Awesome Search KG

2026.31

10 min read

History — 2026 week 31 (Jul 27 – Aug 2, 2026)

Newest first.

2026-07-30 — The Vespa zero-shot ranking series (11 new, 16 updated)

A dense retriever that beats BM25 in its training domain loses to it nearly everywhere else — that is Jo Kristian Bergum’s problem statement across three Vespa posts, and the fix is unglamorous. Part two fuses a tuned BM25 baseline with a 22M-parameter distilled ColBERT under min-max normalization, reaching 0.481 average nDCG@10 on BEIR against 0.453 for BM25 alone, winning 12 of 13 datasets with no fine-tuning and under 60 ms on CPU. The sting is in the baseline: their BM25 beats the BM25 numbers published with BEIR, so some published neural gains are baseline artifacts. Part three builds training data from three labeled queries via Consistency Filtering.

Articles — Improving Zero-Shot Ranking with Vespa Hybrid Search · Improving Zero-Shot Ranking with Vespa Hybrid Search - part two · Improving Search Ranking with Few-Shot Prompting of LLMs · Vespa - Ranking Without Labels on CORD-19 Concepts — Synthetic Query Generation · Consistency Filtering · Dense Passage Retriever · PROMPTAGATOR · FLAN-T5 Datasets — TREC-COVID · Natural Questions Updated — Zero-Shot Retrieval · Score Normalization (distributed min-max via match-features) · Hybrid Search · BM25 · ColBERT · Cross-Encoder · Knowledge Distillation · Hard Negative Mining · Linear Score Combination · Late Interaction in Vespa · Presentation Bias · ONNX · BEIR · MS MARCO · Jo Kristian Bergum · Vespa

2026-07-30 — Bayesian BM25 stub corrected (0 new, 3 updated)

Bayesian BM25 is logistic calibration of a BM25 score into a probability, which is what makes scores from different retrievers comparable enough to fuse. Both Doug Turnbull sources — Bayesian BM25 is Cool and Can BM25 be a Probability — agree on that, and the note now carries the 2026 formulation it was missing along with the alternatives Turnbull weighs against it.

Corrections

  • BM25’s stub description of Bayesian BM25 contradicted the dedicated note; rewritten to match it.
  • The claim that Bayesian BM25 is “useful for out-of-vocabulary and low-frequency terms” appears in neither source — removed from BM25 and from Out-of-Vocabulary, which had inherited it.

Updated — BM25 · Bayesian BM25 · Out-of-Vocabulary

2026-07-30 — Tokenization and Out-of-Vocabulary (2 new, 14 updated)

Two kinds of tokenization get confused constantly and fail in opposite ways. Tokenization now separates them: the analyzer kind has an open vocabulary, is configurable per field, and fails by silent recall loss; the model kind has a vocabulary frozen at pre-training and fails by fragmenting unknown terms into subwords. Out-of-Vocabulary owns the consequence and anchors the SPLADE / miniCOIL contrast — what a learned sparse model can represent is bounded by the vocabulary it was trained with, which is why domain terms it never saw behave differently from terms it merely saw rarely.

Concepts — Tokenization · Out-of-Vocabulary Updated — SPLADE · miniCOIL · Sparse Embeddings · Learned Sparse Retrieval · BM25 · Zero-Shot Retrieval · Full-Text Search · Query Understanding · Query Understanding in Practice · Multilingual Search · Search using PostgreSQL · Collocations · Query Segmentation · Query Understanding - Tokenization

2026-07-30 — Pooling split out of Token Pooling (1 new, 12 updated)

Six unrelated operations share the name pooling, and Token Pooling had been standing in for all of them. Pooling now owns sequence pooling specifically — the lossy collapse of per-token vectors into one, which is the step every Bi-Encoder depends on and the point where term-level detail is discarded. Bi-Encoder gained the pooling section it was missing, and Token Pooling is back to its narrow multi-vector-compression scope behind a disambiguation callout.

Corrections

Concepts — Pooling Updated — Token Pooling · Bi-Encoder · Personalization · SPLADE · Judgment Lists · Late Interaction · ColBERT · Vector Similarity Metrics · Vector Search Tradeoffs · Sentence Transformers · Three mistakes when introducing embeddings and vector search · Patterns for Personalization

2026-07-30 — MongoDB hybrid search series parts 1–2 (3 new, 7 updated)

Erik Hatcher’s hybrid search series treats measurement as a precondition rather than a follow-up: without judgments in place first, there is no way to tell whether fusion helped. Survey of the Hybrid Search Landscape supplies that framing and its rankability spectrum; Reciprocal Rank Fusion and Relative Score Fusion the arithmetic of the two fusion families. Score Normalization gets its own note, with sigmoid saturation as the core failure mode — scores far from the curve’s centre collapse toward 0 or 1, so a raw distance of 85.0 and a rating of 4.2 both normalize to roughly 1.0 and a whole pipeline’s gradations vanish.

Articles — Survey of the Hybrid Search Landscape · Reciprocal Rank Fusion and Relative Score Fusion Concepts — Score Normalization Updated — Reciprocal Rank Fusion · Relative Score Fusion · Hybrid Search · Semantic Boosting · Hybrid Search Blueprint Series Semantic Boosting · Erik Hatcher · MongoDB

2026-07-30 — User Behavior Insights (1 new, 5 updated)

User Behavior Insights is a clickstream capture standard — a JSON Schema for events plus a reference plugin in OpenSearch. The query_id is the load-bearing part: without a stable id linking a query to the events that followed it, click data cannot be attributed to the ranking that produced it. It is not the same thing as the labels derived from that stream — those are Implicit Judgments — and the two had been treated as one.

Corrections

Tools — User Behavior Insights Updated — Search Relevance Workbench · Relevance Evaluation Tools Compared · Implicit Judgments · Quepid · Search Evaluation

2026-07-30 — Brute-force vector search and silent fusion failure (10 new, 27 updated)

Most corpora are small enough that brute force beats an ANN index, and the machinery gets bought before it is needed — Doug Turnbull’s Just brute force your embeddings and Jo Kristian Bergum’s Three mistakes when introducing embeddings and vector search make that case from opposite directions, and Vector Search Tradeoffs holds the synthesis. The failure mode is quieter than the cost: Hybrid Fusion Failure - BM25 Displacing Reference Documents is a production account where adding a lexical branch to fix literal-keyword queries removed the correct document from the candidate set entirely, on exactly the query class it was meant to fix. The two brute-force benchmarks quoted are not comparable — different hardware, dimensions, and measurement — and both notes say so.

Corrections

  • HNSW’s memory figures failed arithmetic; traced to a 128-dim SIFT1M table copied without its dimensionality basis, and corrected there and in Dense Vector Retrieval.
  • Hybrid Search’s Elasticsearch section showed the unnormalized bool/should construct with no caveat; caveat added.

Articles — Just brute force your embeddings · Three mistakes when introducing embeddings and vector search Concepts — Brute-Force Vector Search · Zero-Shot Retrieval Topics — Vector Search Tradeoffs People — Roy Keyes · Davit Khachaturyan Tools — ann-benchmarks Case Studies — Hybrid Fusion Failure - BM25 Displacing Reference Documents Datasets — SIFT1M Updated — Hybrid Search · Linear Score Combination · BM25 · Reranking · Retrieval Pipeline · Relative Score Fusion · OpenSearch · Search Quality Assurance · Approximate Nearest Neighbor Search · Dense Vector Retrieval · Vector Search Evaluation · Vector Filtering · HNSW · IVF · LSH · Vector Quantization · MUVERA · ColBERT · Embedding Fine-tuning · FAISS · BEIR · MS MARCO · Vespa · Doug Turnbull · Jo Kristian Bergum · RRF is Not Enough · Dimensionality Reduction vs Quantization

2026-07-29 — Relevance feedback inside the index (5 new, 6 updated)

Relevance Feedback has always been applied to the query — expand it, reweight it, resubmit. Qdrant 1.17 applies it to the scoring function instead, modifying HNSW hop selection with feedback from previous rounds, which is the leg the literature skips because engines are usually black boxes. Written up from Evgeniya Sukhodolskaya’s Berlin Buzzwords 2026 talk and the two Qdrant posts behind it. The reported gains are measured against the feedback model itself as ground truth (abovethreshold@10) rather than human judgments, and several pairings regress — both caveats carried in the notes. BEIR gets a note of its own.

Articles — Relevance Feedback in Informational Retrieval · Relevance Feedback in Qdrant Videos — Evgeniya Sukhodolskaya - Relevance Feedback Inside the Search Engine Tools — qdrant-relevance-feedback Datasets — BEIR Updated — Relevance Feedback · HNSW · Evgeniya Sukhodolskaya · Qdrant · Berlin Buzzwords · MS MARCO

2026-07-27 — PCA vs t-SNE for retrieval (1 new, 4 updated)

t-SNE is disqualified from retrieval twice over, and PCA vs t-SNE for Retrieval separates the two reasons: it is non-parametric, so there is no learned transform to apply to an incoming query, and its KL objective preserves neighbourhoods rather than distances. UMAP shows the two disqualifiers are independent. Dimensionality Reduction vs Quantization gained the section correcting a false taxonomy — dimensionality reduction and quantization are independent multipliers on bytes = dims × bits/dim, not competing choices, and shrinking one does not stop you shrinking the other.

Topics — PCA vs t-SNE for Retrieval Updated — Dimensionality Reduction vs Quantization · PCA · t-SNE · UMAP

2026-07-27 — SPLADE domain fine-tuning (8 new, 7 updated)

Fine-tuning SPLADE on the Amazon ESCI Dataset reaches +27.5% over BM25 in-domain and loses out-of-domain — a model tuned on one catalog is not a general e-commerce model, and that transfer failure is the result rather than a footnote to it. Thierry Damiba’s five-part Fine-Tuning Sparse Embeddings for E-Commerce Search carries the procedure and the numbers; Evgeniya Sukhodolskaya’s MICES talk makes the same argument and deliberately quotes none. Hard Negative Mining gets its own note — the mining strategy is most of what separates a fine-tune that transfers from one that memorises.

Videos — Evgeniya Sukhodolskaya - Fine-Tuning Sparse Neural Retrievers for E-Commerce Articles — Fine-Tuning Sparse Embeddings for E-Commerce Search Concepts — Hard Negative Mining · miniCOIL People — Evgeniya Sukhodolskaya · Thierry Damiba Tools — Sentence Transformers · qdrant-sparse-finetune Updated — SPLADE · Learned Sparse Retrieval · Amazon ESCI Dataset · Sparse Embeddings · Embedding Fine-tuning · Qdrant · MICES

2026-07-27 — PCA embedding compression (2 new, 5 updated)

PCA now carries a measured recall-versus-dimensions curve instead of a rule of thumb: MiniLM on MS MARCO holds 0.879 recall cutting 384 dimensions to 200, then collapses to 0.2029 at 50. The usable ceiling works out around 1.9×, which tempers the “2–4× compression is common” heuristic in Dimensionality Reduction vs Quantization. The note also gains the flat-spectrum diagnostic — if the eigenvalue spectrum is flat there is no redundancy to harvest and PCA will not help, which is checkable before spending the compute. From Doug Turnbull’s Principal Component Analysis - an embedding shrink-ray; MS MARCO gets a note too.

Articles — Principal Component Analysis - an embedding shrink-ray Datasets — MS MARCO Updated — PCA · Dimensionality Reduction · Dimensionality Reduction vs Quantization · Doug Turnbull · Courses

Graph View

Backlinks