This repository contains code, datasets, visualizations, and analysis for the research project, LEDGE : Leveraging Dependency Graphs for Enhanced Context Aware Documentation Generation. a novel framework that integrates dependency graphs with large language models (LLMs) to automate the generation of structured, context aware software documentation
├── README.md
├── LICENSE
├── datasets/
│ ├── average_evaluation_results.csv
│ ├── evaluation_results.csv
│ └── quantitative_analysis.csv
├── diagrams/
└── notebooks/
├── evaluation.ipynb
└── graphs.ipynb
LEDGE addresses the problem of incomplete or outdated documentation in large codebases. It does so by:
- Constructing dependency graphs from source code using custom parsers.
- Storing these graphs in MemGraph using Cypher queries for fast retrieval.
- Embedding code symbols for semantic similarity matching.
- Using a HybridRetriever that combines graph traversal and vector search to provide the LLM with rich context.
- Generating accurate, structured documentation with GraphRAG-based retrieval-augmented generation.
We evaluated LEDGE on five diverse open-source repositories:
| Project | URL | Description |
|---|---|---|
| Next.js | https://github.com/vercel/next.js | React framework for server-side rendering |
| MobX | https://github.com/mobxjs/mobx | State management library |
| Turborepo | https://github.com/vercel/turborepo | Monorepo build system |
| Prisma | https://github.com/prisma/prisma | Type-safe database toolkit for TypeScript |
| Gatsby | https://github.com/gatsbyjs/gatsby | React-based static site generator |
LEDGE outperforms traditional documentation in both qualitative and quantitative evaluations using metrics like BLEU, ROUGE-L, BERTScore, and Jaccard similarity.
The notebooks/ folder includes interactive visualizations and reproducible analysis:
evaluation.ipynb— Analyzes BLEU, ROUGE, and BERTScore comparisons between LEDGE and baseline documentation.graphs.ipynb— Plots for research paper.
These notebooks replicate core results from the paper and demonstrate the impact of graph-based retrieval in documentation generation.
This project is licensed under the MIT License. See the LICENSE file for full details.
For questions, collaborations, or feedback, feel free to reach out:
Mihir Panchal – mihirpanchal5400@gmail.com
Arnav Deo – arnav.nitin.deo@gmail.com
Varad Prabhu – varadprabhu111@gmail.com
Prinkal Doshi – prinkaldoshi@gmail.com
Chetashri Bhadane – chetashri.bhadane@djsce.ac.in
Pranit Bari – pranit.bari@djsce.ac.in