Skip to main content

Arize integration guide

Arize AI provides a machine learning platform helps data scientists and machine learning engineers monitor, explain, and improve their models in production. You can integrate Arize with Anyscale Private Endpoints for enhanced LLM observability.

Step 0: Install dependencies

pip install arize[AutoEmbeddings]
pip install openai==1.3.2
pip install langchain>=0.0.341

Step 1: Set up Arize client

To log data from an app that uses Anyscale Private Endpoints, you need to set up an ArizeCallBackHandler to use. Create and account or log in to Arize, and retrieve your API_KEY and SPACE_KEY from the console to paste into the code below:

from langchain.callbacks.arize_callback import ArizeCallbackHandler


arize_callback = ArizeCallbackHandler(

Step 2: Initialize chat models with Anyscale Endpoints

To facilitate interaction with various models through Anyscale Private Endpoints, initialize chat models using the LangChain library. This step involves creating a list of initial messages to set the context and utilizing the ChatAnyscale class to prepare for communication with each available model.

For a quick demo, you can paste in your API base and key, but for development, follow best practices for setting your API base and key.

from langchain.chat_models import ChatAnyscale
from langchain.schema import SystemMessage, HumanMessage


# Messages for initiating the chat context
messages = [
content="You are a helpful AI that shares everything you know."
content="Come up with a HEX code that captures November."

# Retrieve available models and creating a ChatAnyscale instance for each one
chats = {
model: ChatAnyscale(
for model in ChatAnyscale.get_available_models(anyscale_api_base=ANYSCALE_BASE_URL, anyscale_api_key=ANYSCALE_API_KEY)

Step 3: Log predictions and monitor with Arize

After setting up the chat instances, send the prompt messages to the chat model endpoints and log the responses through Arize AI. The ArizeCallbackHandler captures each model's response and sends it to the Arize console. This allows you to monitor the model's performance and ensure the quality of predictions in production.

for model_name, endpoint in chats.items():
response = endpoint.predict_messages(messages, callbacks=[arize_callback])
print(model_name, "\n", response.content)