Adaptive RAG writeup & Gradio issue

[rmusser01]

title.

[rmusser01]

https://github.com/yale-nlp/MCTS-RAG

https://arxiv.org/abs/2503.21729
https://github.com/hzy312/knowledge-r1/blob/main/README_Search-R1.md
https://arxiv.org/pdf/2505.22095

 Having built systems like this at the production scale; Most likely, something like this:

    fast pdf parsing (doesn't have to be too accurate, just fast)

    check number of tokens in parsed pdf

    if less than 60k, put entire file as system prompt/context to answer on (this is important, because system prompt stays the same as user prompts/conversations can continue to develop/change)

    have an X turn chat history (after X user-assistant conversation pairs, keep only a max of X recent messages in the conversation chain). This is to avoid things going out of context

You'd add bells and whistles with semantic QA by chunking data and retrieving 10k-30k tokens at most to answer queries; if the uploaded data is more than 60k tokens. 

[rmusser01]

https://substack.com/home/post/p-161922553

https://github.com/aizip/Rag-Eval-flow

https://arxiv.org/pdf/2407.12883

https://huggingface.co/lightonai/Reason-ModernColBERT

https://github.com/facebookresearch/ReasonIR/tree/main/synthetic_data_generation
https://gist.github.com/NohTow/d563244596548bf387f19fcd790664d3

https://engineeringblog.yelp.com/2025/02/search-query-understanding-with-LLMs.html
https://github.com/ibm-granite/granite-io/blob/main/docs/design.md

https://universalrag.github.io/
https://patann.dev/