The system clones and parses repositories with Tree-sitter, produces syntax-aware chunks, and connects natural-language questions to relevant source code through vector retrieval.