Skip to content

PRD.md

Latest commit

 

History

History
112 lines (92 loc) · 5.6 KB

PRD.md

File metadata and controls

112 lines (92 loc) · 5.6 KB

PRD — Office Recorder

1) Summary

Build a local, always-on recording system for a shared office room that captures full-day audio, splits it into manageable segments, transcribes locally, batches conversations, and delivers a daily summary of topics, decisions, and action items. Control must be available via a local web UI and via OpenClaw text commands ("start recording", "stop recording").

Recommendation choices

  • Platform: Local web app on the Mac Studio (fastest build + easy access in any browser).
  • Transcription: On-prem only (no cloud). Use faster-whisper (local) with optional upgrade to whisper.cpp/Metal.
  • OpenClaw control: Skill + Exec tool calling local control script; optional webhook delivery for summaries.
  • Start/Stop: Manual one-click start + optional schedule (not in v1, design-ready).
  • Speaker handling: Start without diarization; add diarization (whisperX + pyannote) as Phase 2.
  • Retention: Keep raw audio 30–90 days, transcripts/summaries longer; all local with configurable pruning.
  • Compute split: Record on Mac Studio A (attached to mic interface), transcribe/summarize on Mac Studio B after hours to avoid daytime CPU load.

2) Goals

  • Capture all in-room conversations during work hours with minimal friction.
  • Provide reliable transcription and daily summaries without sending data off-prem.
  • Enable quick control via UI + OpenClaw text commands.
  • Keep the system stable for full-day sessions (8–12 hours).

3) Non-Goals

  • Real-time transcription/summary during the day (out of scope for v1).
  • Full speaker identification with names (out of scope for v1; requires diarization + enrollment).
  • Cloud storage or cloud LLM use (explicitly avoided in v1).

4) Personas

  • Founders / partners: Need a daily digest of discussions, decisions, and action items.
  • Visitors: Must be informed via consent/signage; expect recordings to be local and private.

5) User Stories

  1. As a founder, I click “Start Recording” and the room records all day without mic fiddling.
  2. As a founder, I can text OpenClaw “start recording” and it starts immediately.
  3. As a founder, I get a daily summary with topics, decisions, and action items.
  4. As a founder, I can review the raw transcript per day if needed.

6) Functional Requirements

6.1 Recording

  • Start/stop recording from UI and from OpenClaw command.
  • Continuous recording for full day without file size issues (segment every N minutes).
  • Store audio as WAV (16 kHz mono) or FLAC.
  • Save session metadata (start time, device, format, sample rate).

6.2 Transcription

  • Batch transcription for a given day.
  • Use local whisper model (faster-whisper).
  • Store transcript per segment and a merged transcript.

6.3 Conversation Batching

  • Group transcript segments into conversation blocks.
  • Default rules: new block when silence gap > 7 minutes or block length exceeds max words.

6.4 Summaries

  • Produce JSON summary and Markdown summary per day.
  • Daily summary includes: overview, top topics, decisions, action items (with owner/due if mentioned), questions, risks, follow-ups.
  • Optional delivery via OpenClaw webhook.

6.5 OpenClaw Integration

  • Provide a custom skill file to map “start recording / stop recording / status / summarize” to local API calls.
  • Support sending daily summary via OpenClaw webhook (if configured).

6.6 Admin/Config

  • Environment variables for mic input, sample rate, segmentation, LLM endpoint, and retention.
  • Optional retention job (Phase 2) to prune raw audio after X days.

7) Non-Functional Requirements

  • Local-only data: No external upload or cloud processing.
  • Reliability: Must tolerate a full day of continuous capture.
  • Performance: Recording must be lightweight; transcription can be scheduled after hours.
  • Security: Data stored locally; optional disk encryption recommended.

8) System Architecture

8.1 Components

  • Recorder Service: ffmpeg-based audio capture with segmenting.
  • API Server: FastAPI for control and pipeline orchestration.
  • Transcriber: faster-whisper (local) producing segment JSON.
  • Summarizer: Local LLM via OpenAI-compatible API (Ollama/LM Studio).
  • UI: Local web dashboard (Start/Stop, pipeline buttons, summary view).
  • OpenClaw Skill: Maps text commands to local control script.

8.2 Data Flow

  1. Start recording ? ffmpeg segments audio into /OfficeRecorder/YYYY-MM-DD/audio.
  2. Transcribe ? JSON per segment in /transcripts.
  3. Batch ? conversation blocks based on time gaps.
  4. Summarize ? daily JSON + Markdown in /summaries.
  5. Optional OpenClaw webhook ? summary overview text.

9) Privacy & Consent

  • All participants informed of recording with signage.
  • Ensure clear internal policy for visitors.
  • Local data only; define retention (e.g., 90 days audio, 1 year transcript).

10) Metrics

  • Recording uptime (hours recorded/day).
  • Transcription completion time.
  • Summary delivery success rate.
  • “Missed decision” feedback from founders.

11) Risks & Mitigations

  • Mic misconfigured: Provide device listing and test recording flow.
  • Long audio files: Segment files every 5 minutes.
  • Model drift: Allow model updates and prompt tuning.
  • No diarization: Add Phase 2 diarization module.

12) Phase Plan

  • Phase 1 (MVP): Recording + transcription + summary + OpenClaw commands + UI.
  • Phase 2: Diarization + auto-schedule + retention pruning + analytics.

13) Acceptance Criteria

  • Start/stop recording from UI and OpenClaw text.
  • Successfully record a full business day without failure.
  • Generate a daily summary with topics and action items.
  • All artifacts stored locally.