I've been uploading PDFs to Gemini 2.5 and ChatGPT (o3), but I notice they often hallucinate when I ask questions about the content. It seems like ChatGPT o3 is especially prone to this. Is there a way to give them a persistent memory so that when I upload a document or paste text, they can recall it accurately? I want to be able to ask for specific parts of the PDF or get the full text back exactly as it is in the original document.
2 Answers
Hallucinations are pretty common with large language models, especially the more complex ones. The best practice is to enforce guardrails and use known reference data. If you're getting offbeat responses, consider trying the API multiple times to see if you get a more accurate output that way.
I totally get your frustration! From what I've found, Claude.ai seems to handle documents without going bonkers, unlike ChatGPT. It could be worth giving it a try if you're looking for accuracy with PDFs.
Related Questions
xAI Grok Token Calculator
DeepSeek Token Calculator
Google Gemini Token Calculator
Meta LLaMA Token Calculator
OpenAI Token Calculator