Index / Development · 2026-07-07 dev ai automation

zvec-search-setup

Daftar isi

Summary

Mesin retrieval vault: chunk markdown → embedding Cloudflare Workers AI @cf/baai/bge-m3 (dim 1024, multilingual) → zvec (vector DB in-process, db/zvec-notes) + metadata db/chunks.json.

Detail

  • python scripts/brain.py index — scan notes/ library/ social/ daily/; hanya file berubah (mtime) yang di-embed ulang; chunk per section H2, khusus library/ per blockquote (kind=quote + nomor halaman).
  • python scripts/brain.py query "teks" [-k 8] [--folder library] [--kind quote].
  • python scripts/brain.py cite "teks" — khusus kutipan buku.
  • zvec 0.5.1: wheel Windows py3.10–3.14, HNSW cosine. Kalau zvec absen → fallback cosine murni dari chunks.json (hasil sama, lebih lambat).
  • [2026-07-07] Update: di laptop ini DLL zvec diblokir Windows Smart App Control → fallback cosine yang aktif. Detail & opsi: docs/SETUP.md §B.
  • Kredensial: .env root (CF_ACCOUNT_ID, CF_API_TOKEN scope Workers AI Read). Free tier 10k neurons/hari — cukup banget karena delta-only.
  • Upgrade opsional: FTS + MultiQuery hybrid zvec (v0.5+) untuk keyword+semantic sekali jalan.

Key Takeaways

  • Tanpa VPS, tanpa torch lokal — embedding numpang free tier CF.
  • db/ gitignored; index bisa dibangun ulang kapan pun dari markdown.
  • Format kutipan di book-format-rules menentukan mutu /cite — blockquote + — hal. N wajib.
Terhubung dari
book-format-rules · Index — Peta Vault · second-brain-workflow