TL;DR
- Azure SQL made DiskANN vector indexes generally available across Azure SQL Database/MI and Fabric SQL, bringing on-disk ANN to a mainstream relational engine.
- Weaviate shipped multiple 1.39.x patches that harden HNSW/ACORN/HFresh and released 1.40.0-rc.1 with preview features for drop‑vector‑index, HFresh MUVERA, and 4‑bit RQ.
- Cohere launched Embed 5 (Pro/Fast) with a shared embedding space and published RCP‑nDCG@10, a new retrieval metric; Cohere also announced Compass “coming to the cloud,” a managed search/retrieval offering.
Azure SQL: DiskANN vector indexes reach GA (on-disk ANN inside SQL)
- Key facts and current state of the topic
- Microsoft announced general availability of DiskANN Vector Index & Search in Azure SQL Database, Azure SQL Managed Instance (auto‑update policy), and Fabric SQL. It enables approximate vector search directly in the SQL engine with T‑SQL composability. (devblogs.microsoft.com)
- Important context and background information
- DiskANN reduces RAM dependence by using SSD‑backed indexes with small memory footprints, important for billion‑scale catalogs and filter‑heavy workloads where candidate budgets are tight. Prior DiskANN integrations (e.g., SQL Server preview) were covered earlier; this GA brings it to managed PaaS. (devblogs.microsoft.com)
- Recent developments or changes
- GA includes optimizer‑driven plan selection (exact vs. ANN), iterative filtering with relational predicates, async maintenance for INSERT/UPDATE/DELETE, and a T‑SQL “WITH APPROXIMATE” pattern—simplifying hybrid (vector + SQL) pipelines. (devblogs.microsoft.com)
Weaviate September updates: stability patches + next‑gen preview
- Key facts and current state of the topic
- Weaviate 1.39.x continues hardening high‑QPS vector/hybrid serving; 1.40.0‑rc.1 previews retrieval/compression features. (github.com)
- Important context and background information
- Late‑interaction and multi‑vector stacks are sensitive to HNSW/ACORN traversal, disk‑oriented HFresh behavior, replication, and backup paths; small engine changes materially affect p95/p99. (github.com)
- Recent developments or changes
- 1.39.6 (Sep 22) and 1.39.7 (Sep 25) ship LSM‑store performance work and targeted fixes across HNSW (ACORN), HFresh, async replication, and backup; 1.40.0‑rc.1 (Sep 15) introduces preview support for “Alter Schema → Drop vector index,” HFresh MUVERA, and 4‑bit Rotational Quantization. Evaluate patches in prod and trial RC features in non‑prod. (github.com)
Cohere Embed 5: new embedding family (Pro/Fast) with shared space
- Key facts and current state of the topic
- Cohere released Embed 5, positioning it as a new SOTA embedding family; Pro and Fast share one embedding space so a corpus indexed with one can be queried with the other. (docs.cohere.com)
- Important context and background information
- Shared spaces reduce re‑indexing cost and unlock dynamic latency/quality trade‑offs in production (e.g., index with Fast, query critical paths with Pro). Documentation also highlights multimodal capabilities across recent embed families. (cohere.com)
- Recent developments or changes
- Launch materials emphasize retrieval quality and operational flexibility; assess Embed 5 vs. your current models on domain corpora, especially under compression and filters. (cohere.com)
Cohere Compass “coming to the cloud”: managed search/retrieval
- Key facts and current state of the topic
- Cohere announced that Compass, its search/retrieval engine, is “coming to the cloud,” offering managed access without running infrastructure. (cohere.com)
- Important context and background information
- For teams standardizing on external retrieval services, Compass‑as‑a‑service could simplify pilots and A/Bs versus self‑hosted vector/lexical stacks, especially if combined with Embed 5 and managed reranking. (cohere.com)
- Recent developments or changes
- Announcement post is a product signal; validate roadmap, filter semantics, and latency/throughput SLAs before considering production migration. (cohere.com)
Retrieval evaluation: Cohere proposes RCP‑nDCG@10
- Key facts and current state of the topic
- Cohere published RCP‑nDCG@10, arguing it’s a more complete metric for enterprise retrieval relevance (relative coverage penalty accounting). (cohere.com)
- Important context and background information
- Standard metrics (e.g., nDCG) can miss coverage gaps in multi‑facet/business‑constraint scenarios; improved metrics help tune early‑stage recall and re‑ranking budget. (cohere.com)
- Recent developments or changes
- Consider adding RCP‑nDCG@10 to your offline evals alongside existing metrics; compare conclusions vs. nDCG/Recall to detect under‑coverage that impacts downstream ranking. (cohere.com)