Answers you can check
An answer you can verify in one click is worth more than a better sounding answer you cannot. Every citation is a real second in the video, and the player seeks there when you click it.
Learn a whole courseby asking it questions.
Point it at a lecture series. It watches every video, slides and code and diagrams included, then answers from anywhere in the course and shows you the exact moment it read.
Open the appBring your own key. We spend nothing, and neither does anyone else on your behalf.
?
lecture 1readylecture 24
- 1readthe video is transcribed with timestamps
- 2chunksplit into passages that keep their seconds
- 3embedvectors computed in your browser, cached
- 4retrievethe few passages that answer the question
- 5answerstreamed, citing the moments it used
Cost stops tracking length
Sending whole videos with every question means every question pays for the entire runtime again. Retrieval sends the handful of moments that actually answer it, so a thirty hour course costs no more per question than a single lecture.
That is the entire architectural argument, and it is why an entire lecture series is as practical to interrogate as a single clip.
Input tokens per question
one question, against a 24 lecture course
Retrieval is 117x smaller than sending every transcript, and the top bar does not fit in a context window at all. Reading the course costs once; every question after that is flat, however many lectures you add.
Your key, your browser
Keys are typed into your browser and sent only to the provider you picked. There are no server routes in this app that could receive one, which is a stronger guarantee than a promise not to look.
Read once, ask forever
Reading a long video is the expensive step, so the transcript, the embeddings and your answers are all kept in your browser. Reopening a video you have already read costs no provider requests at all.
Follow ups actually work
“And the second one?” carries almost no signal on its own. Each question is rewritten against the conversation before retrieval runs, so a follow up finds the right passages instead of noise.
Pick your own models
Nothing is hardcoded. The app asks your key which models it can use and lets you choose, so a model being retired is a dropdown away from fixed rather than a rebuild.
Honest about limits
Free tiers are counted in requests per minute, and long videos hit them. When a part fails you keep everything that worked, see exactly what the provider said, and retry only the parts that are missing.
Why this exists
I built this for my own knowledge base. I watch a lot of lectures and conference talks, and the useful part is never findable later: I remember that someone explained a thing well, but not which video, and certainly not where in it. Scrubbing through a twenty hour course to find one four minute stretch is not learning, it is archaeology.
So it reads the whole course once, keeps the transcript in my browser, and lets me ask it things. The citation matters more than the answer: I want to be taken to the moment and judge it myself, not trust a paraphrase.
It turned out useful enough that I am sharing it. It runs entirely on your own API key with nothing stored on a server, so using it costs me nothing and costs you only whatever your provider charges. Point it at a course you have been meaning to get through.
Take it for a run
You need a lecture or playlist URL and a key from one provider. Reading a course is a one time cost; after that it is instant and free, forever.
Open the app