video-ingest

Auto-generated from project docs

North Star

Booklet Status

BookletStatusAppetiteNotes
B1DONE3 hoursCore pipeline working. YouTube captions via cmd-aorus, summary via LiteLLM, corpus indexed.
B2PLANNED3 hoursMin Min /ingest command + MeTube web UI + media file watcher
B3PLANNED3 hoursRSS auto-ingest + playlist batch + speaker diarization
IDTypeDeliverablePriority
---------------------------------
B1-D1infraInstall yt-dlp + youtube-transcript-api + ffmpeg on VPSP0

Recent Decisions

DateDecisionRationale
---------------------------
2026-04-14Project kickstartedReplace Gemini manual workflow with automated pipeline
2026-04-14yt-dlp over reclip/MeTube/cobaltBest API for programmatic use, 1700+ sites (RESEARCH-247)
2026-04-14faster-whisper over Ollama WhisperFull control, VAD filter, word timestamps, int8 GPU (RESEARCH-247)
2026-04-14large-v3-turbo modelBest speed/accuracy for RTX 3050 8GB (~45s/20min, 5.5% WER)
2026-04-14Tiered transcription (captions → Whisper → Deepgram)Skip GPU when good captions exist
2026-04-14Bucket BCorpus feeds agent quality for revenue work
2026-04-14YouTube captions via cmd-aorus SSHVPS Hetzner IP blocked by YouTube bot detection. Residential IP works.
2026-04-14gpt-4.1-nano for summaries (not Haiku)Fast + cheap, good structured output quality
2026-04-14B1 DONEYouTube caption pipeline working end-to-end. Whisper path tested but non-YouTube blocked from VPS.

Source: /root/projects/video-ingest/