Skip to content

How to customize model usage

Supported models

Vizro-AI currently supports OpenAI models as follows, although we are working on supporting more vendors:

  • gpt-3.5-turbo (default model)
  • gpt-4-turbo (recommended for best model performance)
  • gpt-3.5-turbo-0125
  • gpt-3.5-turbo-1106
  • gpt-4-0613
  • gpt-4-1106-preview
  • gpt-3.5-turbo-0613 (to be deprecated on June 13, 2024)

These models offer different levels of performance and cost to Vizro-AI users:

  • The gpt-3.5 model series have lower price point and higher speeds for providing answers, but do not offer sophisticated charting.

  • Consider upgrading to the gpt-4 models for more demanding tasks. While they are part of a more capable GPT model series, their response time is slower than gpt-3.5 models, and they come at a higher cost.

Refer to the OpenAI documentation for more about model capabilities and pricing.

Default initialization

VizroAI can be initialized without any arguments, in which case it uses "gpt-3.5-turbo" by default, with a temperature setting of 0. "gpt-3.5-turbo" offers enhanced speed and accuracy, and generates responses in requested formats while maintaining cost-effective performance.

Customization at initialization

To customize the model, you can pass VizroAI a single argument named model, which can either be a string that specifies the name of a ChatOpenAI model or an instantiated ChatOpenAI model.

The example below uses the OpenAI model name in a string form:

Customize with string

1
2
3
from vizro_ai import VizroAI

vizro_ai = VizroAI(model="gpt-3.5-turbo-0125")

The example below customizes the ChatOpenAI instance further beyond the chosen default from the string instantiation. We pass the "gpt-3.5-turbo-0125" model from OpenAI as model_name for ChatOpenAI, which offers improved response accuracy, we also want to increase maximum number of retries. It's important to mention that any parameter that could be used in the openai.create call is also usable in ChatOpenAI. For more customization options for ChatOpenAI, refer to the LangChain ChatOpenAI docs

To ensure a deterministic answer to our queries, we've set the temperature to 0. If you prefer more creative (but potentially more unstable) responses, you can raise the temperature to a maximum of 1.

Customize with instantiated model

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
from vizro_ai import VizroAI
from langchain_openai import ChatOpenAI

import vizro.plotly.express as px
df = px.data.gapminder()

llm = ChatOpenAI(
    model_name="gpt-3.5-turbo-0125",
    temperature=0,
    max_retries=3,
)
vizro_ai = VizroAI(model=llm)
vizro_ai.plot(df, "describe the composition of gdp in continent")

Azure OpenAI models

To set up Azure OpenAI with VizroAI, you'll need to configure the AzureOpenAI instance by specifying your deployment name and model name using LangChain. You can also set your environment variables for API configuration, such as OPENAI_API_TYPE, OPENAI_API_VERSION, OPENAI_API_BASE and OPENAI_API_KEY. Authentication can be done via an API key directly or through Azure Active Directory (AAD) for enhanced security. For a detailed walk-through, refer to the LangChain documentation.

Here is an example of how to set the LLM model to be an AzureOpenAI model:

Use Azure OpenAI model

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
from langchain_openai import AzureOpenAI
from vizro_ai import VizroAI

# Create an instance of Azure OpenAI
# Replace the deployment name with your own
llm = AzureOpenAI(
    deployment_name="td2",
    model_name="gpt-3.5-turbo-instruct",
)
vizro_ai = VizroAI(model=llm)