LLMs benefit from different levels of external context depending on the task—from basic prompting to retrieval-augmented generation with causal reasoning—and choosing the right approach requires understanding your specific use case and deployment constraints.
This survey examines how LLMs can be enhanced with external information at inference time, progressing from simple prompting techniques to sophisticated retrieval systems that incorporate causal reasoning. It provides a framework for understanding when and how to add structured context to improve model performance.