Quickstart¶

Installation¶

To install TableGPT Agent, use the following command:

In [ ]:

%pip install tablegpt-agent

%pip install tablegpt-agent

TableGPT Agent depends on pybox to manage code execution environment. By default, pybox operates in an in-cluster mode. If you intend to run tablegpt-agent in a local environment, install the optional dependency as follows:

In [1]:

%pip install tablegpt-agent[local]

%pip install tablegpt-agent[local]

This tutorial uses langchain-openai for the chat model instance. Please make sure you have it installed:

In [ ]:

%pip install langchain-openai

%pip install langchain-openai

Setup the LLM Service¶

Before using TableGPT Agent, ensure you have an OpenAI-compatible server configured to host TableGPT2. We recommend using vllm for this:

python -m vllm.entrypoints.openai.api_server --served-model-name TableGPT2-7B --model path/to/weights

NOTES:

• To analyze tabular data with tablegpt-agent, make sure TableGPT2 is served with vllm version 0.5.5 or higher.
• For production environments, it's important to optimize the vllm server configuration. For details, refer to the vllm documentation on server configuration.

Create TableGPT Agent¶

NOTE: TableGPT Agent fully supports aync invocation. If you are running this tutorial in a Jupyter Notebook, no additional setup is required. However, if you plan to run the tutorial in a Python console, make sure to use a console that supports asynchronous operations. To get started, execute the following command:

python -m asyncio

In the console or notebook, create the agent as follows:

In [ ]:

from langchain_openai import ChatOpenAI
from pybox import AsyncLocalPyBoxManager
from tablegpt.agent import create_tablegpt_graph


llm = ChatOpenAI(openai_api_base="YOUR_VLLM_URL", openai_api_key="whatever", model_name="TableGPT2-7B")
pybox_manager = AsyncLocalPyBoxManager()

agent = create_tablegpt_graph(
    llm=llm,
    pybox_manager=pybox_manager,
)

from langchain_openai import ChatOpenAI from pybox import AsyncLocalPyBoxManager from tablegpt.agent import create_tablegpt_graph llm = ChatOpenAI(openai_api_base="YOUR_VLLM_URL", openai_api_key="whatever", model_name="TableGPT2-7B") pybox_manager = AsyncLocalPyBoxManager() agent = create_tablegpt_graph( llm=llm, pybox_manager=pybox_manager, )

Start Chatting¶

In [3]:

from datetime import date
from langchain_core.messages import HumanMessage

message = HumanMessage(content="Hi")

_input = {
    "messages": [message],
    "parent_id": "some-parent-id",
    "date": date.today(),
}

state = await agent.ainvoke(_input)
state["messages"]

from datetime import date from langchain_core.messages import HumanMessage message = HumanMessage(content="Hi") _input = { "messages": [message], "parent_id": "some-parent-id", "date": date.today(), } state = await agent.ainvoke(_input) state["messages"]

[HumanMessage(content='Hi', additional_kwargs={}, response_metadata={}, id='34fe748c-81ab-49ea-bec6-9c621598a61a'), AIMessage(content="Hello! How can I assist you with data analysis today? Please let me know the details of the dataset you're working with and what specific analysis you'd like to perform.", additional_kwargs={'parent_id': 'some-parent-id'}, response_metadata={}, id='a1ee29d2-723e-41c7-b420-27d0cfaed5dc')]

You can get more detailed outputs with the astream_events method:

In [4]:

async for event in agent.astream_events(
    input=_input,
    version="v2",
):
    # We ignore irrelevant events here.
    if event["event"] == "on_chat_model_end":
        print(event["data"]["output"])

async for event in agent.astream_events( input=_input, version="v2", ): # We ignore irrelevant events here. if event["event"] == "on_chat_model_end": print(event["data"]["output"])

content='Hello! How can I assist you with your data analysis today? Please let me know what dataset you are working with and what specific analyses or visualizations you would like to perform.' additional_kwargs={} response_metadata={'finish_reason': 'stop', 'model_name': 'TableGPT2-7B'} id='run-525eb149-0e3f-4b04-868b-708295f789ac'