Tristan D.

Hi team! I'm working with Arize Phoenix for observability and I'm wondering: what's the recommended way to render markdown text in a pretty format when it's part of an LLM's structured output? I'm using Langchain with auto_instrument=True in phoenix.otel.register to register a provider. Currently, I'm getting raw JSON in the traces, but I'd like to display nicely formatted markdown instead. I've explored a few workarounds: Option 1: Tool-based tracing

from langchain_core.tools import tool

@tool
def trace_pretty(content: str) -> str:
    """Trace some content in a pretty way."""
    return content

def run_agent(...):
    result = agent.invoke(...)
    structured_output = result["structured_response"]
    trace_pretty.invoke(structured_output.text_field)

This works but feels like a workaround that pollutes the traces. Option 2: Direct span tracing

from opentelemetry import trace

tracer = trace.get_tracer(__name__)

def run_agent(...):
    with tracer.start_as_current_span("run_agent", openinference_span_kind="tool"):
        result = agent.invoke(...)
        structured_output = result["structured_response"]
        span.set_output(structured_output.text_field)

This creates a new root span that's decoupled from the current tracing context, and it also feels like a workaround (we declare a tool span only to have pretty traces...). Option 3: Custom SpanProcessor

from phoenix.otel import SimpleSpanProcessor

class MySpanProcessor(SimpleSpanProcessor):
    def on_end(span):
        # Span is read-only here, can't modify it

This doesn't work because the span is read-only in on_end , and in on_start the span has not begun to reference the LLM answer. Is there a cleaner, officially supported way to achieve this? Thanks for any guidance!

Hi, I'm Tristan. I use Arize Phoenix as AI tracing stack for my company, in which I'm working as a freelance AI Engineer. Based from France 🇫🇷 Thanks for this great product !