Skip to content

enzyme v0.13.4

Latest

Choose a tag to compare

@jshph jshph released this 11 Nov 19:57

Enzyme release notes - v0.13.4

Short version

Catalyst links are now scored when you search instead of being precomputed,
so indexes are about half the size and refresh after edits is roughly 10×
faster. Every note is embedded by default, stored vectors record which model
made them, and the embedding backend moves to encoders 3. A new
enzyme compact command returns the freed space to disk.

What shipped

Catalyst links on demand

  • The precomputed catalyst-to-note similarity tables are gone. Search scores
    catalyst links from the stored vectors when it needs them: the catalyst shown
    beside each catalyze result, --routing catalysts, --target search, and
    the desktop explore views. (4804ada)
  • catalyze returns the same notes in the same order. On a 4,700-note index
    the catalyst shown beside each result matched the old one 98.6% of the time;
    the rest differ because stored catalyst vectors are 8-bit.
  • The same index drops from 389 MB to 174 MB once compacted, and a refresh
    after editing 25 notes goes from 2.6 s to 0.19 s. Searches take about
    10–25 ms longer, because each loads the stored vectors.
  • --target folders now hold only their own notes and vectors, so preparing a
    target no longer depends on catalyst generation.

Compacting the index

  • enzyme compact rewrites the index file so freed space goes back to disk,
    and reports the size before and after (--json available). The old tables
    are dropped automatically on the first refresh after upgrading, but the file
    only shrinks when you compact it. (bc717cd)

Embeddings

  • Every note is embedded by default. The per-run cap (2,048 notes, adaptive up
    to 8,192) is gone, so a refresh catches up in one pass. An explicit
    prepare up to N documents / max_embedding_files = N still caps a run to
    the newest N. (20b267b)
  • Stored vectors are tagged with the model, width, and quantization that made
    them, so a future model change re-embeds instead of mixing vectors. The
    first refresh after upgrading re-embeds every note once, locally, with no
    model calls; questions are kept. (af3d7ad)
  • encoders 2 → 3: faster encoding, identical vectors for the model Enzyme uses.

Behavior changes to check

  • refresh JSON no longer has a similarity object or
    similarities_recomputed; it reports catalysts_embedded instead (normally
    0). The similarities_incrementally_maintained change reason is now
    catalyst_vectors_embedded.
  • ingest --quiet reports catalysts_embedded instead of similarity.
  • init --json-progress emits a catalyst_vectors stage instead of
    similarity.
  • The desktop explore warm-up (POST /explore/warm) reports chunks instead
    of similarities.
  • The first refresh after upgrading re-embeds the whole vault. That takes
    about a second for a few thousand notes.