Repository navigation
Highlights
Two commands that duplicated others are gone, and every hsearch hit now says how it was found, so loosely related results are easy to spot.
Breaking Changes
bkmr sem-searchis removed. Usebkmr hsearch --mode semantic QUERY. It runs the same vector search and adds whatsem-searchlacked: tag filters,--json,--fzfand--stdout.bkmr clear-embeddingsis removed. Usebkmr backfill --clear. It still asks first (exit 130 if declined) and needs no embedding model. It cannot be combined with--force,--pruneor--dry-run.- Piped
hsearchoutput changes shape. It now always has six tab-separated columns:id,title,url,rrf_score,similarityandkeyword_match. The last two are empty where they do not apply. Intitleandurl, backslash, tab, newline and carriage return are written as\\,\t,\nand\r, so multi-line snippet and markdown content stays on one line. Scripts that read the url column as raw content must unescape it.
New
hsearch --mode semanticreturns only the bookmarks closest in meaning, ranked by similarity.- Every hit reports
similarity(cosine similarity to the query, 1 = same meaning) andkeyword_match(whether the keyword search found it). Both are new--jsonkeys, left out where they do not apply, and the interactive list shows them as[kw]andsim 0.812. Semantic search has no cutoff, so a hit withkeyword_match: falseand a low similarity is only loosely related.
Behaviour changes
- After switching the embedding model,
hsearchfails with a message that names the fix,bkmr backfill --force, instead of sqlite-vec's raw dimension error. hsearch --mode semanticwithout stored embeddings suggests runningbkmr backfill.- The
hsearchopen prompt rejects IDs that are not numbers or not among the results, and opens nothing until every ID is valid. Before, such IDs were skipped silently. backfill --clearresets content hashes even when no vectors are stored, so the nextbackfillembeds everything again.
What's Changed
Full Changelog: v8.8.0...v8.9.0