The ability of modern machine learning systems to generate contextually relevant and coherent text is often challenged when external knowledge must be integrated into the generative process. Addressing this challenge, a novel approach was developed to dynamically reduce token-level uncertainty during retrieval-augmented text generation. By modifying the Mistral model to incorporate a token-level uncertainty estimation layer, the generation process was significantly enhanced in both precision and fluency. Experiments conducted on multiple retrieval-based tasks demonstrated substantial improvements in retrieval precision, perplexity, and uncertainty reduction, showing the modified model's ability to generate more coherent outputs, particularly in scenarios involving ambiguous or incomplete retrieved information. The recalibration of token predictions during the generation process directly contributed to the model's robustness and reliability, making it more adaptable to the complexities of retrieval-augmented contexts. Results indicated that the proposed method offers a significant advancement in improving the alignment between retrieved documents and the generated responses, leading to more accurate and contextually appropriate outputs across diverse datasets.