History — 2026 week 27 (Jun 29 – Jul 5, 2026)
Newest first.
2026-07-05 — Conferences MOC + wikilink sweep
Added a Conferences MOC indexing the four conference notes (Haystack US, Haystack EU, Berlin Buzzwords, MICES) and linked it from global_toc’s Maps of Content. Swept previously plain-text conference mentions into wikilinks across curated hub notes (Search Communities, Search Consultancy, How to Start a Career in Search, Women of Search, Search Result Diversity, Events and Conferences, The Search Juggler, Audrey Lorberfeld, Rene Kriegler) and two factual article references (Block-Max WAND / Metarank citing Berlin Buzzwords). Deliberately left untouched: the unrelated deepset Haystack framework references (HyDE notes, django-haystack), the Haystack Live meetup mentions, markdown-hyperlinked citations, Clippings/, and the link-dump README.
2026-07-05 — Haystack US + Haystack EU conference notes (2 notes)
Added Haystack US and Haystack EU to the Conferences/ folder — the two editions of OpenSource Connections’ search relevance conference (US original in Charlottesville, VA; European edition in Berlin). Cross-linked to each other, to Doug Turnbull and Charlie Hull (organizer), the Haystack Live meetup talk, and Berlin Buzzwords / MICES as the other Berlin search events. Wikilinked from Events and Conferences and Charlie Hull; registered in global_toc Conferences.
2026-07-05 — MICES + Berlin Buzzwords conference notes + new Conferences content type (2 notes)
Created MICES (Mix-Camp E-Commerce Search) in a new Conferences/ folder — the Berlin e-commerce search conference organized by Rene Kriegler, co-located with Berlin Buzzwords and overlapping E-commerce Search. Consolidates references previously scattered as plain text across Events and Conferences, Andreas Wagner (MICES talk on result diversity / basket composition), and the Kriegler notes. Backlinked from Rene Kriegler, Rene Kriegler - Query Relaxation, Andreas Wagner, and Events and Conferences (all now wikilinking MICES); registered in global_toc under a new Conferences section. Also added Berlin Buzzwords — the open-source Berlin search/data conference (co-located with MICES) that recurs across the vault as plain text — linked to Lester Solbakken’s Hybrid search > sum of its parts? Berlin Buzzwords 2022 talk and wikilinked from Events and Conferences and the Kriegler notes.
2026-07-05 — René Kriegler’s “Query Relaxation” talk (2 notes)
Added the video Rene Kriegler - Query Relaxation — a 2019 OpenSource Connections talk (introduced by Charlie Hull) reframing Query Relaxation not as query repair but as a query recommendation problem: the aim is to keep the user in the conversation, not reconstruct exact intent. Walks the zero-result strategy spectrum (Query Expansion → spelling → low-quality fields → Boolean loosening → hypernyms → related queries → recommendations) and benchmarks term-drop predictors on freq vs. session co-occurrence datasets: random baseline → shortest/digit word-shape heuristics → least-frequent index term → (entropy fails) → Word2Vec cosine similarity → Grbovic-style query embeddings → a multi-layer neural network on word2vec + word-shape features (best, ~0.9 P/R), running real-time via TensorFlow. Created the new person Rene Kriegler (e-commerce relevance consultant, maintainer of Querqy, organizer of MICES). Backlinked from Query Relaxation, Querqy (added Kriegler as maintainer), Charlie Hull, Daniel Tunkelang (blog-post inspiration), Mihajlo Grbovic, OpenSource Connections, E-commerce Search, and Query Understanding in Practice; registered in global_toc (Videos + People), the Videos MOC, Index, and All about Information Retrieval & Search.
2026-07-05 — FAISS index tutorial video + ANN/FAISS/LSH hubs (4 notes)
Added the video Choosing Indexes for Similarity Search (Faiss in Python) by James Briggs (Pinecone) — a hands-on FAISS walkthrough of the four index families (Flat, LSH, HNSW, IVF) benchmarked on Sift1M, the video companion to the existing Nearest Neighbor Indexes for Similarity Search article. Filled three long-standing hub gaps discovered while linking: FAISS (tool — previously referenced by 30+ notes as plain text with no page), LSH (concept), and Approximate Nearest Neighbor Search (the parent ANN concept that HNSW/IVF linked to but never existed). Backlinked from James Briggs, HNSW, IVF, and the Pinecone article; registered in global_toc (ANN Indexing, Search & Vector Engines, Videos), Concepts, Tools, and the Videos MOC.
2026-07-05 — Videos section + Max Irwin’s “Search Engine Migration Circus” (1 note, new content type)
Introduced a new Videos content type for conference talks and recorded presentations, with its own folder (Videos/) and a section across all indexes. First entry: Max Irwin - The Search Engine Migration Circus, a Haystack Live talk by Max Irwin (OpenSource Connections) drawn from his Wolters Kluwer / MediRegs migrations. Covers the OSC migration playbook (define success → measure baseline → “Hello Search” PoC → feature-parity analysis → risk register → gradual customer migration), the observation that content is the timeline killer and 100% feature parity is impossible, the “damage” metric (top-N result-set diff between legacy and target engines), and practical war stories (FAST ESP deprecation, the AWS lift-and-shift mistake, staggered query-parser/highlighter releases, forced-cutover vs. gradual migration). Q&A touches re-rank-only latency fixes and the Quepid / Rated Ranking Evaluator tooling. Companion to the Migration between Search Engines topic. Backlinked from Max Irwin (new Talks & Videos section) and Migration between Search Engines (Related Notes); registered in global_toc (new Videos section + MOC pointer), Index, and All about Information Retrieval & Search.
2026-07-03 — NDCG Variants topic (the “Flavors of NDCG”)
Created NDCG Variants, a Topic hub synthesizing Doug Turnbull’s “flavors of NDCG” across two independent axes: the gain function (Järvelin grade-gain vs. Burges exponential 2^rel−1) and the normalization target (“normalized to what?” — local / recall / global / max), plus @k cutoffs and per-library defaults (scikit-learn, RankLib, LightGBM, MS MARCO / BEIR, Quepid). Consolidates the two existing article clippings (Flavors of NDCG, Flavors of NDCG - normalized to what) that each covered only one axis. Backlinked from the NDCG concept, both source articles (new Related Topics sections), global_toc, Index, Topics, and All about Information Retrieval & Search.
2026-07-03 — Hooking Quepid to Vespa (1 note)
Added How to Securely Hook Up Quepid to Vespa by Charlie Hull (The Search Juggler). A proof-of-concept for the vault’s recurring theme — driving Quepid against a non-Lucene backend as a custom search API. Here the backend is Vespa Cloud, authenticated with a read-only token (Quepid can’t use Vespa’s self-signed client certs) and parsed via JavaScript result mappers. Fills Vespa’s gap in interactive offline relevance testing and points toward exporting Quepid ratings as re-ranking training data. Companion to a Maven Lightning Lesson co-presented with Trey Grainger. Backlinked from Charlie Hull, Quepid, Vespa, and the Search QA MOC (Quepid in Practice).
2026-07-03 — OpenSearch Search Relevance Workbench & Elastic Relevance Studio (4 notes)
Filled the vault’s biggest relevance-tooling gap: the vault had deep Quepid coverage but nothing on the engine-native evaluation tools that have emerged to rival it. Added both, plus a comparison note anchoring the three-way choice. Cross-engine takeaway: Elastic and OpenSearch are absorbing Quepid’s offline-evaluation loop into the engine itself, adding behavior-driven (UBI/click) and agent-driven (LLM/MCP) judgments — and notably OpenSource Connections (Quepid’s authors) also drove the OpenSearch tool, so it’s the same community pushing the practice into the engines.
Tools — Search Relevance Workbench (new — OpenSearch 3.1+ native tool: query sets / search configurations / judgments / experiments; three judgment sources incl. UBI COEC implicit judgments and LLM-as-judge; unique hybrid-search grid-search optimization; imports Quepid CSV; Dashboards visualizations in 3.2); Elasticsearch Relevance Studio (new — Elastic’s experimental React+Flask lifecycle app over the ES Search API: scenarios / judgements / strategies / benchmarks; drag-slider judging; NDCG/Precision/Recall/MRR + unrated-doc reporting; headline MCP server for agentic AI workflows); Rated Ranking Evaluator (new — Sease’s open-source, CI/CD-oriented offline evaluation library for Solr/Elasticsearch; JVM/Maven modules, corpus→topics→query-groups→queries domain model, immutable version-over-version delta tracking, broad metric set incl. ERR/F-measure, RRE Server dashboard + RRE Enterprise; resolves a long-standing dead link).
Topics — Relevance Evaluation Tools Compared (new — four-way Quepid vs. Workbench vs. Relevance Studio vs. RRE; at-a-glance / judgment-sources / experiment-capabilities / metrics tables + a “how to choose” guide).
People — Daniel Wrigley linked (authored the OpenSearch SRW/judgments blog as well as the Quepid guide).
Updated — Quepid (Related Tools → SRW/ESRS/RRE + comparison pointer), OpenSearch (Related Tools → SRW), Elasticsearch (Related Tools → ESRS), Search Evaluation (new Tools section), Sease (added RRE as its product), Relevance Program Setup (RRE wikilinked), Tools, global_toc, index.md / Index / All about Information Retrieval & Search (Tools highlight lists).
Correction — OpenSource Connections previously mis-listed Rated Ranking Evaluator as an OSC product; RRE is Sease’s. OSC’s products corrected to Quepid + the Elasticsearch LTR plugin, with a note pointing RRE to Sease.
2026-07-01 — Solr↔Vespa Onboarding & AI-Assisted Migration (7 notes)
Processed two Clippings articles that both orbit cross-engine migration — and created a Migration between Search Engines topic to anchor them alongside the existing ES→Vespa case studies.
Articles — How I learned Vespa by thinking in Solr (new — Sujit Pal, Elsevier Labs, 2021-02-24; a Solr→Vespa mapping table — core≈application, managed-schema≈.sd, MLT≈nearestNeighbor ANN, YQL≈SQL — MVP on CORD-19 + SPECTER embeddings via HNSW); Amazon OpenSearch Service now offers AI-assisted migrations (new — Amazon Web Services, 2026-06-23; Migration Assistant’s agent-guided (Kiro / Claude Code) workflow for Solr/ES/OpenSearch → Amazon OpenSearch, now with live-traffic capture/replay for Solr).
Topics — Migration between Search Engines (new hub — schema/analyzer parity, historical + live-traffic sync, relevance validation; groups the hand-built Vinted/Kleinanzeigen ES→Vespa migrations, AWS’s tool-assisted OpenSearch migration, and Solr→Vespa mental mapping).
Tools — Solr (new — Apache Lucene-based engine; cores, managed-schema/solrconfig.xml, MLT, function queries; common migration source).
People — Sujit Pal (new — Technology Research Director, Elsevier Labs).
Companies — Elsevier (new — Elsevier Labs), Amazon Web Services (new — Amazon OpenSearch Service / Migration Assistant).
Updated — Vespa + OpenSearch (article lists), Vinted - Migrating Search from Elasticsearch to Vespa + Kleinanzeigen - Vespa Migration for Homepage Feed (linked to the migration topic), People, global_toc (Tools/Companies/Topics/People), index.md (Tools/Companies/Stats: 7 tools, 30 companies, 28 topics).
2026-06-29 — Vinted Dense Retrieval & Billion-Scale (2 notes)
Finished the Vinted Vespa arc from the “Search Scaling” series — importing only the Vespa-era, retrieval-relevant posts (all pre-Vespa/ops chapters deliberately skipped).
Articles — Dense Retrieval at Vinted (new — Laurynas Jasiukėnas & Dainius Jocas, 2025-11-18; frozen multilingual-CLIP two-tower query/item towers, 256-dim, hybrid ANN-supplements-lexical capped to top-K, contrastive training 7–10k negatives/positive over >100M pairs, HNSW on 30 content nodes/group × 3 market indices, 500ms budget w/ 350ms approx + 150ms exact fallback, GraalVM/ZGC, <0.02% error, ~50 A/B tests).
People — Laurynas Jasiukėnas (new — Vinted; dense-retrieval co-author).
Folded in (no new note) — Search Scaling Chapter 9 “Billion-Scale Search” (Dainius Jocas, 2025-01-10; 1B docs by Nov 2024, ~10× since 2019, mean <20ms at data layer, ~2× headroom) added as a Postscript section to Vinted - Migrating Search from Elasticsearch to Vespa.
Updated — Dense Vector Retrieval (Vinted as a production example), Vinted + Dainius Jocas (article lists), global_toc (People L).
2026-06-29 — Vinted Vespa Match-Features (3 notes)
Captured Vinted’s Vespa match-features post — a sharp engineering result tying into Vespa Learning to Rank: using match-features as an in-engine feature store (replacing Redis round-trips) and, more surprisingly, to cut latency by skipping the document-summary .fill() fetch (two-phase scatter-gather → single round-trip; P99 ~9ms→3ms, mean ~430µs at 7.5k RPS). The companion 2023 recommendation-retrieval post was already in the vault — only its authorship was filled in.
Articles — Optimizing Vespa Latency with Match-Features at Vinted (new — Dainius Jocas, 2025-11-06; match-features declaration/tensors, Vespa-as-feature-store vs Redis, summary-fetch elimination, latency results).
People — Dainius Jocas (new — Vinted; author of both Vinted Vespa posts), Aleksas Kateiva (new — Vinted; co-author of the recommendation post).
Updated — Vespa Learning to Rank (match-features bullet → online-serving/latency note + link), Adopting Vespa for Recommendation Retrieval at Vinted (real authors + People section), Vinted (article list + people), global_toc (People A/D).
2026-06-29 — Vinted ES→Vespa Search Migration (2 notes)
Processed Vinted Engineering’s “Search Scaling Chapter 8: Goodbye Elasticsearch, Hello Vespa” into a case study — a rare concrete account of a billion-item, 20k-RPS production search platform moving off Elasticsearch to Vespa. Decisive wins were operational (shard toil eliminated, no hot nodes, change-visibility 300s→5s, server fleet halved to 60), enabled by porting Lucene text analyzers into Vespa.
Case Studies — Vinted - Migrating Search from Elasticsearch to Vespa (new — scale, motivations, 60 content / 12 container / 3 config node topology, Flink + open-sourced Vespa Kafka Connect sink, Go “search contract” middleware with 12 query patterns, 3× ranking depth to 200k candidates, May 2023→Apr 2024 timeline, results).
People — Ernestas Poškus (new — Vinted search-platform engineer; author of the Search Scaling series).
Updated — Vinted (new Search Platform Migration section + frontmatter), Case Studies, MOC - Case Studies (Architecture & Platform Migration), global_toc (Case Studies + People), Index / All about Information Retrieval & Search (Case Studies list + counts 9→10).
2026-06-29 — Vespa Learning to Rank (2 notes)
Gave Vespa the same LTR coverage as Elasticsearch Learning to Rank. Framing: Vespa doesn’t add a dedicated LTR subsystem — GBDT LTR, neural reranking, and MaxSim are all just expressions in its one tensor/ranking-expression engine, runnable across first/second/global phases and ensemble-able.
Topics — Vespa Learning to Rank (new — phased ranking first-phase/second-phase(content-node, rerank-count)/global-phase(container, cross-hit); native GBDT import via xgboost("model.ubj") / lightgbm("model.json") and onnx(...) neural models in models/; rank features bm25/nativeRank/fieldMatch/attribute/closeness + tensor MaxSim; feature dumping via match-features/summary-features/rank-features; train-offline/serve-in-engine workflow; Vespa-vs-Elasticsearch LTR table).
Tools — ONNX (new — Open Neural Network Exchange + ONNX Runtime; how neural rankers/cross-encoders are served in-engine, esp. Vespa global-phase).
Updated — Learning to Rank and Elasticsearch Learning to Rank (sibling cross-links; Vespa examples linked), Vespa (company note: LTR capability + Concepts links), XGBoost / LightGBM (Related links to Vespa LTR), global_toc (new ML & Model Serving tools group), Concepts, Index / All about Information Retrieval & Search indices.
Reclassification — both engine-specific LTR notes (Elasticsearch Learning to Rank, Vespa Learning to Rank) moved concept → topic (folder Concepts/ → Topics/) to match the Late Interaction in <engine> topic pattern; indices and counts adjusted (−2 concepts, +2 topics).
2026-06-29 — Late Interaction in Vespa (1 note)
Completed the four-engine late-interaction set, bringing Vespa to the same coverage as Elasticsearch / OpenSearch / Qdrant. Framing: Vespa is the origin engine for production late interaction (Jo Kristian Bergum’s native ColBERT embedder + the sign-bit binarization the field later adopted), and the only one that expresses MaxSim as a general tensor ranking expression rather than a dedicated field/function.
Topics — Late Interaction in Vespa (new — mixed tensor<int8> storage for ColBERT tokens / long-context windows / ColPali patches; MaxSim written as reduce(...,max,...)+sum tensor expressions; hamming + unpack_bits 32× binary quantization; native multi-phase ranking with nearestNeighbor/BM25 first phase + query-token pruning; billion-scale ColPali by computing on content nodes — “speed of memory not network”; four-engine comparison table).
Updated — Vespa (company note: ColBERT embedder → topic pointer + Concepts link), Late Interaction and the three sibling topics (Late Interaction in Elasticsearch, Late Interaction in OpenSearch, Late Interaction in Qdrant) cross-linked, global_toc, Topics, Index / All about Information Retrieval & Search indices.
2026-06-29 — Late Interaction in OpenSearch & Qdrant (3 notes)
Extended the late-interaction cluster sideways from Late Interaction in Elasticsearch to the other two engines, so the vault now covers how ColBERT/ColPali multi-vector reranking is implemented across all three. Cross-engine takeaway: all three converge on late interaction as a reranker over a cheap first stage, differing mainly in packaging.
Topics — Late Interaction in OpenSearch (new — native lateInteractionScore painless fn in OpenSearch 3.3+, object+float multi-vector storage, Lucene LateInteractionField/LateInteractionRescorer 10.3+, ml-inference ingest/search processors, two-phase retrieval, 10–100× storage); Late Interaction in Qdrant (new — first-class multivector_config with MultiVectorComparator.MAX_SIM, hnsw_config m=0 for rerank-only vectors, single-call prefetch+query retrieve-then-rerank via the Query API, MUVERA/dense/sparse first stage).
Concepts — MUVERA (new — Multi-Vector Retrieval via Fixed Dimensional Encodings; collapses a ColBERT multivector to one ANN-indexable vector for stage-1 retrieval; the Qdrant analogue of Elastic’s “average vectors”).
Updated — Late Interaction in Elasticsearch and Late Interaction (sibling cross-links), OpenSearch and Qdrant Vector DB (late-interaction capability + links), Elasticsearch vs OpenSearch (late-interaction feature row + links), global_toc, Topics, Concepts indices.