From about 9000 astronomy preprints on arXiv in all of 2005 to a staggering 5000 in just the first third of 2025, the relentless pace of discovery leaves little time to absorb it all.

Currently, you can find the papers most relevant to your work among the dozens published daily using keywords or rating systems. Keywords are too simple, and most of these rating systems do not allow you to define personally what you feel like reading on a given day.

Cosmic Compass is here to make your life easier, fetching papers based on what you define as your interests and summarising the results for you.

Cosmic Compass uses the power of GenAI to maximize your efficiency, allowing you to get a list of papers that are relevant to your interests, plus a short summary of their key results, so that you can truly look at the headlines you care about over your morning coffee before you delve into dissecting the full texts. Through the combination of prompt engineering, embeddings, few-shots prompts and document parsing, Cosmic Compass can look at the most recent papers, and find the ones that match what you're looking for faster than you can imagine. You can access it here.

What does Cosmic Compass actually do?

It starts by asking you about your research interests. For example, you could say something like this:

research_interests = "I am interested in globular clusters in an extragalactic context and how they connect to galaxy evolution."

Then, it goes out and grabs information (metadata) about research papers. It uses AI to understand the meaning of these papers (by creating something called 'embeddings') and compares them to your interests. This helps it figure out which papers are most likely to be relevant to you. A code snippet of the section that creates embeddings can be found in the widget below under 'Generating embeddings'.

It then uses a powerful language model (like Gemini Pro, in this case) to review the top papers, give you a curated list of the most interesting ones, and briefly explain why they fit your research. For this we use a carefully curated prompt, as you can see in the widget above, under 'Prompt'.

To facilitate efficient access to these papers, the system then refines the LLM's output, extracting arXiv IDs and generating direct web links. Finally, the code downloads the full-text PDFs, extracts their content, and employs an LLM with few-shot prompting to create concise summaries of the key findings, methodology, and significance. Each one includes the paper's title, authors, and a link back to the original arXiv page.

The end result is a collection of easily digestible, AI-powered summaries (see 'Summaries' code section above), enabling researchers to quickly grasp the essence of the latest arXiv publications. In essence, this behind-the-scenes code acts as an intelligent research assistant, significantly enhancing the efficiency of literature review.

Looking ahead, this insightful arXiv summarization project holds exciting potential for automation and wider dissemination. By integrating with a robust email service like SendGrid, the generated daily digests could be transformed into a fully automatic newsletter. Imagine waking up each morning to a curated email in your inbox, containing concise summaries of the latest relevant research directly from arXiv. SendGrid's reliable email infrastructure would ensure consistent delivery, while its features could be leveraged for further enhancements such as personalized content based on user preferences, improved email formatting, and tracking of engagement. This evolution into an automated email service would significantly enhance the project's usability and reach, making it an indispensable tool for researchers and enthusiasts alike to stay informed effortlessly.