iRAT: Replanning and Controlled Retrieval for Robust LLM Reasoning

Note

Please star ⭐ the repository to show your support.

Architecture:

Setup steps

1. Clone the repository

git clone https://github.com/prane-eth/iRAT

2. Open a terminal in the "iRAT" folder and install the required packages

pip install -r requirements.txt

3. Ensure the environment file (.env) is set up with correct values, based on ".env.example" file.
4. To host the models, run:

python host_models.py

5. To test the pipeline, run the following command in the terminal in "irat" folder:

python pipeline.py

6. To start the web server, run:

python server.py

Demo

Screenshots are available in the Demo folder.

Citation

@misc{vadlapati2025irat,
  author       = {Praneeth Vadlapati and Zeeshan Ali},
  title        = {iRAT: Replanning and Controlled Retrieval for Robust LLM Reasoning},
  year         = {2025},
  month        = {July},
  howpublished = {\url{https://github.com/prane-eth/iRAT}},
  note         = {GitHub repository}
}

Team members:

Praneeth Vadlapati (@prane-eth) <praneeth.vad@gmail.com> (Main coder and author)
Zeeshan Ali (@zeeshan5k)
Aryan Singh (@ekk012) <aryansingh729@gmail.com>
Alvaro Arteaga (@LagrangianPoint)

Contributions:

• Praneeth Vadlapati: Pipeline, result-filter module, evaluation, most of the code and paper, and team leadership.
• Zeeshan Ali: Architecture, uncertainty evaluation, and Chain Evaluator model.
• Aryan Singh: Retrieval module with budget control, dataset analysis, MBPP pre-processing, pipeline wireframe, and bug fixing in budget control.
• Alvaro Arteaga: User input scanning, and the idea of spam website filter.

📧 Contact

For personal queries, please find Praneeth's contact details here: https://prane-eth.github.io/

🪪 License

Copyright © 2025 Praneeth Vadlapati and team
Please refer to the LICENSE file for more information.
To request a permission to use my work, please contact me using the link below.

⚠️ Disclaimer

The code is not intended for use in production environments. This code is for educational and research purposes only. No author is responsible for any misuse or damage caused by this code. Use it at your own risk. The code is provided as is without any guarantees or warranty.

---

Base paper:

RAT: Retrieval Augmented Thoughts Elicit Context-Aware Reasoning and Verification in Long-Horizon Generation

[GitHub] [Website] [Published Paper] [Pre-print]