RonanCodes/ingest-video

Ingest Local Video

Transcribe local video files and create wiki pages with keyframe screenshots.

Dependency: FFmpeg (lazy-install)

which ffmpeg >/dev/null 2>&1 || brew install ffmpeg

Step 1: Extract Audio

ffmpeg -i "$VIDEO_PATH" -vn -acodec pcm_s16le -ar 16000 -ac 1 /tmp/llm-wiki-audio.wav

Step 2: Transcribe (use first available provider)

Tier 1 — Local Whisper (free, no key):

# Try whisper-cpp first, then python whisper
whisper-cpp -m /usr/local/share/whisper-cpp/models/ggml-base.en.bin \
  -f /tmp/llm-wiki-audio.wav --output-format txt -of /tmp/llm-wiki-transcript
# OR: whisper /tmp/llm-wiki-audio.wav --model base.en --output_format txt --output_dir /tmp

Tier 2 — OpenAI Whisper API ($0.006/min, needs OPENAI_API_KEY in env or .env):

curl -s https://api.openai.com/v1/audio/transcriptions \
  -H "Authorization: Bearer $OPENAI_API_KEY" \
  -F file=@/tmp/llm-wiki-audio.wav -F model=whisper-1 -F response_format=text

Tier 3 — AssemblyAI ($0.01/min, has speaker diarization, needs ASSEMBLYAI_API_KEY): Upload file to /v2/upload, POST to /v2/transcript with speaker_labels: true, poll until status: completed. Returns utterances with speaker labels.

No provider? Tell user to install whisper-cpp (brew install whisper-cpp) or set an API key. Stop.

Step 3: Extract Keyframes

mkdir -p "$VAULT/raw/assets"
ffmpeg -i "$VIDEO_PATH" -vf "select=gt(scene\,0.3),scale=1280:-1" \
  -vsync vfr "$VAULT/raw/assets/${VIDEO_SLUG}-frame-%03d.png"

If fewer than 3 frames, fall back to I-frame extraction (select=eq(ptype\,I)). Cap at 20 keyframes.

Step 4: Save Raw Transcript

Save to $VAULT/raw/<video-slug>-transcript.md:

---
source-url: "file://<absolute-path>"
title: "<video-filename>"
date-fetched: <today>
source-type: video
transcription-provider: <whisper-cpp|openai|assemblyai>
has-diarization: <true|false>
keyframes: <count>
---

Include full transcript below frontmatter. With diarization, format as Speaker A: ....

Step 5: Create Wiki Source Note

Create $VAULT/wiki/sources/<video-title>.md with sections:

• Frontmatter: title, date-created/modified, page-type: source-note, domain, tags: [video, ...], sources, source-type: video, raw-file
• Overview — 2-3 sentence summary
• Attendees — speaker list if diarization available
• Key Decisions — extracted from transcript
• Action Items — tasks, follow-ups, deadlines
• Key Takeaways — 3-5 main points
• Screenshots — embedded keyframe images: ![Frame 1](../raw/assets/<slug>-frame-001.png)
• Full Transcript — wrapped in <details><summary> for collapsibility
• Sources — link to original file

Post-Extraction Report

Tell the user which provider was used, keyframe count, diarization availability, and content summary.

Dependencies

• FFmpeg — audio extraction + keyframes (lazy-installed)
• One of: whisper-cpp, whisper, OpenAI API key, or AssemblyAI API key