Set up LLM

PandaAI supports multiple LLMs. To make the library lightweight, the default LLM is BambooLLM, developed by PandaAI team themselves. To use other LLMs, you need to install the corresponding llm extension. Once a LLM extension is installed, you can configure it simply using pai.config.set(). Then, every time you use the .chat() method, it will use the configured LLM.

​

BambooLLM

BambooLLM is the default LLM for PandaAI, fine-tuned for data analysis. You can get your free API key by signing up at app.pandabi.ai.

import pandasai as pai

pai.api_key.set("api-key")

​

OpenAI models

Install the pandasai-openai extension:

# Using poetry
poetry add pandasai-openai

# Using pip
pip install pandasai-openai

In order to use OpenAI models, you need to have an OpenAI API key. You can get one here. Once you have an API key, you can use it to instantiate an OpenAI object:

Configure OpenAI:

import pandasai as pai
from pandasai_openai import OpenAI

llm = OpenAI(api_token="my-openai-api-key")

# Set your OpenAI API key
pai.config.set({"llm": llm})

​

Azure OpenAI models

Install the pandasai-openai extension:

# Using poetry
poetry add pandasai-openai

# Using pip
pip install pandasai-openai

In order to use Azure OpenAI models, you need to have an Azure OpenAI API key. You can get one here. Once you have an API key, you can use it to instantiate an Azure OpenAI object:

Configure Azure OpenAI:

import pandasai as pai
from pandasai_openai import AzureOpenAI

llm = AzureOpenAI(api_base="https://<your-endpoint>.openai.azure.com/",
    api_key="my-azure-openai-api-key",
    deployment_name="text-davinci-003")  # The name of your deployed model

pai.config.set({"llm": llm})

​

How to set up any LLM?

LiteLLM provides a unified interface to interact with 100+ LLM models from various providers including OpenAI, Azure, Anthropic, Google, AWS, Hugging Face, and many more. This makes it easy to switch between different LLM providers without changing your code.

Install the pandasai-litellm extension:

# Using poetry
poetry add pandasai-litellm

# Using pip
pip install pandasai-litellm

Configure LiteLLM with your chosen model. First, set up your API keys as environment variables:

import os
import pandasai as pai
from pandasai_litellm import LiteLLM

# Set your API keys as environment variables
os.environ["OPENAI_API_KEY"] = "your-openai-api-key"
os.environ["ANTHROPIC_API_KEY"] = "your-anthropic-api-key"

# Example with OpenAI
llm = LiteLLM(model="gpt-3.5-turbo")

# Example with Anthropic
llm = LiteLLM(model="claude-2")

# Set your LLM configuration
pai.config.set({"llm": llm})

LiteLLM supports a wide range of models from various providers, including but not limited to:

• OpenAI (gpt-3.5-turbo, gpt-4, etc.)
• Anthropic (claude-2, claude-instant-1, etc.)
• Google (gemini-pro, palm2, etc.)
• Azure OpenAI
• AWS (Bedrock, SageMaker)
• Mistral AI
• Cohere
• Hugging Face

For a complete list of supported models and providers, visit the LiteLLM documentation.

​

Determinism

Determinism in language models refers to the ability to produce the same output consistently given the same input under identical conditions. This characteristic is vital for:

• Reproducibility: Ensuring the same results can be obtained across different runs, which is crucial for debugging and iterative development.
• Consistency: Maintaining uniformity in responses, particularly important in scenarios like automated customer support, where varied responses to the same query might be undesirable.
• Testing: Facilitating the evaluation and comparison of models or algorithms by providing a stable ground for testing.

​

The Role of temperature=0

The temperature parameter in language models controls the randomness of the output. A higher temperature increases diversity and creativity in responses, while a lower temperature makes the model more predictable and conservative. Setting temperature=0 essentially turns off randomness, leading the model to choose the most likely next word at each step. This is critical for achieving determinism as it minimizes variance in the model’s output.

​

Implications of temperature=0

• Predictable Responses: The model will consistently choose the most probable path, leading to high predictability in outputs.
• Creativity: The trade-off for predictability is reduced creativity and variation in responses, as the model won’t explore less likely options.

​

Utilizing seed for Enhanced Control

The seed parameter is another tool to enhance determinism. It sets the initial state for the random number generator used in the model, ensuring that the same sequence of “random” numbers is used for each run. This parameter, when combined with temperature=0, offers an even higher degree of predictability.

​

Example:

import pandasai as pai

# Sample DataFrame
df = pai.DataFrame({
    "country": ["United States", "United Kingdom", "France", "Germany", "Italy", "Spain", "Canada", "Australia", "Japan", "China"],
    "gdp": [19294482071552, 2891615567872, 2411255037952, 3435817336832, 1745433788416, 1181205135360, 1607402389504, 1490967855104, 4380756541440, 14631844184064],
    "happiness_index": [6.94, 7.16, 6.66, 7.07, 6.38, 6.4, 7.23, 7.22, 5.87, 5.12]
})

# Configure the LLM
pai.config.set("temperature", 0)
pai.config.set("seed", 26)

df.chat('Which are the 5 happiest countries?') # answer should me (mostly) consistent across devices.

​

Current Limitation:

​

AzureOpenAI Instance

While the seed parameter is effective with the OpenAI instance in our library, it’s important to note that this functionality is not yet available for AzureOpenAI. Users working with AzureOpenAI can still use temperature=0 to reduce randomness but without the added predictability that seed offers.

​

System fingerprint

As mentioned in the documentation (OpenAI Seed) :

Sometimes, determinism may be impacted due to necessary changes OpenAI makes to model configurations on our end. To help you keep track of these changes, we expose the system_fingerprint field. If this value is different, you may see different outputs due to changes we’ve made on our systems.

​

Workarounds and Future Updates

For AzureOpenAI Users: Rely on temperature=0 for reducing randomness. Stay tuned for future updates as we work towards integrating seed functionality with AzureOpenAI. For OpenAI Users: Utilize both temperature=0 and seed for maximum determinism.