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 itThis 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!
There is no officially supported way in Phoenix to automatically render markdown in a pretty format from structured LLM outputs in traces; Phoenix currently displays the raw JSON or text as-is, and the UI does not provide a markdown rendering toggle for arbitrary output fields. Workarounds like tool-based tracing or manual span creation are common but not ideal, as you noted, and custom SpanProcessors cannot modify span attributes after the span ends because spans are read-only at that stage. This limitation is confirmed in the documentation and multiple GitHub issues, and there is no configuration or API to force markdown rendering for structured outputs in the current Phoenix UI or Langchain instrumentation setup聽see discussion,聽see also,聽see also.If you want markdown rendering, the only current options are to (a) ensure your output is a plain markdown string in the standard LLM output field (not nested in a custom structure), or (b) request a UI feature from the Phoenix team to support markdown rendering for structured outputs. For now, your workarounds are the only available approaches, and there is no "clean" or officially recommended solution for this use case.Would you like more detail on how Phoenix parses and displays output fields, or on possible UI feature requests?Sources: - see discussion - see also - see also
馃挕聽Hint: Mention RunLLM in the thread for followups.
Hey Tristan, HNY. Given today is a company holiday we might be delayed in response. What part of your tracing are you wanting to render in markdown? In general we have support for markdown in the llm messages but depending on the place we might be able to just do it in the UI
Hi Mikyo, happy new year too ! Thank you for your kind help 馃檪 I have a structured output LLM message that has text fields, and I'd like to be able to render such fields as markdown instead of raw text. This happens because Phoenix displays the structured responses as JSON/dictionaries. It's not a problem for most use cases, but when dealing with long generation we lose a lot of readability. If we're talking features, being able to focus on a text element and have a modal/side panel with the prettified text in markdown would be super useful. Also, having long-text truncation + expander/read more ellipsis could truly help for manually reviewing long conversations. Another feedback that I got is that I feel difficult, when dealing with long text generations, to visualize easily the generation metadata : reasoning tokens, output tokens, etc. I spend a lot of time in scrolling, and Ctrl+F often struggles to find the text I want to search.
Thanks for the feedback here. I think I roughly get some of the usability issues but might badly represent them. Copying Srilakshmi C. to help guide our feedback. Keep the feedback coming.
To help us develop more concrete ideas for improving the UI, could you share screenshots (with redactions if necessary) of the specific areas you鈥檙e currently having trouble with? Having visual examples will enable us to move forward more efficiently.