<p>There is a long tradition of using sentiment in economic analysis. Sentiment variables such as consumer confidence, inflation expectations and investor sentiment are often used to gauge the attitudes of consumers and firms about the state of the economy. These measures are typically collected through quarterly surveys, which limits their usefulness for more frequent economic monitoring. The increasing publication of online content has made large text collections easily accessible to computer algorithms for sentiment analysis. This article combines the recent accessibility of large language models with topic modelling and lexicon-based sentiment analysis to construct topic augmented sentiment indices (TaMSI) for the South African economy. The information value of TaMSI is also evaluated in an application where I forecast traditional survey-based confidence indicators using a Mixed-Frequency Bayesian VAR. The results find a reduction in forecasting errors with the inclusion of the sentiment indices for longer term forecast. The model framework has been developed and is currently in implementation at the Bureau for Economic Research, Stellenbosch, South Africa.</p>