TL;DR
Three relevant updates since Dec 18, 2025: (1) Faiss v1.13.2 ships multi‑bit RaBitQ FastScan, Panorama refinements, and new backends (cuVS + Intel SVS) for faster, cheaper ANN at scale; (2) Elastic Stack 9.2.3 (Dec 19) delivers security fixes affecting Elasticsearch/Kibana—recommended for Lucene‑based retrieval stacks; (3) new research proposes quantization that stays accurate under streaming inserts/deletes, reducing costly index rebuilds.
Faiss v1.13.2: multi‑bit RaBitQ FastScan, Panorama refine, cuVS + Intel SVS
- Key facts and current state of the topic
- Faiss is a primary ANN engine used in large‑scale retrieval; recent work (PANORAMA) targets the verification stage, while RaBitQ improves binary quantization fidelity. (github.com)
- Important context and background information
- Lower‑bit, higher‑fidelity quantizers and faster verification let you probe more candidates within fixed latency/memory, improving recall, especially under filters or hybrid pipelines. (github.com)
- Recent developments or changes
- Released Dec 19, 2025: adds multi‑bit RaBitQ FastScan (IVF and Flat variants), IndexRefinePanorama, enables NVIDIA cuVS interop and Intel ScalableVectorSearch, plus backward‑compatible Panorama serialization. Consider A/Bs vs. PQ/previous RaBitQ at target recalls. (github.com)
Elastic Stack 9.2.3 (Dec 19): security patches for Elasticsearch/Kibana
- Key facts and current state of the topic
- Many production retrieval systems run on Elasticsearch/Lucene; 9.2.x also introduced vector‑search improvements in earlier minors. (elastic.co)
- Important context and background information
- Medium‑severity vulnerabilities addressed on Dec 18 (ESA‑2025‑37/38) affect Elasticsearch/Kibana; patching is advisable to avoid DoS or privilege escalation in managed clusters and self‑hosted stacks powering candidate generation or hybrid search. (discuss.elastic.co)
- Recent developments or changes
- 9.2.3 (Dec 19) is the recommended upgrade; review release notes and security advisories, schedule maintenance, and verify plugin/API compatibility. (elastic.co)
Quantization under streaming updates: toward fresh, accurate on‑disk indexes
- Key facts and current state of the topic
- Most ANN systems quantize vectors in RAM and keep full‑precision on disk; data‑dependent quantizers degrade as data evolves or force expensive rebuilds. (arxiv.org)
- Important context and background information
- Freshness is critical for catalogs/ads; the goal is to update codes with bounded I/O and stable recall/latency without full re‑training/re‑building. (arxiv.org)
- Recent developments or changes
- Dec 20 preprint formalizes “dynamic consistency” for streaming quantization and proposes a practical method that adapts codes online, showing superior ANN accuracy under inserts/deletes. Pilot where continuous ingest currently triggers frequent rebuilds. (arxiv.org)