Include timestamps of key moments from live stream

[c12i]

Currently, we are only storing coherent summaries of parliamentary/ senate proceedings transcribed from youtube. To make it easy for readers to navigate to specific timestamps in the live stream, having timestamps included alongside the summaries would be highly beneficial.

The logic to extract the vtt files from the youtube livestream is already available in the ytdlp_bindings crate

[johngitahi]

curious about how to contribute to this. will sending the vtt content along with the audio transcript work? i am thinking an updated prompt to map key moments to timestamps?

[c12i]

Hey @johngitahi, I was thinking of having the timestamp generation part of the core summarization pipeline.

will sending the vtt content along with the audio transcript work?

To answer your question, not really, the transcript tokens already are massive, to the point that they can exceed the token limit of the model in use (GPT-4o). I would prefer if they were handled separately (with the generated summary as context maybe), while also accounting for potential token limit exceeds, which would be handled dynamically, similar to how it's done with transcripts.

[johngitahi]

that means sending the summary along with the vtt contents, i think?

[c12i]

Yes, while paying attention to the token limit. Chunking of the WebVTT content might be required if the limit is exceeded. This is a rare case, but it's worth having this measure just in case.