Table of Contents:

• To Do List Management System
• LLM with RAG system
• Online Courses Summary (In the future)
• Summary

This semester, it's a bit odd that I registered for 5 courses for my master degree at HCMUT along with the job at the bank, so It makes me burn out a bit.  I've only been taking the courses for a week and I'm already feeling miserable.

It's not until now that I've tried applying LLMs in work and study, but really, my tight schedule recently forces me to optimize everything and only prioritize things that brings the most important value

There is never enough time to do everything, but there is always enough time to do the most important thing.  - Brian Tracy.

To Do List Management System

I often use Microsoft To Do because it is simple and convenient.  Normally, just call "Hi Bixby" to create a task from my Samsung phone.  Tasks will be synchronized into the Calendar app.  However, it handles Vietnamese poorly, cannot add multiple tasks at the same time, does not automatically set the importance level... Although Google Assistant are better at handling speech to text problem, it has the same limitations as above.

I decided to go big this time and build my own Telegram chat bot.  Every time you send a voice, it will process it automatically:

• Breaking down the text into multiple tasks.
• Summary, fix tasks
• Set due date, reminder date, important level based on my personal information.

Very convenient.  But the biggest drawback is that I still need to send voice to the app manually, not "Hi Bixby" or "Hey Google", this makes it lose its practicality.  I will find a way to fix this in the future

To do list is integrated into this tele bot:

LLM with RAG system

I used Faiss as vector database and Langchain to get text from any sources I could think of (.pdf, .doc, .ppt, .txt, .md, .epub, link, ebook, youtube...).  If you ask Chatgpt, it will sometime mislead you or create fake answer. Although the answer might be correct, it doesn't follow the lecturer's documents.

This RAG system take the answer from our documents to produce results, so the results are more accurate and we can immediately check its correctness.  I also tested on the Software Testing Quizz, the result was about 7/10.  I also ask chatgpt, but the answer seems bad.

But Questions with the answer "All the answers are correct" are often wrong.  Perhaps it's because the model only finds the answer with the highest probability, so the aggregated answers are not fit to this case. Maybe I should rewriting the prompt.  Additionally, a small part of the learning resources are images, and I'm actually still looking for ways to optimize the explanation for the images.

My LLM + RAG system:

  nougat data/web_data/Growth_of_Functions.pdf --markdown --no-skipping -m 0.1.0-base -o data/nougat

Online Courses Summary (In the future)

During the pandemic, I realized that I'm best at  learning online courses.  I usually watch the video 2 or 3 times, fast forwarding unimportant parts (focusing on the 20% of the time that brings 80% of the value).  I plan to record all the lectures, then ask the model to synthesize and track which time period and what topic is being talked about... Or convert it to text format to save in the database for LLM + RAG system

Summary

I am convinced that individuals adept at utilizing LLMs can elevate their performance by up to 200%. To stand out among others, it requires a combination of diligent effort and the right tools. Even a marginal improvement of 1% can contribute significantly to your success.