Problem
YouTube and audio ingestion (extractors.py's _extract_youtube/_extract_audio, via faster-whisper) currently flattens everything to plain transcript text before chunking. There's no way to jump back to "the part where they explained backprop" — citations point at a chunk index, not a timestamp.
Proposal
Preserve timestamp metadata through ingestion (faster-whisper and the YouTube transcript API both provide per-segment timestamps already) and store it alongside each chunk's payload in Qdrant. Citations for video/audio sources could then link to youtube.com/watch?v=...&t=123s or an equivalent seek point, instead of just naming the source.
Problem
YouTube and audio ingestion (
extractors.py's_extract_youtube/_extract_audio, viafaster-whisper) currently flattens everything to plain transcript text before chunking. There's no way to jump back to "the part where they explained backprop" — citations point at a chunk index, not a timestamp.Proposal
Preserve timestamp metadata through ingestion (
faster-whisperand the YouTube transcript API both provide per-segment timestamps already) and store it alongside each chunk's payload in Qdrant. Citations for video/audio sources could then link toyoutube.com/watch?v=...&t=123sor an equivalent seek point, instead of just naming the source.