SOTA - Generator

SOTA-Generator is an LLM app to create state-of-the-art articles from scientific PDFs. The interface is Streamlit-like, Graph of Thoughts is the prompting technique and Gemini is the brain behind it.

Comment

Inspiration

The start of a SOTA-Generator journey was marked with one simple yet deep observation: The abundance of research, i.e. scientific knowledge, is still rather hard to make accessible and synthesize in a comfortable and fast way. Seeing the huge bulk of unread or not fully used scientific papers due to their density, we had one idea of making a tool that frees this gold mine of knowledge.

What it does

The SOTA-Generator tool makes it a breeze to upload your scientific articles in PDF format using a Streamlit-like interface that steers the user easily through the process of generating a state-of-the-art. The underlying drive for the tool is Gemini 1.5 Pro-LLM combined with the Graph of Thoughts (GoT) prompting technique to interpret, structure, and draft a comprehensive state-of-the-art from the uploaded document.

How we built it

Development of SOTA-Generator was done in a multifaceted manner:

  • Interface Design: Design an intuitive interface akin to what Streamlit has for ease of upload and navigation through PDFs.
  • AI Integration: Integrate Gemini to manage the system's planning and processing activities.
  • GoT Prompting: Develop a graph of complex prompts to guide the LLM to achieve the purpose.
  • Graph Visualization: Implement a visualization of the reasoning process that the LLM carried out.
  • Testing and Refinement: Institute continuous testing and refining of the AI outputs, based on feedback from beta testers in the academic community.

Challenges we ran into

The path is laid with thorns. We were challenged by the massive task of making AI more capable of understanding and interpreting highly technical, niche scientific jargon through advanced prompt engineering. Common AI Chatbot services fail at this task. On the other hand, data extraction for complex layout PDFs appeared like a terrifying experience that required innovative solutions. We also had to adapt the closed GoT framework to be able to visualize the reasoning process of Gemini and, hence, be able to debug and refine the outputs.

Accomplishments that we're proud of

We are proud to have managed to turn our vision into a working tool that enables more accessible access to scientific research. The successful merger of Gemini with the GoT AI technologies gave a better understanding of complex data structures. We have been able to develop a complex graph-reasoning structure that enhances the Gemini's original capabilities to create a complex, higher-quality state-of-the-art document about scientific articles. On top of that, we have created a graph visualization engine for the GoT library that allows the user to inspect the LLM thoughts as the graph is traversed.

What we learned

In this project, we learned about the abilities and potentials of large language models, the details of prompt engineering, and UI optimization. These lessons were highly technical but also exciting to learn, improving our skills and appreciation for AI possibilities in transforming information access.

What's next for SOTA-Generator

In line with our constant improvement of SOTA-Generator, we have set our sights on several ambitious expansions. Firstly, we seek to extend compatibility with a variety of large language models, such as GPTs, Claude, and LLama, among others. Secondly, we intend to get semantic information from images, equations and figures to enrich our summaries. Finally, we are developing a complete exploitation plan that guarantees its widespread community. It will also permit covering the generation's cost and adhering to the regulatory standards. Hence, making SOTA-Generator a "one-stop" tool to access cutting-edge research.

Built With

Share this project:

Updates