This research analyzes the impact of different retrieval sources on AI code generation quality, revealing which contextual information most effectively enhances repository-level coding.

• In-context code and API information significantly improve code generation quality
• Similar code snippets provide limited benefits compared to other retrieval sources
• The proposed AllianceCoder framework strategically combines multiple retrieval sources, outperforming existing approaches
• Intelligent retrieval selection reduces hallucinations and improves code correctness

For engineering teams, this research offers practical guidance on building more effective code assistant tools that better understand codebases and generate more accurate, context-aware code suggestions.

What to Retrieve for Effective Retrieval-Augmented Code Generation? An Empirical Study and Beyond