Retrieval at Scale | Drop for 2025-11-08

TL;DR

Since Oct 31, 2025: Weaviate 1.34.0 shipped with Flat-index Rotational Quantization (1‑bit/8‑bit, preview), server‑side dynamic batching (preview), a new SPFresh vector index, and ACORN as the default filtered‑ANN strategy—useful for multi‑vector/late‑interaction and filter‑heavy workloads. Three fresh research preprints push filtered ANN beyond simple predicates: PathFinder (conj/disj across multiple attributes with cost‑based planning), a filter‑aware learned distance for multi‑filter search, and Compass (a general hybrid execution that couples vector and relational indexes). Plus, the community announced the first Late Interaction & Multi‑Vector workshop (ECIR 2026).

Weaviate 1.34.0 GA: Flat-index RQ, server-side dynamic batching (preview) and SPFresh

  • Key facts and current state of the topic
    • Weaviate 1.34.0 (Nov 5) adds Flat-index Rotational Quantization (1‑bit and 8‑bit, preview) to shrink vector memory; introduces server‑side dynamic batching (preview) to stabilize throughput under bursty QPS; ships SPFresh, a new vector index aimed at fresher, on‑disk–friendly search; and sets ACORN as the default filter strategy, while also landing HNSW snapshot improvements. (github.com)
  • Important context and background information
    • ACORN‑style filtered HNSW and quantization are increasingly standard for filtered ANN at scale; having ACORN as default and Flat‑RQ in Flat indexes broadens the footprint/latency knobs for hybrid and late‑interaction pipelines. (github.com)
  • Recent developments or changes
    • Released November 5, 2025. Evaluate Flat‑RQ vs. BBQ/PQ on your embeddings; pilot server‑side batching for p95/p99 control; and test SPFresh if you need continuous updates with predictable tail latencies. (github.com)

PathFinder: efficient filtered‑ANN with conjunctions and disjunctions

  • Key facts and current state of the topic
    • Many filtered‑ANN approaches are strong only for single‑attribute equality filters; PathFinder targets complex Boolean filters (AND/OR) via a framework of attribute‑specific indexes plus a cost‑based optimizer. (arxiv.org)
  • Important context and background information
    • The method introduces a new optimization metric, two‑phase optimization for complex filters, and “index borrowing” to reuse attribute‑specific indexes for other attributes. (arxiv.org)
  • Recent developments or changes
    • Preprint posted November 2, 2025; authors report up to 9.8× throughput at 0.95 recall vs. strong baselines on four real‑world datasets. Consider this line of work if your retrieval integrates multiple structured predicates. (arxiv.org)

Learning filter‑aware distance metrics for multi‑filter ANN

  • Key facts and current state of the topic
    • Instead of fixed penalties for filter violations, this work learns data‑driven weights that balance vector similarity and filter satisfaction, using them for both graph construction and search. (arxiv.org)
  • Important context and background information
    • Fixed, predicate‑agnostic heuristics often fail to generalize across datasets with different label/feature distributions; learning the trade‑off can better capture real filter semantics. (arxiv.org)
  • Recent developments or changes
    • Preprint posted November 6, 2025; reports 5–10% accuracy gains over fixed‑penalty methods on multi‑filter benchmarks—promising for high‑selectivity ad/search use cases. (arxiv.org)

Compass: a general hybrid (vector + relational) filtered‑search framework

  • Key facts and current state of the topic
    • Compass coordinates candidate generation and predicate evaluation across existing vector (HNSW/IVF) and relational (B+‑tree) indexes to support arbitrary conjunctions, disjunctions, and ranges without specialized “filtered ANN” indexes. (arxiv.org)
  • Important context and background information
    • Generality and DBMS compatibility are the focus; prior “general” systems often struggled with performance or restricted predicates. (arxiv.org)
  • Recent developments or changes
    • Preprint posted October 31, 2025; authors claim consistent wins over NaviX and parity with specialized single‑attribute indexes in their best case—worth tracking if you prefer hybrid execution inside a general DB stack. (arxiv.org)

Late Interaction community milestone: LIR workshop (ECIR 2026)

  • Key facts and current state of the topic
    • “LIR: The First Workshop on Late Interaction and Multi‑Vector Retrieval” was announced for ECIR 2026, reflecting momentum around ColBERT‑style token‑level matching, efficiency, and system integration. (arxiv.org)
  • Important context and background information
    • The focus explicitly includes real‑world outcomes and negative/puzzling results—useful for teams moving late‑interaction models from papers to production. (arxiv.org)
  • Recent developments or changes
    • Call posted November 1, 2025; consider submitting production lessons (e.g., index pruning, memory‑mapped serving, filtered‑ANN stages) or attending to compare approaches with peers. (arxiv.org)