11 articles tagged “Vector search”.
Dot product, cosine, Euclidean, Manhattan and Hamming - what each one actually measures, why most of them collapse into the same ranking once your vectors are normalised, and the handful of mistakes that bite in practice.
A small, self-contained experiment on why single-vector retrieval breaks on compound, high-intent queries - and how late interaction keeps two facets intact where a dense embedding averages them away.
Ben Trent and Thomas Veasey shipped another DiskBBQ optimisation today: quantising queries against coarser parent centroids instead of per-document ones. 5x off the quantisation stack with no meaningful recall loss. The insight underneath is that the query path and the document path don’t have to be treated symmetrically.
A new paper out this week shows the per-token hidden states of off-the-shelf single-vector embedders already carry the information needed for ColBERT-style MaxSim - and you can wire it in at inference time, without retraining. The late-interaction deployment barrier I most underestimated just dropped.
Elasticsearch 9.4 adds query_vector_builder.lookup - a tiny API addition that collapses a two-request vector search into one and runs better than 3x faster. A small change with a big impact, and a look at where that ratio actually comes from.
SID AI’s SID-1 is the first retrieval model trained end-to-end with RL. Some observations through a search-and-IR lens: the middle of the retrieval pipeline collapses into one trained model, the NDCG reward gets deliberately bent toward recall, and the agentic-retrieval loop becomes a subagent you hand to a larger system.
xAI open-sourced the For You feed algorithm today. Three observations through a search-and-IR lens: two-tower’s quiet dominance, the retrieve/rank split surviving the bitter lesson, and recsys converging with search.
TurboQuant landed as a KV cache result, but the more interesting application might be ColBERT-style late interaction. Here’s the case, and the open questions.
A new quantisation method out of Google Research is making the rounds. Qdrant shipped it. Elastic ran the benchmarks and politely declined. Both responses tell you something useful.
DiskBBQ’s new filtered-search optimisation is the same architectural move I wrote about last week, applied one layer deeper. The pattern is fractal - and that’s what makes it useful.
How ColBERT-style token-level matching fits between single-vector dense retrieval and cross-encoders, why MaxSim is the clever bit, and what the storage tax actually looks like in practice.
