Retrieval at Scale | Drop for 2026-09-23

TL;DR

  • Faiss 1.15.1 (Sept 16) lands new quantization and HNSW features (e.g., staged RaBitQ storage, SVE fast paths, IVFPQ distance modes), plus zero‑copy deserialization hooks.
  • Milvus 3.0.2 (Sept 20) focuses on faster filtered search/group‑by and sturdier lake‑native ops (External Collections, Storage V2), with planner and I/O improvements.
  • Weaviate 1.39.x patches (Sept 7–18) bring aggregation speedups, safer GraphQL/module paths, HFresh/HNSW fixes, structured reference selection in the new Search REST API, and RQ4‑centered quantization tweaks.
  • Elastic 9.5.3/9.5.4 (Sept 3/15) resolve a serious 9.5.2 replication divergence issue; 9.5.4 is the current recommended patch for Lucene‑backed vector/hybrid stacks.

Faiss 1.15.1: more levers for quantization and HNSW

  • Key facts and current state of the topic
    • Faiss remains a primary ANN engine; recent lines added TurboQuant/SQ paths, Panorama verification, Metal/ARM kernels. (github.com)
  • Important context and background information
    • Large‑scale retrieval benefits from better binary/low‑bit quantization and faster HNSW verification/storage to probe more candidates at fixed SLAs.
  • Recent developments or changes
    • v1.15.1 (Sept 16, 2026) adds staged RaBitQ storage to HNSW, ZeroCopyIOReader for RaBitQ deserialization, explicit IVFPQ scanner distance modes, SVE low‑dim fast paths, and CI for Windows ARM64. See the changelog for the full list. (github.com)

Milvus 3.0.2: faster filtered search/group‑by, sturdier lake‑native ops

  • Key facts and current state of the topic
    • Milvus 3.0 introduced a lake‑native architecture (External Collections, Snapshots, Storage V3) and multi‑vector ergonomics. (github.com)
  • Important context and background information
    • At ads/search scale, hot paths are filtered ANN, group‑by/aggregation, and index build I/O; external tables and S3‑compat endpoints help reduce ETL and improve freshness.
  • Recent developments or changes
    • v3.0.2 (Sept 20, 2026) removes hot‑path contention in filtered search/group‑by/index builds; adds custom S3‑compatible endpoints for External Collections; unifies Bloom/Roaring membership filters; and trims text/JSON index I/O. (github.com)

Weaviate 1.39.x: aggregation speedups, safer modules/GraphQL, HFresh/HNSW fixes

  • Key facts and current state of the topic
    • 1.39.0 GA brought Namespaces, Hybrid MMR diversity, a Search REST API, and preview RQ4 compression. (weaviate.io)
  • Important context and background information
    • Multi‑tenant, hybrid, and late‑interaction stacks need predictable tails; patch releases that harden aggregation, replication, and HFresh/HNSW behavior matter directly at p95/p99.
  • Recent developments or changes
    • v1.39.3 (Sept 7) adds structured reference selection in the REST API, aggregation routing/count fixes, RQ4 centered quantization, and multiple async‑replication/LSMKV performance tweaks; v1.39.4 adds further HFresh recovery and “batched Contains” filter speedups. (newreleases.io)
  • Key facts and current state of the topic
    • Elasticsearch powers many Lucene‑based vector/hybrid stacks (ACORN‑style filtered HNSW, BBQ/SQ quantization). Patch cadence affects stability and correctness. (elastic.co)
  • Important context and background information
    • A known issue in 9.5.2 could process some bulk items on the primary without replication, causing shard divergence and long recoveries. Elastic advised upgrading to 9.5.3 ASAP. (elastic.co)
  • Recent developments or changes
    • 9.5.3 (Sept 3) resolves the issue; 9.5.4 (Sept 15) is now available with additional fixes. If you’re on any 9.5.x prior to .3/.4, plan an expedited upgrade and validation across vector + lexical paths. (elastic.co)

Weaviate Search 1.39: GA recap (MMR diversity, Boost API) with ongoing hardening

  • Key facts and current state of the topic
    • The 1.39 GA post also finalized MMR diversity for vector/hybrid and GA’d the Boost API for query‑time re‑scoring—useful knobs before heavy re‑rankers. (weaviate.io)
  • Important context and background information
    • Diversity and lightweight boosts can raise early‑stage utility without reindexing, reducing downstream cross‑encoder/LLM spend.
  • Recent developments or changes
    • Combine GA features with the 1.39.3/.4 patches above for lower duplication in results and steadier aggregation under load; track incremental 1.39.x notes as they land. (newreleases.io)