VectorChord v0.5.0

Two big upgrades, focused and pragmatic.

1) Experimental DiskANN (with RaBitQ) — vchordg (preview)

A new disk-backed graph index that keeps memory low while giving you a DiskANN-style option inside VectorChord.

• When it shines: can be faster than IVF+RaBitQ (vchordrq) on some embeddings (e.g., OpenAI/Cohere) — but not always.
• Caveats: slow build, and insert/delete are weaker than IVF. Dataset-dependent: benchmark before switching.
• Try it:

  CREATE INDEX ON items USING vchordg (embedding vector_l2_ops)
  WITH (options = $$
    m = 64
    ef_construction = 128
  $$);
  
  SET vchordg.ef_search = 128;

• Memory knob: bits = 1 halves index memory vs default bits = 2 (better recall/QPS).

We’re shipping this to give you a one-stop vector search toolbox at VectorChord. Feel free to share any thoughts and questions about it!

2) Recall measurement for IVF+RaBitQ — vchordrq_evaluate_query_recall

Approximate ≠ exact. Now you can quantify how close your results are with vchordrq_evaluate_query_recall. It accepts a query that returns row identifiers (e.g., ctid) and returns a recall score.

SET vchordrq.probes = '100';
SET vchordrq.epsilon = 1.0;

SELECT vchordrq_evaluate_query_recall(query => $$
  SELECT ctid FROM items
  ORDER BY embedding <-> '[3,1,2]'
  LIMIT 10
$$);  -- add ", exact_search => true" for table-scan ground truth

Note: recall evaluation targets vchordrq in 0.5 (not vchordg yet).

Other fixes

We fixed some performance regression problems in this release. User can enjoy better performances with it!

Talk to us

Thanks for building with us. If you have any question or thoughts, open an issue, join our discord or start a Discussion. Your notes guide what we fix first. If VectorChord helped you, drop us a ⭐ on GitHub and hit Watch → Releases.