Auto instrumentation guide

In this page we cover how to set up auto instrumentation of LLM calls using OpenTelemetry. This is the recommended approach for most users as it requires minimal code changes and provides rich trace data out of the box. If you need more control, you can also use manual instrumentation with the OpenTelemetry SDK.

Auto instrumentation with OpenAI SDK

If you are using the OpenAI Python SDK, the official OpenTelemetry distro can instrument all LLM calls with zero code changes.

1. Install dependencies

pip install opentelemetry-distro opentelemetry-exporter-otlp
opentelemetry-bootstrap -a install

2. Set environment variables

export OTEL_EXPORTER_OTLP_ENDPOINT=http://localhost:4318
export OTEL_SERVICE_NAME=my-agent
export OTEL_INSTRUMENTATION_GENAI_CAPTURE_MESSAGE_CONTENT=true
export OTEL_TRACES_EXPORTER=otlp
export OTEL_EXPORTER_OTLP_PROTOCOL=http/protobuf

3. Run your application with the auto-instrumentation wrapper

opentelemetry-instrument python my_agent.py

That is it. Every openai.chat.completions.create() call will emit gen_ai.* spans automatically.

Auto instrumentation with other SDKs

For broader provider support (Anthropic, Cohere, Mistral, Bedrock, VertexAI, and more), use OpenLLMetry (Traceloop SDK) or OpenLIT. These require only two lines of code added to your application entry point.

Option 1: OpenLLMetry (Traceloop SDK)

pip install traceloop-sdk

from traceloop.sdk import Traceloop

Traceloop.init()  # Call once at application startup

Option 2: OpenLIT

pip install openlit

import openlit

openlit.init()  # Call once at application startup

Both libraries auto-patch supported LLM client libraries and emit standard gen_ai.* OTel spans. Configure the OTLP exporter endpoint using the same environment variables shown in Path A.

Verify data ingestion

Once your application is running, verify that trace data is arriving in Parseable:

SELECT COUNT(*) AS span_count,
       MIN(p_timestamp) AS first_seen,
       MAX(p_timestamp) AS last_seen
FROM "genai-traces"
WHERE p_timestamp > NOW() - INTERVAL '1 hour';

If span_count is greater than zero, your LLM call instrumentation is working.

Next steps

Explore the Schema Reference and SQL Query Templates to start querying your data.