Rishikesh Ksheersagar

Rishikesh Ksheersagar

Masters in Data Science at University of Michigan
NLP Researcher at LIT Lab @ UMich
Previously - Data Science Manager at Mu Sigma Inc.

PapeRet: Research Paper Retrieval System

PapeRet is a specialized Information Retrieval (IR) system designed to address the challenges of efficiently retrieving academic research papers in machine learning, AI, and NLP. The system combines traditional IR techniques with advanced NLP methods to enhance search precision and deliver meaningful results.

Key Features:
  1. Recursive Metadata Collection:
    • Utilized the OpenAlex API and ArXiv API to gather metadata for ~98,000 papers, starting from 30 influential seed papers and recursively expanding the dataset.
    • Implemented web scraping and PDF parsing to extract full-text content, overcoming challenges like rate limits, CAPTCHA protections, and diverse website structures.
  2. Advanced Query Processing:
    • Integrated TF-IDF and BM25 rankers with enhanced query augmentation techniques.
    • Leveraged zero-shot inference using LLaMA for query tokenization, enabling structured extraction of keywords, authors, and temporal constraints.
    • Introduced weighted scoring, author/year boosting, and re-ranking strategies for improved retrieval.
  3. Summarization and UI:
    • Developed concise, retrieval-augmented summaries using LLaMA’s Retrieval-Augmented Generation (RAG).
    • Built a Streamlit-based UI for intuitive interaction, enabling users to perform parameterized searches and view AI-generated summaries.
Results:
  • Processed ~98,000 research papers with a final curated dataset.
  • Achieved a MAP@10 of 0.539 and NDCG@10 of 0.81, significantly outperforming baseline approaches.
  • Enabled precise searches for complex queries, such as author-specific works or multi-modal topics.
Future Directions:
  • Explore supervised fine-tuning of LLaMA for improved query understanding.
  • Refine PDF parsing methods for higher accuracy in full-text extraction.
  • Extend the system to visualize citation networks and identify research gaps.

PapeRet provides a robust framework for navigating the growing corpus of academic literature, offering researchers a powerful tool to streamline their workflows and focus on innovation.

You can find the codebase on GitHub.