LLM-Powered Research Paper Discovery Tool

A lightweight tool that retrieves relevant research papers from Arxiv based on user queries, leveraging a fine-tuned open-source LLM and Cursor IDE integration via MCP.

!Demo

Key Features

• ✅ Lightweight: Powered by a custom fine-tuned Meta Llama 3.2 1B Instruct model, which can be loaded unquantized on a 16GB Macbook Air.
• ✅ Open Source: Model available on 🤗 - model id Shaikh58/llama-3.2-1b-instruct-lora-arxiv-query. Papers are retrieved from open source Arxiv API, citation counts from OpenCitations. Avoids the need for paid proprietary model APIs.
• ✅ Cursor Integration: Seamless workflow through MCP server integration. No need to leave the IDE or load the model yourself.
• ✅ Custom Search: Supports various search constraints and filtering options such as citation count and publication date

User Guide

First, create a Hugging Face account and obtain an API token.

Request access to the Llama 3.2 1B Instruct model on Hugging Face: https://huggingface.co/meta-llama/Llama-3.2-1B-Instruct

Clone the repository:

git clone https://github.com/shaikh58/llm-paper-retriever.git
cd llm-paper-retriever

Create a .env file in the root of the repo and add your Hugging Face API token:

HF_TOKEN=<your_hugging_face_api_token>

This project uses uv to manage dependencies. Install uv:

curl -LsSf https://astral.sh/uv/install.sh | sh

Install dependencies:

uv sync

Add cursor mcp json file and cursor rules file to your project. Update the path in the mcp.json file with the path to this repository on your local machine.

mkdir -p /your/cursor/project/.cursor
cp .cursor/mcp.json /your/cursor/project/.cursor/mcp.json
cp .cursorrules /your/cursor/project/.cursorrules

Check whether Cursor recognizes the server: From the Menu bar, go to 'Cursor' -> 'Settings' -> Cursor Settings -> MCP.
If everything works, you should see the server in the list. You may need to refresh the list.

Its ready to try out!

Note: This tool loads in a fine tuned 1B parameter model from the Hugging Face hub. It is small enough to fit unquantized on a 16GB M3 Macbook Air.

Tools in Cursor only work in Agent mode. Switch to Agent mode in your chat window, and ask it to find you research papers! It should connect to this tool and ask for your permission to run it. Wait a few seconds and it should return a list of papers.

For best results, only enter your query in the chat window. Do not include any other text. e.g. "Find me papers on transformer architectures published in since 2023 with at least 100 citations".

Deep dive: LLM Fine-tuning

• Base model: Meta Llama 3.2 1B Instruct
• Structured Output: The model is trained to output structured markdown instead of conversational text. This makes it possible to parse the output and construct a query for a search API.
• Generate realistic training data: Synthetically generated dataset designed to teach the model to output structured markdown. Added variance to the training data by passing generated user queries through an LLM to make the user queries more conversational and realistic.
• Dataset size: 50,000 synthetically generated queries enough to achieve satisfactory performance.
• LoRA fine tuning: Fine-tuning process uses LoRA (Low-Rank Adaptation) to efficiently adapt the base model for our specific structured output task while maintaining its general knowledge capabilities.

Dataset Sample

Input query	Label
"Find recent papers on transformer architectures in NLP published since 2023 with at least 100 citations"	"## QUERY PARAMETERS\n\n- **Topic**: NLP\n\n## CONSTRAINTS\n\n- **Citations**: (>=, 100)\n- **Keyword**: transformers\n- **Year**: (>=, 2023)\n\n## OPTIONS\n\n- **Limit**: 10\n- **Sort By**: relevance\n- **Sort Order**: descending"

During training, the input query is also augmented with a system prompt (not shown) to guide the model to output structured markdown.

License

This project is released under the MIT License - see the LICENSE file for details.