LangSmith/Q/Is it possible to log a string to a trace in langsmith

• Is it possible to log a string to a LangSmith/Trace?

  • Impetus: it's difficult to get access to log messages in langgraph/Cloud. Could we instead log onto the trace?

  • Research

    • Yes, there are several ways to add custom data to a LangSmith trace:
      • 1. Using metadata (docs):
        • With the @traceable decorator or traceable function:
          • python

            @ls.traceable(
              metadata={"my-log": "my message"}
            )
            def my_function():
              # Your code here
              pass

        • Dynamically within a traced function (docs):
          • python

            @ls.traceable
            def my_function():
              rt = ls.get_current_run_tree()
              rt.metadata["log_message"] = "my log message"

      • 2. Using the context manager (docs):
        • python

          with ls.trace(
            metadata={"log": "my message"}
          ) as rt:
            # Your code here
            rt.metadata["another_log"] = "another message"

      • 3. At invocation time using langsmith_extra (docs):
        • python

          my_traced_function(
            langsmith_extra={
              "metadata": {"log": "my message"}
            }
          )

    • The metadata can be any JSON-serializable key-value pairs
    • You can view these logs in the LangSmith UI or query them programmatically (docs)
    • This provides a way to add logging information that will be associated with the specific trace/run

  • My Questions - Follow-Up

    • What about if I'm running within langgraph/Cloud or langgraph/Studio? All interactions are automatically logged to langsmith without needing to add ls.trace. should I just use the context manager?

      • When running in LangGraph Cloud/Studio:
        • All LangGraph operations are automatically traced to LangSmith
        • You have two main options for adding custom logs:
          • 1. Use get_current_run_tree() to access the current trace:
            • python

              from langsmith import get_current_run_tree
              
              def my_node():
                  rt = get_current_run_tree()
                  rt.metadata["my_log"] = "some message"
                  # ... rest of node logic

              1. Use the context manager for specific sub-sections:

            • python

              from langsmith import trace
              
              def my_node():
                  with trace(
                      name="sub-operation",
                      metadata={"log": "message"}
                  ) as rt:
                      # ... sub-operation logic
                      rt.metadata["another_log"] = "another message"

        • The first approach is simpler if you just want to add logs to the existing trace
        • The second approach is better if you want to create a separate sub-trace

    • What about if I'm inside a test function marked with LangSmith/PyTest/mark @pytest.mark.langsmith?

      • When using @pytest.mark.langsmith:
        • The test function is automatically traced
        • You can use the testing module to log data (docs):
          • python

            from langsmith import testing as t
            
            @pytest.mark.langsmith
            def test_my_function():
                # Log test inputs
                t.log_inputs({"input_key": "input_value"})
                
                # Run your test logic
                result = my_function()
                
                # Log test outputs
                t.log_outputs({"output_key": result})
                
                # Add custom feedback/logs
                with t.trace_feedback():
                    t.log_feedback(key="my_log", score=1.0)

        • The test results and all logged data will be available in the LangSmith UI
        • You can group tests into test suites using LANGSMITH_TEST_SUITE environment variable
        • Test runs can be cached using LANGSMITH_TEST_CACHE to save resources