Filtering HTTP Request Traces in OpenAI Agent SDK with FastAPI

Filtering HTTP Request Traces in OpenAI Agent SDK with FastAPI | Arize AI Community

RunLLM
APP
·
Yes, you can filter out or disable HTTP request traces by customizing your OpenTelemetry setup. Use a custom SpanProcessor to filter out spans (e.g., those with names like "HTTP POST" or "HTTP GET") before they are sent to Phoenix. This is done by subclassing BatchSpanProcessor and overriding its on_start/on_end methods to skip unwanted spans.Example: LBKSfrom phoenix.otel import register, BatchSpanProcessor from opentelemetry.context import Context from opentelemetry.sdk.trace import ReadableSpan, SpanLBKSclass FilteringSpanProcessor(BatchSpanProcessor): def _filter_condition(self, span: Span) -> bool: # Filter out HTTP request spans by name or attributes return span.name.startswith("HTTP")LBKS def on_start(self, span: Span, parent_context: Context) -> None: if self._filter_condition(span): return super().on_start(span, parent_context)LBKS def on_end(self, span: ReadableSpan) -> None: if self._filter_condition(span): return super().on_end(span)LBKStracer_provider = register( project_name="openai-agents-test", endpoint="http://localhost:6006/v1/traces", auto_instrument=True, ) tracer_provider.add_span_processor(FilteringSpanProcessor( endpoint="http://localhost:6006/v1/traces", protocol="http/protobuf", ))LBKSThis will prevent HTTP request spans from being exported to Phoenix. Would you like more detail on customizing the filter or other approaches?Sources: - Phoenix Docs: Modifying Spans - FilteringSpanProcessor Example - Phoenix Docs: Filter Spans to Export
💡 Hint: Mention RunLLM in the thread for followups.
Daniel
·
Using the code above doesnt solve the issue, the HTTP/POST requests are still visible

can you try setting set_global_tracer_provider=False?

tracer_provider = register(
    project_name="openai-agents-test",
    endpoint="http://localhost:6006/v1/traces",
    auto_instrument=True,
    set_global_tracer_provider=False,
)

Daniel
·
Roger Y. - when I do that, nothing gets traced

Daniel

🔒[private user] custom is a custom wrapper around FastAPI that logs traces HTTP requests using OpenTel

from .server import init_server
from phoenix.otel import register

tracer_provider = register(
    project_name="openai-agents-test",
    endpoint="http://localhost:6006/v1/traces",
    auto_instrument=True)

custom = init_server()
custom.run()

Here is what I see

Roger Y.

not sure why that didn’t work, but for the time being you can try replacing register with raw otel as shown below

from opentelemetry.exporter.otlp.proto.http.trace_exporter import OTLPSpanExporter
from opentelemetry.sdk.resources import Resource
from opentelemetry.sdk.trace import TracerProvider
from opentelemetry.sdk.trace.export import SimpleSpanProcessor

from openinference.instrumentation.openai_agents import OpenAIAgentsInstrumentor

endpoint = "http://127.0.0.1:6006/v1/traces"
tracer_provider = TracerProvider(resource=Resource({"openinference.project.name": "openai-agents-test"}))
tracer_provider.add_span_processor(SimpleSpanProcessor(OTLPSpanExporter(endpoint)))

OpenAIAgentsInstrumentor().instrument(tracer_provider=tracer_provider)

Daniel
·
Sorry! It worked Roger Y. - I was looking at spans instead of traces! With the setup you suggested it worked
🙌1
Daniel
·
If I would like to implement it without relying on phoenix functionality than the code you suggested is equivalent ?
Daniel
·
i.e. to do it on opentel level
Roger Y.
·
yea, it’s equivalent
👀1
✅1
Daniel
·
Thank you very much Roger - you saved me a lot of time - appreciate it! 💪
Roger Y.
·
you’re welcome. let us know if you have other questions. thanks for using phoenix!
Daniel
·
I am curious if that approach with openinference and opentelemetry could work with other providers such as langfuse - is there openinference slack channel I can use?
Roger Y.
·
We don’t have a separate channel for openinference: you can just use this same one. As for other providers, I haven’t tested it myself, but with OpenInference being open standards, it should technically work: we’re open by design, so flexibility is built in.
✅1
Daniel
·
Oki, thanks Despite Phoenix lacking some features than langfuse has, we might stick with phoenix because of the amazing support

16 comments