Highlight, a platform providing full-stack visibility into application errors, has implemented a new feature using a language learning model to effectively distinguish between significant and inconsequential errors. The system employs a 1024-dimension open-source model to tag errors, such as "authentication error" or "database error," and group similar errors, even if they have different stack traces, by using Euclidean distance between their embeddings. This approach, which leverages the OpenAI embeddings API and is hosted on Hugging Face, allows developers to quickly identify and address issues by reducing noise and focusing on actionable errors. The first version of this error grouping logic has been integrated into their cloud product, and the Highlight team is open to further suggestions from the community for expanding the use of LLM tooling in observability.