ARMS provides OpenTelemetry Auto instrumentation for various LLM providers, frameworks, and VectorDBs, providing you with insights into the behavior and performance of your AI applications.

This documentation covers tracing settings, understanding semantic conventions, and interpreting span attributes to enhance the monitoring and observability of your LLM applications.

Using an existing OTel Tracer

You have the flexibility to integrate your existing OpenTelemetry (OTel) tracer configuration with ARMS. If you already have an OTel tracer instantiated in your application, you can pass it directly to elsai_arms.init(tracer=tracer). This integration ensures that ARMS utilizes your custom tracer settings, allowing for a unified tracing setup across your application.

Example:

# Instantiate an OpenTelemetry Tracer
tracer = ...

# Pass the tracer to ARMS
elsai_arms.init(tracer=tracer)

Add custom resource attributes

The OTEL_RESOURCE_ATTRIBUTES environment variable allows you to provide additional OpenTelemetry resource attributes when starting your application with ARMS. ARMS already includes some default resource attributes:

• telemetry.sdk.name: elsai-arms
• service.name: YOUR_SERVICE_NAME
• deployment.environment: YOUR_ENVIRONMENT_NAME

You can enhance these default resource attributes by adding your own using the OTEL_RESOURCE_ATTRIBUTES variable. Your custom attributes will be added on top of the existing ARMS attributes, providing additional context to your telemetry data. Simply format your attributes as key1=value1,key2=value2.

For example:

export OTEL_RESOURCE_ATTRIBUTES="service.instance.id=YOUR_SERVICE_ID,k8s.pod.name=K8S_POD_NAME,k8s.namespace.name=K8S_NAMESPACE,k8s.node.name=K8S_NODE_NAME"

Disable Tracing of Content

By default, ARMS adds the prompts and completions to Trace span attributes.

However, you may want to disable this logging for privacy reasons, as they may contain highly sensitive data from your users. You may also simply want to reduce the size of your traces.

Example:

elsai_arms.init(capture_message_content=False)

Limit Content Length

By default, ARMS captures full prompt and completion content without any truncation. If you want to limit the size of captured content (for example, to reduce trace storage costs or avoid excessively large spans), set max_content_length to a positive integer. Content exceeding the limit will be truncated with ... appended.

A value of None (the default), 0, or -1 disables truncation entirely.

Example:

# Truncate all captured content to 500 characters
elsai_arms.init(max_content_length=500)

You can also set this via the environment variable:

export ELSAI_ARMS_MAX_CONTENT_LENGTH=500

Disable Batch

By default, the SDK batches spans using the OpenTelemetry batch span processor. When working locally, sometimes you may wish to disable this behavior. You can do that with this flag.

Example:

elsai_arms.init(disable_batch=True)

Disable Instrumentations

By default, ARMS automatically detects which models and frameworks you are using and instruments them for you. You can override this and disable instrumentation for specific frameworks and models.

Example:

elsai_arms.init(disabled_instrumentors=["anthropic", "langchain"])

Manual Tracing

Using elsai_arms.trace, you get access to manually create traces, allowing you to record every process within a single function.

@elsai_arms.trace
def generate_one_liner():
    completion = client.chat.completions.create(
        model="gpt-3.5-turbo",
        messages=[
            {
                "role": "system",
                "content": "Return a one liner from any movie for me to guess",
            }
        ],
    )

The trace function automatically groups any LLM function invoked within generate_one_liner, providing you with organized groupings right out of the box.

You can do more with traces by running the start_trace context generator:

with elsai_arms.start_trace(name="<GIVE_TRACE_A_NAME>") as trace:
    # your code

Use trace.set_result('') to set the final result of the trace and trace.set_metadata({}) to add custom metadata.

Full Example

@elsai_arms.trace
def generate_one_liner():
    completion = client.chat.completions.create(
        model="gpt-3.5-turbo",
        messages=[
            {
                "role": "system",
                "content": "Return a one liner from any movie for me to guess",
            }
        ],
    )

def guess_one_liner(one_liner: str):
    with elsai_arms.start_trace("Guess One-liner") as trace:
        completion = client.chat.completions.create(
            model="gpt-3.5-turbo",
            messages=[
                {
                    "role": "user",
                    "content": f"Guess movie from this line: {one_liner}",
                }
            ],
        )
        trace.set_result(completion.choices[0].message.content)