Tom Sobolik, Barry Eom, and Shri Subramanian discuss the proliferation of managed LLM services like OpenAI, Amazon Bedrock, and Anthropic, which have introduced possibilities for generative AI applications. However, introducing non-deterministic LLM services can increase the need for comprehensive observability, as debugging LLM chains can be challenging. The authors argue that tracing your LLM chains can help examine each step across the full chain execution to more quickly spot errors and latency and troubleshoot issues. They also explain how chains are essential for integrating LLMs into application workflows, helping structure interactions between the model and other components of the application. Chains extend LLM functionality to overcome limitations such as limited knowledge or ability to execute tasks not related to text generation. The authors discuss challenges and goals of monitoring LLM chains, including tracking code and request errors, pinpointing root causes, identifying hallucinations, and tracking token consumption. They also explain how instrumenting LLM chains presents unique considerations, requiring different approaches for instrumentation. Datadog's LLM Observability SDK can be used to instrument your LLM chains for debugging and performance analysis, allowing you to create trace spans within your code, add metadata and context, and monitor incoming requests within a central view.