bytelyst/learning_ai_notes
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity
a697752d15
learning_ai_notes/docs/roadmaps/ADR-2026-03-31-hybrid-search-without-embeddings.md
Saravana Achu Mac a697752d15 feat: implement WEB_AI_FAST_ROADMAP (web + backend + docs) 
Phase 1: Command palette (⌘K), editor autosave with quiet auto-saves, dashboard
saved views from API + quick links + onboarding seed CTA, explicit task scan panel.

Phase 2: Context pack formatter with YAML frontmatter, copy on note + workspace .md export.

Phase 3: ADR for hybrid search without embeddings; POST /notes/search (lexical +
ranked hybrid); search UI mode toggle.

Phase 4: POST copilot + suggest-title; in-editor copilot actions; /chat retrieval
answers with citations (backend chat.rag_enabled).

Phase 5: Settings MCP snippet, offline queue note, API token deferral; DEEP_LINKS.md.

Phase 6: Note shares + public GET; share page; POST onboarding-seed.

Phase 7: note_versions on PATCH; version panel; create-note templates; PWA manifest.

Flags: search.hybrid_enabled, copilot.enabled, chat.rag_enabled, onboarding.seed_enabled.
Made-with: Cursor
2026-03-31 13:00:36 -07:00

1.4 KiB
Raw Blame History

ADR 2026-03-31 — Hybrid note search without vector embeddings (NoteLett)

Status

Accepted — implemented in NoteLett backend POST /notes/search.

Context

Full semantic search requires embedding models, batch indexing, and higher infra cost. NoteLett needed ranked search with explainability (match kind: title, body, tag) for the AI-fast roadmap without blocking on a shared RAG platform.

Decision

  1. Phase A (shipped): Implement lexical + in-process re-ranking on decrypted note text:

    • Load candidate notes via existing datastore query (search substring filter when query non-empty; recent notes when empty).
    • Score with token overlap: title > tags > body; multi-word queries use AND semantics.
    • Return score, matchKind, and snippet for each hit.

  2. Phase B (future): Optional integration with a shared embedding / RAG service or Cosmos vector index when product and cost model allow. The ADR in this doc should be updated to reference the chosen provider and index update path (on note write vs batch).

Consequences

  • Pros: No new services; works with current Fastify + datastore; explainable results.
  • Cons: Not true semantic similarity; large workspaces may need stricter candidate limits or Phase B.

Feature flags

  • search.hybrid_enabled — when false, POST /notes/search lexical mode only (substring list).