Hacker News surfaced a DuckDB community extension project that takes aim at a very practical vector search problem: filtered nearest-neighbor queries that stop being useful once the database applies WHERE clauses after the HNSW index has already chosen candidates. The linked GitHub repository is a fork of duckdb-vss and argues that SQL-native retrieval needs to respect filters during graph traversal, not after it.

The extension adds ACORN-1 filtered search so predicates are pushed into HNSW traversal. In plain language, that means a query such as "top 10 vectors inside category X" can return a real top 10 inside that subset instead of a smaller, distorted result set. The README describes a selectivity strategy: high-selectivity queries can stay on standard HNSW, mid-selectivity queries use ACORN-1 two-hop expansion, and extremely selective queries fall back to brute-force exact scan. That kind of switching matters in production retrieval pipelines because filters are often as important as vector distance.

The second addition is RaBitQ quantization. The project claims the index can store vectors at 1 bit per dimension, then rescore candidates against original F32 vectors for final ranking. Reported memory savings range from roughly 21x at 128 dimensions to 30x at 768 dimensions, with benchmark tables showing recall gains when oversampling and rescoring are enabled. Even if users treat those numbers as repository benchmarks rather than neutral evaluations, the direction is clear: bring vector compression and filtered search into the analytical database instead of exporting everything to a dedicated vector store.

Hacker News paid attention because this is exactly the kind of plumbing that determines whether retrieval-augmented systems remain simple or split into multiple infrastructure layers. The project still has constraints, including RAM-resident indexes, FLOAT-only arrays, and some query shapes that fall back to sequential scan, but it targets a pain point many RAG teams have hit already.

Key points

ACORN-1 pushes filters into HNSW traversal so filtered queries can return full limits.
RaBitQ adds aggressive vector compression with exact-distance rescoring.
The project uses different search strategies based on filter selectivity.
It strengthens the case for keeping retrieval workloads inside DuckDB and SQL.

#acorn

Hacker News highlights DuckDB vector search fixes with ACORN-1 and RaBitQ