Channels that are not yet indexed or imported can be done so by signing up, and imports are typically pretty fast for small to medium sized channels (~1 minute). Channels are the first-class citizen instead of individual videos, since summarization of individual videos is already a thing in YouTube.
---
I spend a lot of time watching and listening to long-form content on YouTube (livestreams, podcasts), which often contain nuggets of information or wisdom, but would have trouble finding where it was mentioned after closing the tab.
Was also bored with my full stack engineering day job, and wanted to scratch the itch of building something with AI, to hopefully make myself more industry-relevant.
---
What sets this apart from other tools is that I haven't really found anything similar, or they are very paywalled, or the experience is mostly a Chrome extension. There's Gemini which has access to a huge corpus of YouTube data, but I've often found the results to be inaccurate.
My goal here is to make the best possible feature-set on top of YouTube, but currently a little stuck, so looking for some feedback. For example, I recently added a mindmap widget to explore topics and come up with suggested questions, e.g. https://askchannel.ai/channels/@highyield
---
The tech:
A channel and all its transcripts are inserted into a PostgreSQL database, and also chunked and embedded into a vector database (Pinecone). Responses are generated with a LLM.
There's also API access, but that is a bit out of scope, and is currently quite basic.