Retrieval at Scale | Drop for 2026-02-03

TL;DR

Post–Jan 26, 2026 highlights: Weaviate 1.35.5/1.35.6 land async‑replication and backup/ops fixes (on top of 1.35’s TTL and flat‑RQ GA); Faiss mainline adds RaBitQ FastScan IVF support and other low‑level improvements—useful ahead of the next tag; Elastic published a practical guide to query rewriting for recall/precision gains in hybrid/LLM‑assisted search; OpenSearch 3.5 entered its release window (Jan 27–Feb 10)—watch for vector/agentic updates; new GRIT‑VQ work proposes a differentiable framework for vector quantization that may inform future ANN quantizers.

Weaviate 1.35.5–1.35.6: async replication and ops hardening (plus 1.35 feature recap)

  • Key facts and current state of the topic
    • Weaviate shipped 1.35.5 (Jan 28) and 1.35.6 (Jan 29) with improvements to async replication targets, backup metrics, and safety around compressed‑vector directories—reducing tail‑latency risks and smoothing cluster operations. (github.com)
  • Important context and background information
    • The 1.35 GA post (Jan 29) re‑highlights Object TTL (lifecycle control), flat‑index Rotational Quantization (RQ) at GA for footprint reduction, and operational modes—useful for multi‑vector/late‑interaction pipelines and cost control. (weaviate.io)
  • Recent developments or changes
    • If you’re running filtered ANN + late‑interaction, consider upgrading to 1.35.6 for replication stability and pair it with TTL/RQ to bound memory and storage. (github.com)

Faiss mainline (late Jan): RaBitQ FastScan IVF support and verification/infra tweaks

  • Key facts and current state of the topic
    • Faiss main merged several changes after 1.13.2: InvertedListScanner support for IndexIVFRaBitQFastScan, generic result handlers, SVS runtime packaging, and a default qb change for RaBitQ—continuing the recent push on binary/multi‑bit quantization and verification speed. (github.com)
  • Important context and background information
    • These come on top of December’s Panorama integration and multi‑bit RaBitQ; building from main can preview gains before the next tagged release—relevant if you’re quantizing to probe more candidates at fixed latency. (github.com)
  • Recent developments or changes
    • Plan A/B tests for FastScan IVF + RaBitQ versus your current PQ/RaBitQ settings at target recall; watch for the next Faiss tag to consume via wheels. (github.com)

Elastic: practical query rewriting patterns for recall/precision (Jan 30)

  • Key facts and current state of the topic
    • Elastic’s Search Labs outlined concrete query‑rewriting strategies that measurably improve results for hybrid search and LLM‑assisted systems (e.g., synonym/expansion, intent‑aware rewrites). (elastic.co)
  • Important context and background information
    • For ads retrieval, disciplined rewriting often lifts first‑stage recall without inflating ANN budgets, and provides stronger candidates for re‑rankers. (elastic.co)
  • Recent developments or changes
    • Treat the post as an implementation guide: couple rewritten lexical queries with vector candidates, then apply reciprocal rank fusion or learned re‑rankers. (elastic.co)

OpenSearch 3.5 release window opened (heads‑up)

  • Key facts and current state of the topic
    • OpenSearch 3.5’s release window runs Jan 27–Feb 10, 2026. Expect incremental vector/agentic‑search enhancements following 3.4’s UX/perf changes. (opensearch.org)
  • Important context and background information
    • If you depend on Amazon OpenSearch Service, track 3.5 GA to align managed‑service upgrades and regression tests for vector/hybrid workloads. (forum.opensearch.org)
  • Recent developments or changes
    • Prepare upgrade playbooks (index snapshots, A/B on recall/latency, filter selectivity tests) so you can adopt promptly once 3.5 lands. (opensearch.org)

GRIT‑VQ (Feb 1): differentiable vector‑quantization framework

  • Key facts and current state of the topic
    • A new preprint proposes GRIT‑VQ, a differentiable VQ framework with radius‑based updates and integrated codebook transforms; authors report better reconstruction/codebook utilization across tasks. (arxiv.org)
  • Important context and background information
    • Although framed around generative/representation models, the ideas (stable gradient flow, coordinated codebook evolution) could inform high‑fidelity ANN quantization schemes. Validate on retrieval embeddings before adoption. (arxiv.org)
  • Recent developments or changes
    • Track for potential ANN index‑build improvements (e.g., faster encoding, higher‑accuracy low‑bit codes) and compare against RaBitQ/SAQ/NVQ baselines in your stack. (arxiv.org)