Deep Dive
RAG · LLMLike NotebookLM, but pointed at my own content: the podcasts I follow and any YouTube video I throw at it. It reads everything, pulls it into a single summary, then lets me ask questions to dig in.
It runs on the home server, so there's no public instance, but I'm happy to give you a tour.
What it does
I get through a lot of podcasts and the odd long YouTube video, and the good bits never stick the first time round. This pulls a week's worth of that content into one place, writes one analysis across all of it, and then lets me sit and talk to it (“what was the argument about X?”, “how does this connect to last week?”) until I actually understand the thing.
Because it only ever draws on the sources I've picked, the answers stay tied to those instead of drifting into whatever the model already knows.
How it works
Each source, whether a podcast feed or a YouTube link, gets transcribed and chunked into a vector store. When I start a session, the relevant passages are retrieved and a language model writes that summary, then answers follow-ups using only those sources.
Sessions and bookmarks are saved, so I can come back to a topic later and keep pulling on the thread.