interaction Paradigms Topic (1 note)
Extracted the broader idea behind the “Interaction Timeline Comparison” section of Late Interaction into a dedicated Topic hub spanning the no / late / early interaction spectrum, which organizes Bi-Encoder, ColBERT / Late Interaction, and Cross-Encoder as three points on one axis (when query and document are allowed to interact).
Topics — Interaction Paradigms (new hub — the no/late/early interaction spectrum; both timeline diagrams; comparison table across bi-encoder / late-interaction / cross-encoder covering pre-encoding, granularity, speed, quality, pipeline role; an orthogonal sparse-vs-dense axis tying in Learned Sparse Retrieval / SPLADE / ELSER / Hybrid Search).
Updated — Late Interaction (pointer added under the comparison section + Related Topics entry), Bi-Encoder (backlink — the no-interaction endpoint), Cross-Encoder (backlink — the early-interaction endpoint).
2026-06-25 — Search Communities & Women of Search (3 notes)
Added a community / careers cluster, distinguishing standing people-networks from the events where they gather. Decision: Women of Search is a community, not a conference — it does not belong under Events and Conferences/Haystack, so a dedicated Search Communities hub was created as its parent.
Topics — Search Communities (new hub — identity/mentorship communities, meetups, and Slack/online communities; why communities are the field’s learning + hiring pipeline; explicitly distinct from Events and Conferences), Women of Search (new — international community for women in search, founded May 2021; mission, core leadership, mentorship program, Haystack talks; sourced from women-of-search.org).
People — Audrey Lorberfeld (new — librarian-turned-software engineer; founder of Women of Search; AI-driven IR). Seven WoS leadership names (Erika Cardenas, Atita Arora, Elzbieta Jakubowska, Meghan Boyd, Olena Gorbatiuk, Samdisha Kapoor) left as ready-to-fill wikilinks.
Updated — Events and Conferences frontmatter repair; Weaviate referenced as employer of WoS core member Erika Cardenas.
2026-06-23 — Federated Search Concept (1 note)
Added a dedicated concept note for federated search in its information-retrieval sense — distributed IR / metasearch — deliberately kept distinct from the “federated graph” data-unification sense (Netflix), which is entity modeling rather than a retrieval problem.
Concepts — Federated Search — querying multiple independent collections/engines and merging results; the three sub-problems (resource/collection selection, query routing, results merging), cooperative vs. uncooperative environments, and its relationship to Hybrid Search (federation over representation spaces vs. over collections).
Updated — Multilingual Search (per-language-index fan-out linked), Enterprise Search (multi-source retrieval linked), Search Architecture (Netflix “federated graph” disambiguated), Reciprocal Rank Fusion (added as the standard merger for fan-out), Knowledge Graph Search (callout disambiguating federated-graph vs. federated-search) cross-linked into the new note.
2026-06-22 — Region-Based Embeddings & Generative Retrieval / Semantic IDs (~27 notes)
Two related embedding-geometry clusters. First, the region-based representation lineage — words/items as regions (boxes, Gaussians, hyperbolic balls) rather than points — and its search payoff for set-theoretic/compositional queries. Second, the generative retrieval cluster — discrete Semantic IDs as the identifier scheme behind sequence-to-sequence retrieval, anchored to its IR-native origin Differentiable Search Index rather than framed as pure recsys.
Articles — Express Words in a Box - Understanding Box Embedding from the Basics (by Shun Tsukagoshi, Behitek / State of AI Guide, 2022-12-16; paywalled — processed from supplied text). From-the-basics tutorial on box embeddings as a Region-Based Representation, walking the lineage from point embeddings to Word2Box. Answering Compositional Queries with Set-Theoretic Embeddings (by Shib Sankar Dasgupta, Andrew McCallum, Steffen Rendle, Li Zhang, arXiv:2306.04133, 2023-06-07). The search/IR payoff of box embeddings — representing item–attribute relations as “learnable Venn diagrams” answers AND/OR/NOT Compositional Queries better than dot-product vectors. Semantic IDs for Recommendation Systems (by Janu Verma, Incomplete Distillation, 2025-08-04). Hands-on build of Semantic IDs from vector quantization through RQ-VAE, reproducing a TIGER-style generative pipeline on Amazon Beauty.
Concepts — region-based family: Box Embedding, Region-Based Representation, Gaussian Embedding, Poincaré Embedding, Word2Box, Set-Theoretic Embeddings, Compositional Embeddings, Compositional Queries. Generative-retrieval family: Generative Retrieval, Differentiable Search Index, Semantic IDs, RQ-VAE, TIGER. Foundations & interpretability: Word2Vec, Concept Vectors, Steering Vectors.
People — Shib Sankar Dasgupta (Word2Box / set-theoretic embeddings lead author), Andrew McCallum (UMass; box-embedding line), Steffen Rendle and Li Zhang (compositional-queries co-authors), Luke Vilnis (box/order embeddings), Shun Tsukagoshi (Nagoya; box-embedding tutorial), Tomas Mikolov (Word2Vec), Janu Verma (Incomplete Distillation; semantic IDs).
Updated — Embeddings, Dense Embeddings, Faceted Search, Vector Quantization cross-linked into the two new clusters.