Pandasai

PandasAI is the smallest possible version of the idea behind every AI data tool: put a .chat() method on a dataframe. You load a CSV, configure a model, ask a question in English, and the library writes Python, runs it, and hands back the answer. Nothing is hosted, nothing is a platform, and the whole thing fits in a notebook cell.

That makes it a good fit for analysts who already work in Python and want conversational analysis inside a script or notebook rather than in a vendor's web app. It also makes the security question yours to answer, because the code that runs is code a language model just generated.

Highlights

• Chat over one dataframe or several. df.chat("What is the average revenue by region?") returns a value; pai.chat(question, df_a, df_b) asks across multiple frames and relates them. Ask for a plot and you get one.
• An experimental semantic layer. pai.create() saves a dataset with a path, a description, and typed, described columns, so the model reads business meaning rather than guessing from headers. PandasAI marks the semantic layer as experimental and aimed at advanced users.
• Data beyond files. CSV and Excel load directly; a sql extension covers SQL, PostgreSQL, MySQL, CockroachDB, and Microsoft SQL Server. Snowflake, Databricks, and the other cloud connectors sit in the enterprise-licensed extensions.
• Model-agnostic. Install pandasai-litellm for a unified interface to OpenAI, Anthropic, Google and others, or pandasai-openai for OpenAI and Azure OpenAI directly, then set it once with pai.config.set().
• A real sandbox, not a promise. pandasai-docker runs generated code in an isolated Docker container that operates entirely offline, with strict resource limits and filesystem isolation.
• An Agent for multi-turn work. pai.chat() is for single-session exploration; the Agent class keeps conversation state so follow-up questions like "and which one has the most deals?" resolve correctly.

In an analyst's workflow

The pattern worth adopting is to describe your columns once and then never explain them again:

import pandasai as pai
from pandasai_litellm.litellm import LiteLLM
 
pai.config.set({"llm": LiteLLM(model="gpt-4.1-mini", api_key="...")})
 
df = pai.create(
    path="company/sales-data",
    df=pai.read_csv("data/sales.csv"),
    description="Sales data from our retail stores",
    columns={
        "sale_date": {"type": "datetime", "description": "Date and time of the sale"},
        "quantity":  {"type": "integer",  "description": "Number of units sold"},
        "price":     {"type": "float",    "description": "Price per unit, pre-discount"},
    },
)
 
df.chat("Which product had the highest revenue last quarter?")

Those column descriptions do most of the accuracy work, and they are cheaper to maintain than a prompt. Read the code it generates before you trust a number, the same discipline the checking an AI data analysis guide applies to any generated result.

Good to know

PandasAI is published by Sinaptik GmbH. Version 3.0.0 is the current release on PyPI, uploaded on October 7, 2025, and requires Python 3.8 or later and below 3.12 — a narrow window worth checking against your environment. Development on the public repository has been quiet since late 2025, so treat it as a stable library rather than a fast-moving one, and pin your version.

For SQL-first work against a warehouse, Vanna is the closer open-source comparison, and the best text-to-SQL tools in 2026 sets both against the hosted options. If you want the same conversational experience without maintaining any of it, Julius is the hosted equivalent, and Claude for data analysis covers doing it with a general assistant. The best AI tools for data analysts in 2026 roundup has the full comparison.