The tool extracts 13 similarity features from both snippets (token overlap, length ratios, control flow, normalized structure, token sequences, operators, etc.). A small neural network trained on labeled pairs outputs a score; above your chosen threshold the pair is flagged as plagiarism. The model improves with each analysis and with your feedback.