DEXA: Data Exploration via eXplanatory AI

Contributors:

• Syed Ali Haider
• Rao Daud Ali Khan

DEXA (Data Exploration via eXplanatory AI) is a framework designed to translate natural language queries into safe, explainable, and executable code for dataset analysis (multimodal analysis). It leverages large language models (LLMs) alongside schema-aware prompt engineering to generate code, visualizations, and textual explanations of data.

DEXA supports both zero-shot and few-shot prompting modes, with a feedback loop for safety and interpretability. It has been evaluated on real-world datasets through both automated metrics (code execution) and human feedback using multiple LLMs (GPT-4, LLaMA-4).

---

📁 Repository Contents

File	Description
BostonHousing.csv	Real-world dataset used for benchmarking (regression task).
Titanic_Dataset.csv	Real-world dataset used for benchmarking (statistics task).
no_shot.ipynb	Notebook implementing zero-shot prompting with DEXA.
fewshot.ipynb	Notebook implementing schema-guided few-shot prompting with DEXA.
query_eval.csv	Results from human evaluations across LLMs and prompting modes on both datasets.
NLP Final Report DEXA.pdf	Final report detailing methodology, experiments, and findings.

---

🚀 Getting Started

1. Clone the repository:
2. Add OpenAI and Together.AI API in .env file