You are tackling an interesting topic which I hit every day too. Especially if I read the next “RAG is dead”.
The main problem we have right now when working with knowledge (especially in enterprise), you don’t want to send the document every time to the llm to answer one question.
That’s just horrible inefficient.
There must be a way that keeps the document close to the llm (like your disk idea), so I can send several questions including chat history.
You mentioned curse how they do it. Do you have any interesting links?
Cheers Daniel. I don't think Cursor's approach is public, but you can see the tool reading "lines 1-99", and this is my speculation on how they do it. I'll aim to clarify that in the next vid. Thanks.
You are tackling an interesting topic which I hit every day too. Especially if I read the next “RAG is dead”.
The main problem we have right now when working with knowledge (especially in enterprise), you don’t want to send the document every time to the llm to answer one question.
That’s just horrible inefficient.
There must be a way that keeps the document close to the llm (like your disk idea), so I can send several questions including chat history.
You mentioned curse how they do it. Do you have any interesting links?
Cheers Daniel. I don't think Cursor's approach is public, but you can see the tool reading "lines 1-99", and this is my speculation on how they do it. I'll aim to clarify that in the next vid. Thanks.
Do you plan to share your code somewhere? Happy to work with you on the topic. It is an super interesting one.
I also have another one which I can share with you in common, which I think is not tackled yet, but could be a game changer 😁
Howdy , code is a paid membership. Is on Trelis.com . And Trelis has a colab program! See Trelis.com !