Intelligent Change Impact Analyzer

Project Overview

A graph-based system for analyzing the impact of code changes in software repositories.
It extracts structural information from the code, builds a dependency graph, and identifies components affected by a given change.

The system integrates machine learning to classify and rank impacted components, helping developers focus testing efforts and reduce regression risks.

---

Features

• Fetch commits from a GitHub repository
• Select and compare two commits
• Analyze code structure using AST
• Build dependency graph of functions/modules
• Predict impact severity using ML

---

Setup

1. Clone the repository

2. Create a virtual environment:

   python -m venv venv

3. Activate it:

   • Linux/macOS:

     source venv/bin/activate

   • Windows:

     venv\Scripts\activate

4. Install dependencies:

   pip install -r requirements.txt

5. Create a .env file:

   FLASK_DEBUG=1
   MONGO_URI=<MongoDB connection string>
   GITHUB_TOKEN=<GitHub personal access token>
   HF_TOKEN=<HuggingFace access token>
   GEMINI_API_KEY=<Gemini API key>
   GEMINI_MODEL=gemini-2.5-flash-lite

   Optional:

   # Disable all ML tagging entirely.
   DISABLE_ML_TAGGER=1
   
   # Or disable only the local model path.
   # With USE_LOCAL_MODEL=0, the app can still use Hugging Face Spaces.
   # If the Space is unavailable, it will not fall back to the local model.
   DISABLE_ML_TAGGER=local_only
   
   # Defaults to local model inference.
   # Set to 0/false to use the Hugging Face Spaces Gradio client instead.
   USE_LOCAL_MODEL=1
   
   # Optional when USE_LOCAL_MODEL=0
   HF_SPACE_ID=VantaTree/MLCodeTagger

---

ML Setup (Dataset + Training)

1. Build dataset (local only, not stored in repo)

python ml_tagger/build_raw_dataset.py
python ml_tagger/dataset_builder.py

2. Configure training (one-time)

accelerate config

Recommended settings:

If you have an NVIDIA GPU (RTX / GTX)

This machine
No distributed training
Do you want to run on CPU only? → NO
torch dynamo → NO
DeepSpeed → NO
GPU ids → 0
NUMA efficiency → NO
Mixed precision → fp16

If you do NOT have a GPU

This machine
No distributed training
Run on CPU only → YES

---

3. Train model

python ml_tagger/train.py

Model and dataset are stored locally and are ignored by git.

---

Run Application

python app.py

Open: http://127.0.0.1:5000

---

Notes

• ml_tagger/data/ and model/ are excluded from git
• Dataset is generated locally for reproducibility
• Only the final trained model is required for inference

AI Summary

The analysis page can now generate an optional AI summary on top of the deterministic graph analysis.

How it works:

• services/analyzer.py builds the normal impact result
• services/ai_summary.py converts that result into a compact LLM prompt
• If GEMINI_API_KEY is configured, the app requests a structured summary from the Gemini API
• The summary is rendered in the AI Summary section of the analysis page

If AI config is missing or the API call fails, the rest of the analysis still works normally.