NoveltyAgent

NoveltyAgent is the official codebase for our paper:

NoveltyAgent: Autonomous Novelty Reporting Agent with Point-wise Novelty Analysis and Self-Validation

It is designed to help users quickly understand how novel a research paper is by retrieving related literature, comparing the paper's main contributions with prior work, and automatically generating a structured novelty analysis report.

⭐ If you find this project useful, please consider giving us a star — it helps others discover the project and motivates continued development!

Overview

Given a research paper, NoveltyAgent will automatically analyze the main contributions of the paper, retrieve relevant related work, compare the paper against prior literature, and generate a structured novelty report.

The output helps users quickly understand what the paper mainly proposes, which parts are similar to existing work, which parts are potentially novel, and an overall novelty judgment.

Core Advantages

Compared with general-purpose review or deep research systems, NoveltyAgent is designed specifically for paper novelty analysis. Its main advantages include the following:

• Point-by-point novelty analysis — It decomposes a paper into individual novelty points and retrieves evidence for each one separately, enabling a more comprehensive analysis, instead of treating the entire paper as a single query.

• Full-text literature grounding — It builds its reference database from the full text of related papers rather than relying only on abstracts, providing richer details for comparison and helping avoid shallow or overly generic analysis.

• Better faithfulness — It includes a self-validation step to reduce unsupported claims and improve the reliability of the generated report.

Demo

We provide a short demo video to illustrate the usage of NoveltyAgent:

Insight.Agent.11.-.-.Microsoft.Edge.2026-03-12.18-16-27_2_20260316_14452621.mp4

Note: This demo is for demonstration purposes only. In real-world usage, the analysis process takes significantly longer than shown in the video, as it involves extensive literature retrieval, comparison, and validation steps.

Installation & Setup

Prerequisites

• Python 3.9+
• Docker & Docker Compose
• Git
• NVIDIA GPU with CUDA support (for RAGFlow GPU mode and Reranker inference)

Step 1: Clone the Repository

git clone https://github.com/SStan1/NoveltyAgent.git
cd NoveltyAgent

Step 2: Install Python Dependencies

pip install -r requirements.txt

Step 3: Download the Reranker Model

Download the Qwen3-Reranker-4B model from HuggingFace using huggingface_hub:

python -c "
from huggingface_hub import snapshot_download
snapshot_download(
    repo_id='dengcao/Qwen3-Reranker-4B',
    local_dir='./dengcao/Qwen3-Reranker-4B'
)
"

Step 4: Deploy RAGFlow

4.1 Check and Configure vm.max_map_count

Elasticsearch requires vm.max_map_count to be at least 262144. Check the current value:

sysctl vm.max_map_count

If the value is less than 262144, update it:

sudo sysctl -w vm.max_map_count=262144

To make this change permanent, add or update the following line in /etc/sysctl.conf:

vm.max_map_count=262144

4.2 Clone RAGFlow

Clone the RAGFlow repository into your working directory:

git clone https://github.com/infiniflow/ragflow.git

4.3 Replace Configuration Files

Replace RAGFlow's default .env and docker-compose-base.yml with the customized versions provided in this project's Setup/ directory:

cp Setup/.env ragflow/docker/.env
cp Setup/docker-compose-base.yml ragflow/docker/docker-compose-base.yml

Step 5: Start Docker Services

5.1 Start RAGFlow

cd ragflow/docker
docker compose -f docker-compose-gpu.yml up -d
cd ../..

5.2 Start the Reranker Service

Navigate to the Reranker model directory and start the service:

cd dengcao/Qwen3-Reranker-4B
docker compose up -d
cd ../..

Step 6: Configure the Reranker in RAGFlow

After both Docker services are up and running, you need to manually register the Reranker model inside RAGFlow's web UI:

1. Open your browser and navigate to http://localhost:9380.
2. Log in to the RAGFlow admin panel.
3. Go to Model Providers settings.
4. Add a new Rerank model with the following settings:
   • Model type: Reranker
   • Model name: Qwen3-Reranker-4B
   • Provider / Inference backend: VLLM
   • Point the model URL to the Reranker service endpoint launched in Step 5.2.
5. Save the configuration.

Step 7: Configure API Keys

Before running the system, configure the required API keys and related environment variables.

Step 8: Launch the Application

Start the Streamlit frontend from the project root:

streamlit run NoveltyAgent/app.py

The application will be available by default at:

http://localhost:8501

Citation

If you find our work helpful, please consider citing our paper:

@misc{hou2026noveltyagentautonomousnoveltyreporting,
      title={NoveltyAgent: Autonomous Novelty Reporting Agent with Point-wise Novelty Analysis and Self-Validation}, 
      author={Jiajun Hou and Hexuan Deng and Wenxiang Jiao and Xuebo Liu and Xiaopeng Ke and Min Zhang},
      year={2026},
      eprint={2603.20884},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2603.20884}, 
}