Large language models (LLMs) through their training process learn and store a lot of general knowledge. However, the distribution of knowledge has a long-tail pattern of many infrequently used facts. This poses a challenge for LLMs when a query requires information that is on the long-tail. It is on such queries that LLMs have a tendency to hallucinate. Retrieval-augmented generation (RAG) improves an LLM's ability to answer such questions by retrieving the needed information and adding it to the LLM's context. Part of this thesis proposal looks at this retrieval algorithm and analyzes how it works. A crucial part of RAG is the retrieval corpus itself. Most RAG benchmarks use Wikipedia or Wikipedia-like texts as their retrieval corpus. These texts are written in a neutral and factual tone. However, when RAG systems retrieve internet-based content, they encounter text with diverse tones and linguistic styles, introducing challenges for downstream tasks. This thesis addresses this problem by constructing and validating datasets that introduce sarcasm and emotional variation into retrieved passages, and by developing methods that enable LLMs to better comprehend such pragmatically inflected inputs. In doing so, it explores both prompt-based and translation-based approaches for adapting text tone, and analyzes how emotions are represented in LLMs’ latent spaces, showing how these insights can be leveraged to improve RAG reading performance.