March 24, 2026
Acorns, outrage, and vector karma
Show HN: DuckDB community extension for prefiltered HNSW using ACORN-1
Fork makes 'top 10' actually return 10—HN chants 'upstream!' while others beg for binary support
TLDR: A community-made DuckDB add‑on fixes a long‑annoying filtered search bug by pushing filters inside the index search, so “top 10” finally returns 10. Commenters are clamoring to merge it upstream, while others pivot to a wishlist for efficient binary vectors—setting up a “ship it now vs roadmap” tug‑of‑war.
HN just got a new toy: a community fork of DuckDB’s vector search extension that fixes the “top 10” heartbreak. Instead of filtering results after a search (which often left you with 0–2 matches), this fork pushes the filter into the search itself using ACORN‑1, a trick that keeps the graph connected and delivers the full set. Translation: ask for 10 similar movies in Japanese? You finally get 10. Benchmarks show a clean sweep across tough filters, and people are buzzing.
The loudest chant is simple: “Upstream it!” One commenter tossed the first stone, and the crowd piled on with “merge it yesterday” energy, eyeing the duckdb‑vss repo. Another camp immediately turned the thread into a wishlist: what about efficient binary vectors? Cue the roadmap tug‑of‑war. Meanwhile, the meme factory kicked in—ACORN jokes (“squirrels saving your neighbors”), Matrix puns (red‑pill your search), and a healthy dose of “this feels like a fix, not a feature.”
Under the hood made simple: HNSW (a fast “shortcut map” for nearest neighbors) now respects your WHERE filters as it searches. It even switches strategies based on how rare your filter is, and falls back to exact search when things get super picky. For users, it means fewer facepalms and more results that actually make sense. For DuckDB fans, it’s drama worth watching.
Key Points
- •Fork adds ACORN-1 filtered HNSW traversal to DuckDB’s vector search, pushing WHERE predicates into graph search.
- •Selectivity-based strategy: >60% uses post-filter, 1–60% uses ACORN-1, <1% switches to brute-force exact scan.
- •Two-hop expansion through failed neighbors (ACORN-1) and a Lucene-inspired 90% rule improve connectivity and efficiency.
- •Optimizer auto-detects WHERE + ORDER BY distance + LIMIT patterns; prepared statements supported; subquery vectors fall back to scan.
- •Benchmarks (228k movies, 768-dim Nomic) show upstream returns 0–1/10 under selective filters, ACORN-1 returns 10/10.