Distill Knowledge

Convert Recording → Knowledge Markdown

Emit outbox/{meeting-slug}/transcript.md; screenshots inline when useful; structured docs only on request. {meeting-slug} = kebab-case-topic-YYYYMMDD. Never touch inbox/ or knowledge/.

References (load on demand): setup · output templates · ffmpeg · transcribe CLI · prep audio CLI · chunked transcription · structured docs · cleanup CLI · spot-check: scripts/extract_clip.py.

Scripts run via uv run --script (PEP 723). All support --help. On first run verify setup prerequisites.

Workflow

Step 0 — Intake (mandatory, before any API call)

Ask verbatim:

"Before I process this, three quick things:

1. Language of the conversation? (e.g. en, ru, fr — used as --language hint)
2. How many speakers?
3. Topic / domain in one line, plus any proper names or specialized terms."

Record answers; sensible defaults if user skips a field. Never silently transcribe without asking.

Technical / multilingual / mumbled audio → VTT-aligned path strongly preferred (if VTT available). Without VTT, diarize fallback with 8-min chunks; warn user about quality.

Steps 1–6

1. Inventory + probe — ls inbox, ffprobe media. VTT present → parse via parse_vtt.py → tmp/prep/<slug>/vtt_cues.json; sample cues, assess quality (speaker count, garble, gaps, proper-noun fidelity). Non-English VTTs: screen-reference detection won't fire — read cues directly.

2. Gate 1 — present findings + plan + {meeting-slug}. Determine transcription path:

   Scenario	Path
   VTT good quality	Render VTT directly; skip prep + API
   VTT exists, text garbled	VTT-aligned retranscription: VTT as skeleton (speakers + timestamps) + gpt-4o-transcribe for text quality → agent aligns
   No VTT, single speaker	gpt-4o-transcribe directly
   No VTT, multi-speaker	Diarize fallback: gpt-4o-transcribe-diarize at 8-min chunks (known unstable — warn user)

3. Preprocess + transcribe — run prep_audio.py on input (audio or video; extracts audio in-pass; source video retained for screenshots).

   VTT-aligned path (primary for retranscription):

   • VTT provides speaker labels + turn timestamps; transcription provides clean text
   • Run gpt-4o-transcribe on stripped.ogg (or per-chunk if >8 min) with --prompt (vocab + 1-line topic only)
   • Agent aligns clean text to VTT turns — this is language work you perform directly. Match transcribed text to VTT turn boundaries using VTT text as positional guide. Preserve VTT speaker labels and timestamps. See VTT-aligned merge.

   Single-speaker path:

   • Run gpt-4o-transcribe on stripped.ogg (or per-chunk if >8 min) with --prompt
   • Output is the transcript directly; no alignment needed

   Diarize fallback (no VTT, multi-speaker):

   • Chunks at 8 min max (diarize model unstable on longer audio)
   • Per-chunk: transcribe_diarize.py --manifest --chunk-index N; then merge_chunks.py
   • Warn user: diarization quality is unreliable; may need manual correction
   • See chunked transcription

   All paths: --language required. On non-zero exit → surface stderr Error [<category>]:, ask wait/cancel. See exit codes.

4. Speaker labelling —

   • VTT-aligned: speakers come from VTT; confirm with user (VTT labels may be generic like "Speaker 1").
   • Diarize fallback: render_transcript.py --samples <json> → user names speakers.
   • Single-speaker: user provides name or default.
   • Then render via render_transcript.py --speakers ... --out outbox/{slug}/transcript.md.
   • For diarize long path: cleanup pass as language work on merged.json → polished.json. See Cleanup Pass.

5. Transcript (mandatory artifact) — produced by step 4. For exact markdown shape see output templates § transcript.md. Faithful to meaning; repair recoverable garble; never invent. See Fidelity rule.

6. Screenshots — skip if no screen content (faces only → zero screenshots). Take frame at timestamp + 2s, -q:v 2. Source video from manifest.json source field. For format see output templates § screenshots.

   • UI / slides / docs → screenshots in scope
   • Diagram on screen → screenshot and Mermaid
   • Data table → screenshot and markdown table

Steps 7–10 — Structured docs (conditional)

Gate 2 — ask: structured docs or transcript only? If transcript only → cleanup + report + stop.

Otherwise → structured docs reference: plan topics, emit summary.md + topics/{slug}.md, cleanup, report, stop.

Step 11 — Cleanup (mandatory, after user accepts)

Once the user confirms the results are acceptable, remove temporary prep artifacts via cleanup.py:

uv run --script scripts/cleanup.py --slug {meeting-slug}

The script verifies outbox/{meeting-slug}/transcript.md exists before deleting tmp/prep/{meeting-slug}/. Use --dry-run to preview. Never use rm directly — all temp removal goes through this script.

Fidelity rule

Transcript captures what was said and meant, not the literal sound stream. In priority order:

1. Preserve every substantive turn — decisions, claims, questions, objections, reactions. If spoken, it appears.
2. Repair recoverable garble — when ASR returns gibberish and intent is recoverable from parallel clean pass or context, replace with faithful version. Selecting better evidence ≠ paraphrasing.
3. Drop fabrications — sentences introducing concepts absent from any source signal (prompted model hallucinations during silence).
4. Never invent — don't fill silences with plausible speech.

When two transcripts agree on meaning → faithful. When they disagree and neither is recoverable → mark unclear.

Anti-patterns

• Don't fabricate. Inventing dialog is the worst failure mode — worse than garble. See fidelity rule.
• Don't write language-processing scripts. No difflib, regex tables, word-overlap heuristics, jq/sed/awk for cleanup. The cleanup pass is language work you perform directly. Scripts = I/O plumbing only.
• Don't load rich --prompt on non-diarize model. Prompts leak as fabrications. Vocab list + 1-line topic max. See prompt-hallucination warning.
• Don't auto-pick VTT decisions. Always surface assessment; user confirms at Gate 1.
• Don't write outside outbox/{meeting-slug}/. Temp files → tmp/; finals → outbox/.
• Don't run rm on temp files. Always use cleanup.py — it validates output exists and only operates under tmp/.