OpenInference Portkey AI Instrumentation

0.1.8 verified Fri May 01 auth: no python

OpenInference instrumentation for Portkey AI gateway, enabling tracing of LLM calls made via Portkey. Version 0.1.8 supports Python 3.10-3.14. Part of the Arize AI OpenInference ecosystem for observability of AI applications.

pip install openinference-instrumentation-portkey

Common errors

error ImportError: cannot import name 'PortkeyInstrumentor' from 'openinference_instrumentation_portkey' ↓

cause The package may not be installed correctly, or you are using an older version that had a different export.

fix

Verify installation: pip install openinference-instrumentation-portkey. Check your Python environment and reinstall if needed.

error AttributeError: module 'openinference_instrumentation_portkey' has no attribute 'PortkeyInstrumentor' ↓

cause Similar to above; the module exists but the instrumentor class is not exported (possible version mismatch).

fix

Ensure you have the latest version: pip install --upgrade openinference-instrumentation-portkey

Warnings

gotcha Instrumentation must be called before any Portkey client is created. If you import or instantiate Portkey before calling .instrument(), traces will not be captured. ↓

fix Ensure the .instrument() call happens at module-level before any Portkey usage, ideally right after setting up OpenTelemetry.

gotcha If no TracerProvider is set, instrumentation silently fails (no spans emitted). Ensure opentelemetry-api and a TracerProvider (e.g., from opentelemetry-sdk) are configured. ↓

fix Set a global tracer provider via trace_api.set_tracer_provider(TracerProvider()) and add a span processor with an exporter.

gotcha This package requires wrapt. If you have an older version of wrapt or wrapt is missing, you may see import errors. Versions >=1.14 are recommended. ↓

fix Install or upgrade wrapt: pip install 'wrapt>=1.14'

Imports

PortkeyInstrumentor

from openinference_instrumentation_portkey import PortkeyInstrumentor

Correct import for the instrumentor class

Quickstart

Initialize tracing by setting up an OpenTelemetry TracerProvider and calling PortkeyInstrumentor().instrument() before any Portkey usage.

import os
from openinference_instrumentation_portkey import PortkeyInstrumentor
from opentelemetry import trace as trace_api

# Set up a TracerProvider (e.g., ConsoleSpanExporter for debugging)
from opentelemetry.exporter.otlp.proto.http.trace_exporter import OTLPSpanExporter
from opentelemetry.sdk.trace import TracerProvider
from opentelemetry.sdk.trace.export import SimpleSpanProcessor

trace_provider = TracerProvider()
trace_provider.add_span_processor(
    SimpleSpanProcessor(OTLPSpanExporter(endpoint=os.environ.get("OTEL_ENDPOINT", "http://localhost:4318/v1/traces")))
)
trace_api.set_tracer_provider(trace_provider)

# Instrument Portkey
PortkeyInstrumentor().instrument()

# Now use Portkey normally - calls will be traced
from portkey_ai import Portkey

client = Portkey(api_key=os.environ.get("PORTKEY_API_KEY", ""))
response = client.completions.create(prompt="Hello", model="gpt-3.5-turbo")
print(response.choices[0].text)