Блог пользователя KenZuuuu

Автор KenZuuuu, история, 5 месяцев назад, По-английски

Hi everyone, I'm looking for some advice on how to build or configure a local AI system to act as an automated personal tutor. Here is the workflow I’m trying to achieve: Daily Input: Every day, I will feed the AI my study notes, which include specific math/science formulas, question types, and core theories (Physics, Chemistry, Math, etc.). Idea Generation & Quizzes: I want the AI to analyze these notes, understand the underlying patterns, and create brand-new practice questions or theoretical quizzes based on what I just learned.

The "Big Exam" Compilation: Over time, as my knowledge database grows, I want to be able to prompt the AI to generate a complete mock exam file (like a text or Markdown file) that pulls exactly one question from each topic/type I have ever submitted.

I am running a local setup on my laptop (Lenovo LOQ with an NVIDIA RTX GPU), using open-source models like Qwen 2.5 via LM Studio / AnythingLLM, so everything needs to be 100% offline.

Has anyone built a similar system? Do I need to write a custom Python script using LangChain/LlamaIndex, or is there an existing tool/RAG setup that can handle this systematic accumulation and test generation?

Thanks a lot for your help!

  • Проголосовать: нравится
  • -21
  • Проголосовать: не нравится

»
5 месяцев назад, скрыть # |
← Rev. 3  
Проголосовать: нравится 0 Проголосовать: не нравится

Maybe you do not need to get an offline model, which can be very annoying when you try to build and configure them. I think some subjects on Github can meet your need. They are skills for AI agents. All you need is just a platform for vibe coding, and AI agents will do the rest.

I have not tried an offline model, but I am using agent skills from Github to help with my training these times. So I advise you to try this way. Just get on Github and find some subjects that suit you.

»
5 месяцев назад, скрыть # |
 
Проголосовать: нравится 0 Проголосовать: не нравится

Try ollama

»
5 месяцев назад, скрыть # |
 
Проголосовать: нравится 0 Проголосовать: не нравится

"Nvidia RTX" isn't entirely useful. It can range from being able to only run a 7b model at q4 or larger 30b models.

Also for studying I really do not suggest using offline models. Since you probably want up to date information, you need web search. Web Search either requires a API, which is expensive, or self-hosted like searxng which might not be the most accurate. You can try opencode which is primarily for coding but there is (I believe) free web-search and other useful things like context compaction. This technically works but again offline models that can run on Nvidia RTX won't be too great. If you really need offline, I recommend Qwen3.6 35B-A3B as its pretty fast and decently smart too. llama.cpp works fine here.

Seriously just use NotebookLM or another AI and save all you learned to a file and upload it to its context every time. It's much easier and better.