Skip to content

enzyme v0.13.3

Choose a tag to compare

@jshph jshph released this 11 Nov 05:52
· 5 commits to main since this release

Enzyme release notes - v0.13.3

Short version

Large topics get more questions: each topic now earns about one question per
five notes, up to 80, on top of the old share of 15. Upgrades rebuild the
search index without throwing away your questions, and Enzyme tells you when it
rebuilds or regenerates. The embedding backend moved to the encoders crate,
which encodes about 3.5× faster with the same vectors. Agent guidance now
starts searches from the user's words and topic names.

What shipped

Question budgets

  • Each topic gets the larger of its share of 15 (scaled by note count against
    the largest topics, as before) and one question per five notes, capped at
    80. Small vaults keep the allocation they had. On a vault of a few thousand
    notes, total questions grow about 3–4×. (cadf315)
  • question budget N (or min_top_catalysts = N) now sets both the largest
    topics' budget and the cap for every topic. A vault that already sets it
    gets the same allocation as before.
  • Because catalyst logic changed, each vault's next refresh regenerates its
    questions in the background and says so once.

Upgrades

  • When the index schema changes after an update, Enzyme rebuilds the search
    index in place: notes are re-indexed and re-embedded locally with no model
    calls, and questions, epochs, and settings are kept. Before, the agent path
    deleted the whole database and regenerated every question. init --force
    still starts from scratch. (a32b759)
  • Questions are regenerated only when the way Enzyme generates them changes.
  • Enzyme tells you when either happens, in one sentence: refresh --quiet
    returns it once as update.note, interactive refresh and init print it,
    and status shows it for a week as update.notice.
  • After a clean init, the next refresh no longer regenerates every topic a
    second time.

Embeddings

  • The vendored embedding code is replaced by the encoders crate. Vectors are
    unchanged, so existing indexes are not re-embedded; encoding is about 3.5×
    faster. (#21)

Agent guidance

  • The runtime skill tells agents that petri's tag, folder, and people names
    are good search terms. (dd67b21)
  • The Hermes first-session instructions no longer tell agents to build
    searches from catalysts or say search matches "by concept, not keyword".
    They now start from the user's words and topic names, borrow catalyst
    wording only when there is room for one or two searches, and point to grep
    for every occurrence of literal text.
  • The Hermes petri tool documents its real default of 15 topics.

Behavior changes to check

  • The first refresh after upgrading regenerates questions for every topic in
    the background. On hosted generation that costs model calls once per vault.
  • Vaults without a question budget get more questions on large topics, so
    later regenerations take longer and, on hosted generation, cost more.
    Set question budget N to cap it.