Tracking your scientific niche?
AI/lluminator
Technology
Filtering Scientific Literature
Adapting RAG for Tracking Science
AI/lluminator employs a method inspired by retrieval-augmented generation to filter scientific literature efficiently. It works by using an embedding model to generate a vector representation of each paper's title and abstract, capturing the key meaning of these texts. These vectors are then compared to the vector generated from your research interest prompt. The most similar papers are selected and sent to you daily. This approach helps you access the most relevant information from the scientific literature without needing to read the entire paper.
Getting the Embedding Model in the first place
To get the embedding model to begin with, the model was pre-trained on a synthetic dataset generated by prompting a large language model (LLM) to create specific research interests tailored to each paper in the database. For each paper, the LLM was shown the content and then tasked with generating a corresponding research interest. This process was repeated for every paper in the database, resulting in a vast collection of 'interest' -> 'paper' pairs. These pairs were then used to train the embedding model, enabling it to accurately map research interests to relevant papers based on their content.
Fine-Tuning the Embedding Model
Reinforcement learning with human feedback (RLHF) is a technique where a model is refined based on interactions with users. Feedback from users, such as which papers they click on or engage with, is used to improve the model's performance over time. In the case of the embedding model, it is fine-tuned using this feedback to better align its recommendations with user preferences. By analyzing user behavior—specifically which papers are clicked on—the model learns to prioritize more relevant content. For legal and ethical reasons, however, we ensure that this data is anonymized and cannot be attributed to any specific user, preserving privacy while still enhancing the system's effectiveness.
