Retrieval at Scale | Drop for 2026-10-23

TL;DR

  • Oracle added multi‑vector HNSW search, PQ‑compressed HNSW, GraphRAG, and IVF scalar quantization to Oracle AI Vector Search (26ai RU 23.26.4).
  • Milvus 3.0.3 brings function‑chain hybrid search, element‑level partial updates (Array/Struct/JSON), and nullable‑vector filters.
  • Weaviate 1.40.1 shipped security/perf fixes; the 1.40 line (GA earlier this month) continues to harden HNSW/HFresh and the new Search REST API.
  • pgvector 0.8.7 fixes a critical IVFFlat build overflow (CVE), upgrade recommended.
  • DuckDB 2.0.0 LTS is out; combined with the ACORN‑1 extension, it strengthens an embedded SQL path for filtered‑ANN and quantized vector search.

Oracle AI Vector Search (Oct RU): multi‑vector HNSW, PQ for HNSW, GraphRAG, IVF scalar quantization

  • Key facts and current state of the topic
    • Oracle’s 26ai vector stack (inside Oracle Database) now includes: multi‑vector similarity search on HNSW, product quantization (PQ) for HNSW, GraphRAG retrieval, snapshots for distributed HNSW, and scalar quantization for IVF. These features target lower memory/latency at large scale and better multi‑hop retrieval. (docs.oracle.com)
  • Important context and background information
    • Bringing PQ into HNSW shrinks RAM while preserving navigability; multi‑vector HNSW helps token/patch‑level candidates feed late‑interaction rerankers; GraphRAG integrates graph traversal with vector search to improve multi‑hop evidence gathering. (docs.oracle.com)
  • Recent developments or changes
    • All landed in the October 2026 Release Update 23.26.4 (Oracle AI Vector Search docs). Consider A/Bs vs. your current HNSW/PQ or hybrid pipelines. (docs.oracle.com)

Milvus 3.0.3: function‑chain hybrid search, element‑level partial updates, nullable‑vector filters

  • Key facts and current state of the topic
    • Milvus 3.0 focuses on lake‑native retrieval and richer search chains. v3.0.3 adds Function Chain support to Hybrid Search (fusion + reranking), element‑level partial updates (Array/Struct/JSON), and IS NULL/IS NOT NULL filters for nullable vector fields—plus security/stability fixes. (github.com)
  • Important context and background information
    • Function chains help shape multi‑stage retrieval inside the engine (vector, sparse/BM25, reranks); fine‑grained partial updates reduce rebuilds in constantly changing catalogs. (github.com)
  • Recent developments or changes
    • Released October 14, 2026; see official GitHub notes for features and breaking‑change warnings (legacy REST V1 routes, binlog import privileges). (github.com)

Weaviate 1.40.x: security and serving hardening after 1.40 GA

  • Key facts and current state of the topic
    • After 1.40.0 GA (Namespaces, HFresh MUVERA, 4‑bit RQ), Weaviate 1.40.1 shipped mid‑October with targeted security and stability fixes (e.g., dependency updates, queue/ID handling), with mirrors and trackers listing the build. (sourceforge.net)
  • Important context and background information
    • The 1.40 line also finalizes the new Search REST API and continues to harden HNSW/HFresh for high‑QPS hybrid and multi‑vector retrieval. (docs.weaviate.io)
  • Recent developments or changes
    • Pick up 1.40.1 (and subsequent point patches as they appear) to reduce p95/p99 variance; SourceForge and change trackers list artifacts and diffs while the docs page tracks the 1.40 series. (sourceforge.net)

pgvector 0.8.7: critical IVFFlat build overflow fix (security)

  • Key facts and current state of the topic
    • pgvector 0.8.7 addresses a buffer‑overflow vulnerability in IVFFlat index builds (CVE‑2026‑103484) that could allow arbitrary code execution. (postgresql.org)
  • Important context and background information
    • Postgres + pgvector is a common path for hybrid SQL + vector retrieval; unsafe index builds can compromise data and downstream services. (postgresql.org)
  • Recent developments or changes
    • Posted October 5, 2026 on postgresql.org; upgrade clusters and rebuild affected indexes as needed. (postgresql.org)

DuckDB 2.0.0 LTS: stronger embedded SQL foundation for filtered‑ANN (ACORN‑1) and quantization workflows

  • Key facts and current state of the topic
    • DuckDB 2.0.0 LTS was released October 21, 2026, with long‑term support and broad engine upgrades. In the DuckDB ecosystem, the hnsw_acorn extension adds ACORN‑1 filtered‑HNSW traversal and RaBitQ quantization directly in SQL—useful for embedded/lakehouse retrieval. (duckdb.org)
  • Important context and background information
    • ACORN‑class filtered‑ANN maintains recall under selective predicates by pushing filters into graph traversal; pairing it with low‑bit quantization raises candidate budgets at fixed latency. (duckdb.org)
  • Recent developments or changes
    • With 2.0 LTS and the extension, teams can prototype filter‑correct, quantized vector retrieval co‑located with analytics—reducing glue code and data movement. (duckdb.org)