Hacker News Reader

Comment by chermi | Hacker News Reader

Looks very cool! How does it compare to notebooklm?

One feature that may be useful for others that I've wanted is constraining the info the the LLM uses from a given document. I'm self-studying from a math text with pdf I have loaded. Scenario: I'm working on a problem and I want to know if there's any useful theorems or definitions related to the problem, but only from the text up to the problem. I don't want it to give me results from anything in the future of the text, since presumably the author intended the reader to solve the problem using material only up to that point.

You can generalize this, but probably out of scope. Sometimes it's instructive to try to solve a math or physics problem using a restricted set of theorems or ideas. So it would be really cool if you could ask the LLM what results you have at your disposal excepting theorems x,y,z and all results deriving from those.

View on HN

NotebookLM is in my mind an introductory product for deep understanding which does RAG and long podcasts well. However, strong understanding requires much more. We are still building out a coherent experience, but key ingredients are a strong exploration artifact (our map aims to be that), rich source control, context engineering (of the nature you pick up on above), and switching between modalities. We also have a powerful reader which lets us switch to various levels of summaries/original source on a chapter level, and jump between these easily. There's also friction in Notebooklm like you have to find web sources manually; our agent aims to do that from a simple text request. Going forward, we aim to add the ability to have a strong note taking experience (you can take hierarchical notes, AI assisted), and the agent should nudge you to complete what you are trying to achieve in a space.

In the chat agent, there's an option to turn off AI knowledge; and we are adding chapter level context control. I think 'start until upto this point' (and 'from this point to end of text') are great additions. Will add.

Regarding the generalization, I think with strong context engineering and powerful reasoning, this is hopefully achievable. But making a bespoke feature for this feels hard.

Thanks for your feedback!

Cursor and the agenetic IDEs are great for this kind of context control. Although designed for code, I’ve started using them for regular language, by converting large docs to broken-up plain text files (which it can do for you). Then you can @ whatever chapter you like in your questions

I was doing this for epubs in cursor too, which is what led me to building this product! Cursor is just not engineered to support the above across native file formats (pdfs/epubs).

Reply on news.ycombinator.com