I'm back to working on Papersy. After working hard on BeaverHack, now coming back to this project, I basically have to relearn how my program even works and its internal. I have to read through my log and try a few times with some papers to remember where I left off.
Some problems persists:
- It hasn't fetched the name accurately, sometimes it works, other times not. I suspect due to it feeding on the text form of PDF instead of image of PDF. It saves on token, but all the contents are treated indiscriminately by the LLM
- There was a task I left last time "Debug missing tool calls in chat API". This is how it showed up in chat. I remembered the last time when debugging, it showed up as the model thinking instead of calling tool. It might be due to the phrase "If the first retrieval doesn't return enough, try a shorter or rephrased query". The model treats it as a suggestion rather than a command to query database a second time from a different query.
