Afforai might be able to do stuff like this. I haven't tested it myself yet, but the service also seems to have some other features that might be relevant for your use case.

Don't have an answer but I'd be interested in something like that too. I know Microsoft released a freely available lightweight LLM that's supposed to make it easier for people to run it locally, called Phi3. Decent article from ars technica: https://arstechnica.com/information-technology/2024/04/microsofts-phi-3-shows-the-surprising-power-of-small-locally-run-ai-language-models/

I have used this small R package that allows you to read the text content of a PDF and send it to a local llama model via ollama or one of the large LLM APIs. I could use that to get structured answers in JSON format on a whole folder of papers, but the context length of a typical model is only long enough to hold a single (roughly 40-page) paper in the memory. So I had to get separate structurer answers on each paper and then generate a complete summary from those. Unfortunately that is not user-friendly yet.

I don't know of one, but I too would be interested to see what this looks like.

How do you currently store and organize PDFs? I used to use Mendeley during grad school, and honestly I really, really liked it. But being able to ask a question and get a natural language response that suggests which papers might contain insights when taken together would be an incredible asset.

That would likely be a language model finetuned on said material. The problem is feeding PDFs as a structured data source for the model to ingest. The finetuning can't happen with random unstructured PDFs

Chroma is supposed to be able to import a ton of information into a vectorized format that lets you search through it in a way that's semantically meaningful, so you (or your tool) can sort of pick out the stuff from a huge batch of source material that you need to pass to the LLM for any given query.

I played around with it a little bit and I wasn't able to determine if it was a real thing or just a weird AI hype thing, but people seem to take it seriously. I would bet that someone's attempted to make a little system on top of it that lets you do stuff like what you're wanting to do (since that's what it's made for), but IDK how well it would work... might be useful to search for stuff adjacent to Chroma or vector databases to see if there are tools like that, though.

Look into RAG using a vector database, this is exactly what they're for. https://www.linkedin.com/events/buildaragapplicationontheaistac7191489677017649153

I don't think you can use Retrieval Augmented Genaration or vector databases for a task like that. At least not if you want to compare the whole papers and not just a single statement or fact. And that'd be what most tools are focused on. As far a I know the tools that are concerned with big PDF libraries are meant to retrieve specific information out of the library. Relevant to a specific question from the user. If your task is to go through the complete texts, it's not the right tool because it's made to only pick out chunks of text.

I'd say you need an LLM with a long context length, like 128k or way more, fit all the texts in and add your question. Or you come up with a clever agent. Make it summarize each paper individually or extract facts, then feed that result back and let it search for contradictions, or do a summary of the summaries.

(And I'm not sure if AI is up to the task anyways. Doing meta-studies is a really complex task, done by highly skilled professionals of a field. And it takes them months... I don't think current AI's performance is anywhere near that level. It's probably going to make something up instead of outputting anything that's related to reality.)