Skip to content

ADR-v2-062 — On-Device Condense

Status: Accepted (2026-07-02) · Wave H Context links: [[adr-v2-019-meeting-scribe]] (others' meeting vs your own burst), [[adr-013-llm-cleanup]] (abstractive tier), [[adr-011]]

Context

Wave H research (#5) — speak a rambling paragraph, and a "condense" hotkey variant inserts a tightened summary/outline instead of the verbatim transcript. Anchors: NVIDIA Canary-Qwen-2.5B (on-device transcribe+summarize), Omi 0.6B on-device transcribe+summarize.

Distinct from Ambient Meeting Scribe (captures others' multi-speaker meetings) — this condenses your own single dictation burst on demand.

Decision

Add an opt-in On-Device Condense: [condense] enabled=false, max_sentences=2. The pure core condense(text, max_sentences) performs extractive summarization — split into sentences, score each by normalized content-word frequency, and emit the top-N in original order (dependency-free, deterministic). The abstractive small-LLM (Qwen3-0.6B) is deferred behind the existing llm_cleanup extra. OFF by default.

Consequences

  • Ships with no new dependency — frequency-ranked extractive summary.
  • Distinct from Meeting Scribe (own burst vs others' meeting).
  • Privacy (ADR-011): reuses the dormant offline LLM path when abstractive; nothing leaves the machine.
  • Caveat: extractive summaries are blunt → abstractive tier is the quality path; off by default.