ScholarTurbo: Use ChatGPT to chat with PDFs (supports GPT-4)

If you have many papers, perhaps upload the latex source code of all of them into some vector database like pinecone and use them with LangChain for retrieval.

I have nothing to do with it, but to me it's the best, they seem to put some real effort on it since day one.

[1] explainpaper.com

I say reasonable, because in a company growing that fast, everything is on fire all the time and security is the last thing considered.

There appears to be a lot of apathy/total disregard to giving users any clear clue.

It might not matter to me or this audience - we probably know the direction our data is heading - but it matters to less techie people, to ordinary employees and to civil servants and government people at the bottom and the top, who have no notion where their data is processed and who they are trusting with it.

With all these ChatGPT wrappers, the hard bit is pretty much a monopolized commodity, so they better do the annoying bits really really well.

This is the second PDF chatbot I have tried, and they both suffer from the technical constraint of how many tokens you can fit into ChatGPT, meaning they refuse to answer global questions like "are there any mistakes in the document since the ChatGPT may be unaware that there is a document at all, but can answer, for example for a CV, "what jobs might you hire this person for".

This might be fixable by making round trips, for example asking chatgpt YES/NO for does this query apply to this 1000 token section of a larger document, etc. etc.

And of course if they get this right, and do it well, it becomes very useful. If it can answer based on 10000 stored pdfs, word docs, JIRA tickets, confluence pages and slack threads, then it would be super useful for organizations. It would hardly be just a wrapper then though.

But yes Twitter is a wild Wild West right now of people ripping ideas off any trending repos

Having spent quite a lot of professional time working through real-world problems with GPT4, I'll say it very rarely does something I'd call hallucination. It almost never invents crazy things out of whole cloth. Sometimes it makes wrong guesses about things.

You gotta keep your own head and be skeptical. Often you can tell if it's coming unhinged. If it makes mistakes and you notice, it'll definitely own them unlike many LLMs. But of course it's not perfect.

When you add parts of the content of the PDF to the prompt, which is how these tools work, then it is very likely to produce a completion consistent with the PDF.

I say “mainly” and “very likely”, but in my experience I have never had a hallucination when the prompt is augmented with factual data.

Hallucinations occur with regularity when the prompt does not contain what is requested to be completed.

I get that people have an objection to how LLMs invent things, but so do people. You should always verify things.

I also don't use Langchain.

This doesn't make the users idiots, it just means they value their time.