Troubleshooting LangChain Tracing with Phoenix in Python | Arize AI Community

Arize AI Community Icon

Troubleshooting LangChain Tracing with Phoenix in Python | Arize AI Community

Roger Y.
·
Hi Daniel. You can ignore the type warning. But it’s unclear why it’s empty. I’ll take a look.
Daniel
·
Thanks for your response! Just let me know If you need anything else from me
Daniel
·
There is also another option in the docs by using a callback (instead of LangChainInstrumentor):
tracer = OpenInferenceTracer() response = chain.run(query, callbacks=[tracer])
This way it works, so maybe the way by using LangChainInstrumentor(tracer).instrument() is the issue and may be not sufficient?
Roger Y.
·
to step back a bit, were you able to get the spans from get_spans_dataframe? If so, then it should be working as intended
Roger Y.
·
the callbacks= is generally not preferred because it doesn’t trace child spans (this is a limitation of the langchain callback system)
Roger Y.
·
i was not able to replicate the empty output issue. i ran a script and it was able to print out the traces
Daniel
·
Thanks for your efforts. Will try to run the debugger to see what's going on
Roger Y.
·
On the other hand, we are in the process of removing the get_spans() method in a future version. We’ll keep you updated on how to migrate to a new method. Mikyo
✅1
Mikyo
·
Daniel, as Roger mentioned we have a persistence story we are working on as we speak, you just caught us in the middle of making it work end-to-end. The safest thing to do right now is to save data from the running phoenix intance
trace_dataset = px.active_session().get_trace_dataset()
if you call
trace_dataset.save()
It will actually save all the traces and the evals in your PHOENIX_WORKING_DIR
Mikyo
·
you can then actually resume by doing
tds = TraceDataset.load(id)
and resume using
px.launch_app(trace=tds)
Mikyo
·
however we understand that this is really a lot of steps you have to follow right now 😅 we're working on making this simple. For now if you save the trace dataframes, you can always resume on those.
px.launch_app(trace=px.TraceDataset(dataframe=df))
Daniel
·
Thanks for your help. I tried what Mikyo suggested but it seems that px.active_session() returns NoneType The app listens on port 6006 (I tried to launch the app once from a jupyter notebook px.launch_app() and also from the command line python -m phoenix.server.main serve - and I can see the traces traced from the script in the browser), but when I try to access the active session from the script I always get the same result 'None'. When I try to launch the app as part of my script I get an AttributeError. main.py
tracer = OpenInferenceTracer(exporter=HttpExporter(endpoint="http://0.0.0.0:6006")) LangChainInstrumentor(tracer).instrument() chain.invoke(...) # px.launch_app() -> throws an AttributeError print(px.active_session()) # returns 'None' even though I can access the session via notebook <phoenix.session.session.ThreadSession at 0x10965ee10> trace_dataset = px.active_session().get_trace_dataset() trace_dataset.save()
Mikyo
·
Daniel I misread your message and I didn't quite realize that you were running phoenix in a separate process. My apologies. You are right that will not work but it will work very soon via Roger Y.’s PR https://github.com/Arize-ai/phoenix/pull/2154 After this PR you will be able to do the commands I pasted but with px.Client() Thanks for highlighting the confusing flow. I've filed an enhancement request to make it a bit smoother. I agree since you do have access to the instance it does feel like you have access to the session https://github.com/Arize-ai/phoenix/issues/2175
Daniel
·
Mikyo Awesome, this is fantastic news! Can you give a rough estimation when this feature will be available?
Mikyo
·
Daniel Monday we'll cut a release. Def. look out for it in 🔒[private feed]. Thanks for your patience!
🫶1