Sapourdiv Fol

Image credit: dall$\cdot$e

Chatting with My Papers: A Demo ChatGPT

As a researcher in deep learning algorithms for wireless signal processing and localization, I spend most of my days buried in research articles. Reading through academic papers can be a daunting task. I wondered whether it is just me or are these ChatGPT demonstrations starting to feel like a sleight of hand? 🎩🤔 What if there is a way to quickly and easily setup my own chatGPT to chat with my papers? As I chat with chatGPT, I can’t help but feel dissapointed that ChatGPT doesn’t recognize me … Unfortunately, the general limitations of the transformer-based LLM models on the number of input tokens have made it difficult to use the chatGPT-like tools for documents with long pages. (This article already feels outdated, as in less than two weeks, there have been multiple new solutions and the recent introduction of ChatGPT add-ons by OpenAI. The world of LLMs is changing rapidly.)

Inspired by recent community work on llama_index and langchain, I decided to build a simple web app to chat with my research articles. My hope is that others will find my research more valuable and consider citing my work instead of simply avoiding (to cite) it.

Despite I was hoping to gain some more insights on autoregressive self-superivsed approach of LLMs, and the RLHF, I ended up wasting time to patch multiple pieces of making the chat 💬 to work using flask 🌮, langchain 🔗, pinecone 🦔,llama_index 🦙, netlify 🌐, heroku ⚙️, AWS S3 file storage, and not to speak about the front-end and javascript. However, thanks to chatGPT, experience was smooth 🥳. Since I am not a web developer, I do not plan to maintain this app. Hence, I kept only the Q&A part of the app and removed the chain Tool. Anyone who wants to use the app can simply ask questions about my research papers. I can be reached for any questions or comments if you want me to share this piece of demo app for your own research papers. Important is that it works, if you have money to pay for questioning your papers 😅.

As I worked on this project, I realized that customized solutions like this one are more for hobbyists than anything else. There are so many details that go into making an app like ChatGPT state-of-the-art, and most startups and individuals simply don’t have the time or resources to bring it to that level. I am excited to see what the future holds for LLMs and chatbots.

Artan Salihu
Artan Salihu
Ph.D. Student in Wireless Communications Group of TU Wien

PhD student at MCG group of TU Wien. My research interests include deep learning and wireless localization.