I think what a lot of folks are forgetting to answer, and you may not realize yourself: you are not analyzing scientific texts with this sort of approach.. I'm not an "open webui" user myself.. but it's concerning that you're "uploading" the texts in your description.

This is what I assume is happening: You 'add' your documents to the ui, but those aren't being given whole-form to your LLM, they're being split into a bunch of small chunks and indexed with an embedding model - then, when you ask your question, your imput is being similarly processed and compared against all those chunks.. those top N (maybe 5?) Chunks get added to the prompt input when the UI runs the actual LLM request and you get what you get. This is generic, vanilla, RAG; it has its uses, but you need to read up about it so you know what's happening and why it's not answering your question well.

If this is what's happening, you have three problems: one, the llm is given no context about where that chunk came from and it may have chunks from different papers assuming different things. Two, the llm knows zero about the rest of the content you've added, for that particular request, zero. Three, your lookup (embedding model) is garbage - it's trained on general text and knows concepts about general things.. but research papers have highly specific language typically, or use common words in different ways, etc. - so the way that the model sees similarity doesn't actually map all that well into niche research paper land.

Comparing to "ChatGPT" isn't reasonable here, because chatgpt isn't a model, it's a platform, and it's doing many things behind the scenes to compensate for just the model+rag situation. Maybe open webui has similar tools, I don't know, but you'd need to learn them - just attaching a bunch of pdfs and asking questions without the understanding of what's under the hood is like driving a manual in first gear and wondering if your car can't go 60.