Installation

pip install splunk-otel-instrumentation-bedrock

Usage

import boto3
from opentelemetry.instrumentation.bedrock import BedrockInstrumentor

BedrockInstrumentor().instrument()

client = boto3.client("bedrock-runtime", region_name="us-west-2")
response = client.converse(
    modelId="anthropic.claude-3-haiku-20240307-v1:0",
    messages=[{"role": "user", "content": [{"text": "Hello"}]}],
)

Composition With AgentCore

Use this package with AgentCore instrumentation when your application uses BedrockAgentCoreApp and calls Bedrock Runtime from inside the entrypoint. AgentCore owns the agent parent spans, and this package adds the child LLM spans that evaluation callbacks consume.

from opentelemetry.instrumentation.bedrock import BedrockInstrumentor
from opentelemetry.instrumentation.bedrock_agentcore import (
    BedrockAgentCoreInstrumentor,
)

BedrockAgentCoreInstrumentor().instrument()
BedrockInstrumentor().instrument()

Example

In the repository, see examples/manual for a runnable Bedrock Runtime example that can also enable AgentCore instrumentation. It uses console span export by default so you can verify that Bedrock Runtime LLM spans nest under active AgentCore spans when both instrumentors are enabled.

What Gets Instrumented

• bedrock-runtime.Converse -> LLMInvocation

• bedrock-runtime.ConverseStream -> streaming LLMInvocation

• bedrock-runtime.InvokeModel -> provider-aware LLMInvocation

• bedrock-runtime.InvokeModelWithResponseStream -> provider-aware streaming LLMInvocation for supported streamed JSON chunk formats

Agent Runtime calls such as bedrock-agent-runtime.invoke_agent are not instrumented by this package. Agent orchestration spans belong in AgentCore or agent-framework instrumentation.

Configuration

Content capture follows the shared GenAI environment variables:

export OTEL_INSTRUMENTATION_GENAI_CAPTURE_MESSAGE_CONTENT=SPAN_AND_EVENT
export OTEL_INSTRUMENTATION_GENAI_CAPTURE_TOOL_DEFINITIONS=true

Useful flags for Bedrock Runtime and AgentCore composition:

The instrumentation always populates message bodies and tool arguments on the Python invocation objects so evaluations can consume them. The shared GenAI emitters use the content-capture setting to decide whether those values are emitted as span attributes or log events.

For zero-code instrumentation, disable this package with OTEL_PYTHON_DISABLED_INSTRUMENTATIONS=bedrock.

InvokeModel Coverage

InvokeModel and InvokeModelWithResponseStream support generic JSON metadata extraction plus provider-specific shapes aligned with upstream OpenTelemetry botocore Bedrock behavior:

• Amazon Titan

• Amazon Nova

• Anthropic Claude

• Cohere Command and Command R

• Meta Llama

• Mistral

For non-streaming InvokeModel responses, the instrumentation reads botocore.response.StreamingBody only to parse supported JSON response shapes, then replaces the response body with a fresh stream so application code can still read it.

Telemetry Details

The instrumentation sets:

• gen_ai.system = aws.bedrock

• gen_ai.framework = boto3

• gen_ai.request.model from modelId

• gen_ai.provider.name = aws.bedrock

• request params such as temperature, top-p, max tokens, and stop sequences

• response ID, response model, finish reasons, and token usage when available

• gen_ai.request.stream and gen_ai.response.time_to_first_chunk for streaming calls

Requirements

• Python >= 3.10

• boto3 >= 1.34.0

• splunk-otel-util-genai >= 0.1.9