Based on pageIndex, I have built a complete RAG question-answering system with similar performance, supplemented the missing reasoning module, and enabled interaction via a web UI.

[funnamer]

repo：
https://github.com/funnamer/pageIndex-rag-chat
performance：

[bipin-adh]

@funnamer Does it return figures if present as well from pdf ?

[funnamer]

It cannot return images for now. I‘m only using the open-source part of pageIndex for RAG, which only gives the TOC. The full pageIndex should include an extra layer of document splitting to extract per-page source info and images; it then fetches images via an image tool and performs reasoning & QA with a multimodal LLM.

I may implement this part later.

[RajithaBathini]

Hi - could you give me some metrics of how many LLM calls are required in your scenario to Index each doc.
Because when I download and 'run_pageindex' with my docs - its making more than 100 calls to LLM. Am I doing something wrong. Thank you.