I Tried Making the Trending 'GPTs' Myself
Since ChatGPT's GPTs feature was released, I created a GPT for summarizing internal company documents. I added about 30 PDFs to the knowledge base and set the prompt to 'Summarize in 3 lines and extract key keywords.' The accuracy was surprisingly high. However, there are occasional mistranslations in Korean documents, but I think refining the instructions more specifically could improve that. Anyone else who has made GPTs, share your tips!
3 answers
Oh, I've also made a GPT for a similar purpose, and 30 PDFs is quite a lot! I had around 10 and it still performed pretty well. As for the Korean mistranslation issue, I found it improved a bit after adding clear conditions like 'Answer only in Korean' in the Instruction. By the way, can you add other file types besides PDFs to the knowledge base?
I created a GPTs for paper summarization as a personal project, but the search speed slowed down a bit when I added 50 PDFs. One tip I can give is that adding 'cite sources' to the prompt significantly boosts reliability. Korean mistranslations still seem like a challenge. Given the nature of GPTs, I think the key is to repeatedly refine the Instructions.
Wow, I just tried GPTs myself too, and it was surprisingly easy and fun! I made one for summarizing meeting minutes, and the 3-line summary feature is really useful. But I also ran into some Korean mistranslations—have you tried adding examples to the 'System Prompt'? For instance, giving it a pattern like 'Input: ... Output: ...' made it much more stable!