Fixed

  • Memories uploaded before search by meaning are embedded again (ADR-102). The backfill had stopped at 207 of about 8,900 entries in production: full-length entries put one embedding request over the model's 60,000-token limit, and the same request failed on every retry. Requests are now packed under that limit whatever the text contains, a request that still fails is split down to single entries so one entry can never hold up the rest (an entry that cannot be embedded is given up until its content changes, and counted), and a member whose embedding fails as a whole backs off instead of making four failed model calls per upload. Search itself is unchanged.

Update HarnessLink and every installed harness to its current release:

hnl update

New machine? Install with:

curl -fsSL https://harnesslink.sh/install.sh | bash

Newer: 17.21.3 · Older: 17.21.1 · All releases · Atom feed

Flag notifications