Skip to content

1.14.0

Choose a tag to compare

@notoriouslab notoriouslab released this 21 Oct 03:32
· 31 commits to main since this release

Changed

  • Descriptions cover the whole note (#17). AI curation used to read only the first 1,200 and last 800 characters of a note, so points in the middle of a long note were missed and the model filled the gap by guessing. It now reads six evenly spaced passages from start to end, within a fixed budget, with links and inline images left out. In blind side-by-side reviews of 18 descriptions of long notes, the new ones covered about 30% more of each note's main points. With Ollama, description requests now ask for a 6,144-token context so the server never quietly cuts the note short, and wait up to 120 seconds for large models.
  • Descriptions end at a full sentence. English descriptions were cut at 500 characters, usually in the middle of a word. Descriptions now stay within about 500 characters (about 250 Chinese characters) and end at the last complete sentence.
  • English and Chinese prompts now ask for the same thing: 2-3 complete sentences that open with what the note is about and cover the whole note.
  • Canvas labels drop more stock openings such as "This note details…" or 「這份筆記探討了…」 for notes without tags.
  • The model hint is more honest. Settings → AI Curation now suggests models of 4B or larger (for example gemma3:4b) and says that small local models occasionally get a detail wrong, such as who did what or why. Because the new descriptions say more, they also contain a few more of these slips; per point covered the rate stayed about the same, and only a much larger model (27B) made clearly fewer mistakes. The default model is unchanged.

Fixed

  • Descriptions cut off by earlier versions can be repaired. "Generate descriptions for current results" now also rewrites descriptions an earlier version cut off mid-sentence, and a notice says how many were repaired. Other existing descriptions are still left alone, and if you edit a description while a batch is running, your edit is kept.
  • A reply cut off midway is no longer saved as raw JSON in a note's description.