The authors also discuss the cost problem associated with using open-source language models, which are often less accurate and more expensive than GPT-4. They explain how LangSmith helped them identify a bug that was causing token usage to grow uncontrollably, and how fixing this bug reduced their costs by 83%. They conclude by stating that they are looking into using more of LangSmith's features, and recommend it to others who have built applications on LangChain.
Key takeaways:
- Dataherald, an open source natural language-to-SQL engine, has been using LangChain to build their product and recently started using LangSmith to monitor their app in production.
- LangSmith helped Dataherald identify and fix bugs quickly, reducing their token usage and costs significantly.
- LangSmith's features for monitoring, debugging, testing, and evaluating applications powered by LLMs are proving to be invaluable for teams like Dataherald.
- Dataherald is looking to further integrate LangSmith's management and evaluation features into their engineering process, particularly for regression testing.