Aria doesn't seem to give OpenAI access to the PDFs (full text) attached to items in Zotero 7

[inspiredearth]

Hi there,
It's come to my attention when I am using Aria to question a particular paper, it's only giving OpenAI access to the abstract in a parent item, but not to the "Full Text" PDF that's attached to the item.

Is there some special way to have it allow OpenAI to access the PDF? Without that, I get responses like this (bold emphasis mine):

The abstract of the paper "Thirty years of fluoridation: a review" does not mention the consideration of the impact of topical fluoride application through toothpaste. The focus is on the benefits of water fluoridation in reducing dental caries and its role in public health. Without access to the full text, it is unclear if the impact of topical fluoride application is discussed elsewhere in the paper.

[lifan0127]

Hi @inspiredearth, I initially implemented the fulltext feature based on local vector database. However it had serious performance issues and had to be removed. I am working on bringing it back through the OpenAI File Search feature now.

[ww-wolt]

This would be amazing! Without this feature the tool is currently unfortunately mostly useless to me, as I'm also fast in scanning through the abstracts myself. However I would love to have deeper discussions with the content of specific papers and ebooks.

[cfessler]

Dear @lifan0127 I am a big fan of your work with ARIA. And I am also craving to see the full-text feature in ARIA. So thats why I wanted to ask if there is a way to help you to expedite your progress to the feature as you described.
To be honest I dont have unlimited funds :-) but maybe we could find some people willing do some patreon donations or so? Would you be open to it?

[lifan0127]

Hi @cfessler Thanks for offering support! That's very kind of you. I am hoping to release an early version of fulltext Q&A in the next few weeks.

[smartroaddev]

Hi @lifan0127, Thanks for amazing plugin. I believe local vector search will save the cost in long run. Do you trying to use small size embedding model such as nomic-embed or even smaller model 22M model I believe it quite fast for my other project? What do you think?