TL;DR
- Elastic 9.5 is GA with a VectorDB index mode and auto‑calibrated DiskBBQ settings that reduce vector tuning toil and improve out‑of‑the‑box performance.
- OpenSearch 3.8 delivers up to 4.2× faster vector ingestion and 2.1× higher radial‑search throughput, plus broader agent/streaming support.
- Qdrant 1.19.0 adds a 4‑bit TurboQuant “datatype” (no full‑precision fallback) and memory tiers for denser indexes at lower cost.
- PULSAR shows a production late‑interaction system for visual document RAG: pooled multi‑vector indexing cuts median vector latency ~15× and sustains ~88× higher QPS at similar quality.
- A new comparative study benchmarks seven vector engines (Faiss, Qdrant, Milvus, Weaviate, pgvector, LanceDB, Chroma) with actionable system‑selection guidance.
Elastic 9.5 GA: VectorDB index mode + auto‑calibration for DiskBBQ
- Key facts and current state of the topic
- Elastic 9.5 is generally available; highlights include Columnar Mode (tech preview) and new vector features. (elastic.co)
- Important context and background information
- Operating vector search at scale has meant expert tuning (quantization depth, merge policies, cache warmups). Disk‑oriented formats like DiskBBQ help memory but add more knobs. (elastic.co)
- Recent developments or changes
- “VectorDB index mode” ships tuned defaults for vector workloads; “auto‑calibration” for DiskBBQ automatically sets quantization, preconditioning, and oversampling based on index statistics, simplifying rollout while preserving quality. (elastic.co)
OpenSearch 3.8: faster vector ingestion and higher radial‑search throughput
- Key facts and current state of the topic
- OpenSearch 3.8 (Aug 4, 2026) is live with multiple vector‑engine improvements. (docs.opensearch.org)
- Important context and background information
- Ingestion and candidate‑generation throughput bound freshness and p95/p99 in large hybrid/filtered systems; radial (within‑radius) search is common in safety thresholds and de‑duplication. (docs.opensearch.org)
- Recent developments or changes
- Up to 4.2× faster ingestion via Base64‑encoded vectors; up to 2.1× higher radial‑search throughput with improved recall; broader agentic/streaming support (gRPC streaming, MCP tooling). See also project release notes. (docs.opensearch.org)
Qdrant 1.19.0: TurboQuant “datatype,” memory tiers, filter upgrades
- Key facts and current state of the topic
- Qdrant 1.19 (Aug 5) introduces a 4‑bit TurboQuant storage format that doesn’t retain full‑precision vectors, plus memory‑tiering and filtering enhancements. (qdrant.tech)
- Important context and background information
- Lower‑bit codes sharply shrink footprint and can raise probe budgets at fixed latency; removing full‑precision storage cuts cost but shifts dependence to quantizer fidelity. (qdrant.tech)
- Recent developments or changes
- TurboQuant “datatype” targets up to ~9× storage reduction vs. prior TurboQuant; per‑tenant IDF stats and prefix‑match filters land; Helm/operator updates validate 1.19.0. Press coverage summarizes the release. (qdrant.tech)
PULSAR (Aug 28): pooled late‑interaction for visual‑doc RAG in production
- Key facts and current state of the topic
- PULSAR (arXiv) describes a deployed system for visual document RAG using a frozen ColPali‑style backbone and a two‑stage, pooled late‑interaction index. (arxiv.org)
- Important context and background information
- Late‑interaction (multi‑vector) yields strong retrieval on layout‑rich pages but can be memory/latency heavy; pooling and staged rescoring aim to keep MaxSim’s quality while lifting QPS. (arxiv.org)
- Recent developments or changes
- On ViDoRe V3, pooled indexing reduces median vector‑search latency ~15.1× vs. unpooled with <0.01 absolute NDCG@10/Recall@10 loss; in production, median vector latency is 156 ms and the system sustains ~88× higher QPS under load at similar quality. (arxiv.org)
Benchmarking study: practical guidance on engine choices
- Key facts and current state of the topic
- A new empirical evaluation compares Faiss, Qdrant, Milvus, Weaviate, pgvector, LanceDB, and Chroma across six datasets (96–960 dims; >4M vectors). (arxiv.org)
- Important context and background information
- Findings reinforce well‑known trade‑offs: Faiss tops single‑node throughput; Qdrant shows leading latency among full databases; Weaviate provides strong out‑of‑the‑box recall—useful signals for first‑stage candidate stores feeding re‑rankers. (arxiv.org)
- Recent developments or changes
- The paper offers system‑selection guidelines and an open benchmarking framework, enabling A/Bs under your embeddings and filter selectivities before committing to a platform. (arxiv.org)