Is vector store right for my use case?

Hello everyone,
I’m creating an assistant which helps me extract questions from a textbook for a given topic.
I’m a bit confused which would be the better approach:

  1. I create a vector store with my textbooks (I’ve three of them, 10 MB each). Then I ask the assistant to get the questions for a specific topic after reading the vector store.

  2. I pass the text from the index of each of the textbooks and then ask chatgpt to return the page numbers for the lessons which are relevant to the given topic.
    Then I extract the text from the given pages and pass into the prompt asking chatgpt to extract the questions.
    (Since each lesson is a different topic, and none are more than 30 pages long, I don’t see context window being an issue with this approach).

I’m kind of leaning towards 2 because, the assistant doesn’t show what chunks are retrieved from the vector store and I fear chatGPT will add it’s own twist to the questions before returning them.

I’m new to assistants and would greatly appreciate everyone’s feedback.

Thanks for your time!

1 Like

Hi @the.brainiac ,
before going into RAG and vector storage, I’d suggest you trying this easy and simple approach - Routing by semantic similarity. Basically, you can use first a prompt that would classify to which of your topics each user query is relevant and would route it to a variable (system prompt) that would contain that topic. The size of 10Mb doesn’t say much - depending on if it’s a .txt or .PDF it can have different amount of words (and tokens). If your topics do not exceed 100’000 tokens - just use this simple approach and the GPT-4 Turbo 128K model, putting your textbooks directly into the system prompt. Unless you have thousands of queries per day, it would’t have any significant cost impact.

1 Like