This week, I worked on the “book library” MVP. My goal is for the MVP to mimic an entrepreneur who’s an avid reader with a photographic memory. I want to query across multiple books, and I want the MVP to make connections that uncover new insights. I also want it to recall any book’s details quickly.

I’m using retrieval augmented generation (RAG) and Gemini LLMs in the MVP. Last week’s testing yielded responses that weren’t detailed enough and didn’t uncover new insights. I was happy to have something working (it felt like a big milestone), but it still needed work. This week I tried using a different RAG setup—a separate agent and RAG to index each book instead of one agent and RAG to index multiple books. I also tested different prompting and system instructions. The changes didn’t improve the responses, which was frustrating. Still too high-level and unable to make insightful connections.

I’m not sure why this is happening. My developer friend and I have a few theories. It could be a limitation of RAG not being great at indexing entire books. It could be limitations with Gemini LLMs, technical limitations imposed by Google Cloud Platform (GCP), or something else. Given that the output I’ve been able to generate from Google AI Studio for an individual book has been pretty detailed, we think there’s a high probability it could be a GCP limitation.

This wasn’t the outcome I was hoping for at the beginning of the week, but that’s part of the process when you’re building something that hasn’t been done before. Definitely frustrating, but such is life. We’ll do more testing to try to figure this out.