History — 2026 week 25 (Jun 15 – Jun 21, 2026)
Newest first.
2026-06-19 — Frontier-of-Search Period Pages & Late Interaction in Elasticsearch (~4 notes)
Made the “frontier” concept year-anchored and processed Elastic’s two-part ColPali series into a dedicated topic.
Topics — Frontier of Search (new year-by-year index/TOC linking each period page), Frontier of Search 2025 (new period page — late interaction goes multimodal & production-scale, embeddings as a commodity, vector quantization/ANN engineering, the keyword-efficiency renaissance, and the first RL/agentic seeds), Late Interaction in Elasticsearch (new — rank_vectors + maxSimDotProduct in ES 8.18, bit/average vectors, Token Pooling, two-stage rescore).
Articles — ColPali & Elasticsearch - How to Search Complex Documents (by Peter Straßer at Elastic, 2025-03-16). Part 1 of the ColPali series — what ColPali is, the ViDoRe benchmark, the rank_vectors field, and maxSimDotProduct scoring. (Part 2, Late Interaction Models - How to Scale and Optimize in Elasticsearch, already existed.)
Renamed — “Current Frontier of Search” → Frontier of Search 2026 (future-proofing; all 25 wikilinks + frontmatter references updated). Dropped the now-ambiguous “Frontier of Search” alias since it is now a real index note.
Updated — Late Interaction, ColPali, Late Interaction Models - How to Scale and Optimize in Elasticsearch, Peter Straßer, Frontier of Search 2026 cross-linked into the new cluster; index.md Topics + Stats refreshed.
2026-06-18 (evening) — Agentic Search Frontier & Reasoning Reranking (~14 notes)
Built out the “agents as the new user of search” cluster, spanning purpose-built retrieval models, the agentic query workload, and the shift of reranking toward LLMs.
Articles — Agentic Search Models with OpenSearch and Elasticsearch (by Bonsai, 2026-06-05). Hands-on integration of the SID-1 model into a managed OpenSearch cluster as a drop-in relevance layer; tool-based rewrite/retrieve/rerank loop. The Scaling Dimensions of Keyword Search (by Skip Everling at Hornet, 2026-02-06). Benchmark arguing document count is only one scaling dimension — query length and structure matter too, and Block-Max WAND pruning can fall below exhaustive scoring on the long queries agents emit.
Concepts — Purpose-Built Agentic Search Models (LLMs trained specifically on the rewrite/retrieve/rerank loop — smaller/faster/cheaper than frontier models on domain retrieval), Agentic Query Workload (“agents are the new user of search” — long, structured, high-volume, out-of-distribution queries that strain serving infrastructure), RankGPT (listwise LLM reranker, sliding-window permutation, no fine-tuning), RankLLaMA (stub — LLaMA-based pointwise reranker), MonoT5 (stub — T5 pointwise seq2seq reranker).
Companies — Bonsai (managed OpenSearch/Elasticsearch hosting), SID.ai (purpose-built agentic search/reranking models).
Tools — SID-1 (SID.ai’s first agentic retrieval model; OpenAI-compatible API; ~1.9× over embedding-only search, ~24× faster than frontier agentic retrieval).
People — Skip Everling (Hornet; retrieval infrastructure for agentic workloads).
Topics — Frontier of Search 2026 (agents as first-class search users — converging product and research fronts), Reasoning Reranking (LLMs as rerankers: prompted, fine-tuned, or chain-of-thought), RL-Trained Search Agents (training the searcher policy with RL — complement to purpose-built models).
Updated — This Is What Agentic Retrieval Looks Like, When Reranking Becomes a System Boundary, Agentic Search, Reranking, Block-Max WAND, Direct Corpus Interaction, Reinforcement Learning for Search, Search-R1, Doug Turnbull, Vespa, Conversational and Agentic Search, MOC - Agentic Search and Embeddings cross-linked into the frontier cluster.
2026-06-18 (afternoon) — Career & Learning Paths (~2 notes)
Added entry-point topics for people getting into search and curated learning resources.
Topics — How to Start a Career in Search (learning path + the people, books, courses, and communities to follow), Courses (IR/search training and course resources).
Updated — Books, Events and Conferences, Hiring for Search, Search Consultancy, Trey Grainger, OpenSource Connections cross-linked into the career/learning cluster.
2026-06-18 — Search Problem Archetypes (~4 notes)
Processed Atita Arora’s archetype article into a new Topic hub. Decision: this article earned a dedicated topic note — it is a self-slotting diagnostic taxonomy spanning ten domains, not a single concept.
Topics — Search Problem Archetypes — the 10 recurring search-problem patterns (Uniqueness, Complexity Machine, Precision Mandate, Firehose, Extraction, Media Vault, Knowledge Graph, Geospatial, Q-commerce, Code Search); diagnosis-before-solution framing (Ackoff); “same symptom, opposite prescription” (Zero Results); agentic retrieval as a consumption pattern, not an archetype.
Articles — Before You Fix Your Search, Know What’s Actually Broken (by Atita Arora, Unscripted AI Substack, 2026-06-16). Argues most search failures are misdiagnoses — teams solve the wrong problem (and skip defining relevance / judgment lists) before reaching for LTR or vectors.
People — Atita Arora (search relevance consultant; Unscripted AI), Udi Manber (search leader at Amazon/Yahoo/Google/YouTube; “search is essentially a solved problem” misperception).
Updated — Zero Results and Agentic Search cross-linked to Search Problem Archetypes; index.md Topics/People/Stats refreshed.
2026-06-16 — Ranking Objectives & LTR Cluster (~10 notes)
Expanded the Ranking Objectives stub into a full concept note from Clippings sources: the three families (pointwise / pairwise / listwise), the objective-encodes-business-goals framing, the LambdaMART lambda-swap bridge from pairwise training to a listwise metric, and a per-library objective table (LightGBM / XGBoost / CatBoost).
Articles — Pointwise vs Pairwise vs Listwise Learning to Rank (by Nikhil Dandekar, Medium, 2016-09-29). Canonical taxonomy of LTR approaches by how many documents the loss considers; names RankNet/LambdaRank/LambdaMART (pairwise) and SoftRank/AdaRank/ListNet/ListMLE (listwise).
LambdaMART Explained - The Workhorse of Learning-to-Rank (by Tullie Murrell at Shaped, 2025-07-30). RankNet → LambdaRank → LambdaMART lineage, the boosting loop, advantages/limits, library objective strings. Learning to Rank - A Complete Guide to Ranking using Machine Learning (by Francesco Casalegno, TDS, 2022-02-28). Full objective landscape — MAP/NDCG metrics, pointwise/pairwise/listwise, SoftRank, ListNet, and the LambdaLoss generalization. Building a Large-Scale Recommendation System - People You May Know (LinkedIn Engineering, 2024-02-06). Production multi-stage funnel (L0 recall → L1 light ranker → L2 rich ranker → fairness/diversity re-ranker).
Concepts — LambdaLoss (stub) — generalized listwise framework unifying RankNet/LambdaRank/SoftRank/ListNet.
People — Nikhil Dandekar (search / LTR / NLP practitioner), Tullie Murrell (Shaped), Francesco Casalegno (LTR writer).
Updated — Ranking Objectives promoted from stub to full note (three families, business-goal framing, lambda-swap bridge, per-library objective table, LambdaLoss); LambdaMART cross-linked to the two new LTR articles.