openinference-instrumentation-bedrock 0.1.8

OpenInference Bedrock Instrumentation

OpenInference AWS Bedrock Instrumentation

Python autoinstrumentation library for AWS Bedrock calls made using boto3.

This package implements OpenInference tracing for invoke_model and converse calls made using a boto3 bedrock-runtime client. These traces are fully OpenTelemetry compatible and can be sent to an OpenTelemetry collector for viewing, such as Arize phoenix.

[!NOTE]
The Converse API was introduced in botocore v1.34.116. Please use v1.34.116 or above to utilize converse.

Supported Models

Find the list of Bedrock-supported models and their IDs here. Future testing is planned for additional models.

Model	Supported Methods
Anthropic Claude 2.0	converse, invoke
Anthropic Claude 2.1	converse, invoke
Anthropic Claude 3 Sonnet 1.0	converse
Anthropic Claude 3.5 Sonnet	converse
Anthropic Claude 3 Haiku	converse
Meta Llama 3 8b Instruct	converse
Meta Llama 3 70b Instruct	converse
Mistral AI Mistral 7B Instruct	converse
Mistral AI Mixtral 8X7B Instruct	converse
Mistral AI Mistral Large	converse
Mistral AI Mistral Small	converse

Installation

pip install openinference-instrumentation-bedrock

Quickstart

[!IMPORTANT]
OpenInference for AWS Bedrock supports both invoke_model and converse. For models that use the Messages API, such as Anthropic Claude 3 and Anthropic Claude 3.5, use the Converse API instead.

In a notebook environment (jupyter, colab, etc.) install openinference-instrumentation-bedrock, arize-phoenix and boto3.

You can test out this quickstart guide in Google Colab!

pip install openinference-instrumentation-bedrock arize-phoenix boto3

Ensure that boto3 is configured with AWS credentials.

First, import dependencies required to autoinstrument AWS Bedrock and set up phoenix as an collector for OpenInference traces.

from urllib.parse import urljoin

import boto3
import phoenix as px

from openinference.instrumentation.bedrock import BedrockInstrumentor
from opentelemetry import trace as trace_api
from opentelemetry.exporter.otlp.proto.http.trace_exporter import OTLPSpanExporter
from opentelemetry.sdk import trace as trace_sdk
from opentelemetry.sdk.trace.export import SimpleSpanProcessor

Next, we'll start a phoenix server and set it as a collector.

px.launch_app()
session_url = px.active_session().url
phoenix_otlp_endpoint = urljoin(session_url, "v1/traces")
phoenix_exporter = OTLPSpanExporter(endpoint=phoenix_otlp_endpoint)
tracer_provider = trace_sdk.TracerProvider()
tracer_provider.add_span_processor(SimpleSpanProcessor(span_exporter=phoenix_exporter))
trace_api.set_tracer_provider(tracer_provider=tracer_provider)

Instrumenting boto3 is simple:

BedrockInstrumentor().instrument()

Now, all calls to invoke_model are instrumented and can be viewed in the phoenix UI.

session = boto3.session.Session()
client = session.client("bedrock-runtime")
prompt = b'{"prompt": "Human: Hello there, how are you? Assistant:", "max_tokens_to_sample": 1024}'
response = client.invoke_model(modelId="anthropic.claude-v2", body=prompt)
response_body = json.loads(response.get("body").read())
print(response_body["completion"])

Alternatively, all calls to converse are instrumented and can be viewed in the phoenix UI.

session = boto3.session.Session()
client = session.client("bedrock-runtime")

message1 = {
            "role": "user",
            "content": [{"text": "Create a list of 3 pop songs."}]
}
message2 = {
        "role": "user",
        "content": [{"text": "Make sure the songs are by artists from the United Kingdom."}]
}
messages = []

messages.append(message1)
response = client.converse(
    modelId="anthropic.claude-3-5-sonnet-20240620-v1:0",
    messages=messages
)
out = response["output"]["message"]
messages.append(out)
print(out.get("content")[-1].get("text"))

messages.append(message2)
response = client.converse(
    modelId="anthropic.claude-v2:1",
    messages=messages
)
out = response['output']['message']
print(out.get("content")[-1].get("text"))

More Info

More details about tracing with OpenInference and phoenix can be found in the phoenix documentation.