Pandas Query Engine

This guide shows you how to use our PandasQueryEngine: convert natural language to Pandas python code using LLMs.

The input to the PandasQueryEngine is a Pandas dataframe, and the output is a response. The LLM infers dataframe operations to perform in order to retrieve the result.

WARNING: This tool provides the LLM access to the eval function. Arbitrary code execution is possible on the machine running this tool. While some level of filtering is done on code, this tool is not recommended to be used in a production setting without heavy sandboxing or virtual machines.

If you’re opening this Notebook on colab, you will probably need to install LlamaIndex 🦙.

!pip install llama-index llama-index-experimental

import logging
import sys
from IPython.display import Markdown, display

import pandas as pd
from llama_index.experimental.query_engine import PandasQueryEngine


logging.basicConfig(stream=sys.stdout, level=logging.INFO)
logging.getLogger().addHandler(logging.StreamHandler(stream=sys.stdout))

Let’s start on a Toy DataFrame

Here let’s load a very simple dataframe containing city and population pairs, and run the PandasQueryEngine on it.

By setting verbose=True we can see the intermediate generated instructions.

# Test on some sample data
df = pd.DataFrame(
    {
        "city": ["Toronto", "Tokyo", "Berlin"],
        "population": [2930000, 13960000, 3645000],
    }
)

query_engine = PandasQueryEngine(df=df, verbose=True)

response = query_engine.query(
    "What is the city with the highest population?",
)

INFO:httpx:HTTP Request: POST https://api.openai.com/v1/chat/completions "HTTP/1.1 200 OK"
HTTP Request: POST https://api.openai.com/v1/chat/completions "HTTP/1.1 200 OK"
> Pandas Instructions:
```
df['city'][df['population'].idxmax()]
```
> Pandas Output: Tokyo

display(Markdown(f"<b>{response}</b>"))

Tokyo

# get pandas python instructions
print(response.metadata["pandas_instruction_str"])

df['city'][df['population'].idxmax()]

We can also take the step of using an LLM to synthesize a response.

query_engine = PandasQueryEngine(df=df, verbose=True, synthesize_response=True)
response = query_engine.query(
    "What is the city with the highest population? Give both the city and population",
)
print(str(response))

INFO:httpx:HTTP Request: POST https://api.openai.com/v1/chat/completions "HTTP/1.1 200 OK"
HTTP Request: POST https://api.openai.com/v1/chat/completions "HTTP/1.1 200 OK"
> Pandas Instructions:
```
df.loc[df['population'].idxmax()]
```
> Pandas Output: city             Tokyo
population    13960000
Name: 1, dtype: object
INFO:httpx:HTTP Request: POST https://api.openai.com/v1/chat/completions "HTTP/1.1 200 OK"
HTTP Request: POST https://api.openai.com/v1/chat/completions "HTTP/1.1 200 OK"
The city with the highest population is Tokyo, with a population of 13,960,000.

Analyzing the Titanic Dataset

The Titanic dataset is one of the most popular tabular datasets in introductory machine learning Source: https://www.kaggle.com/c/titanic

Download Data

!wget 'https://raw.githubusercontent.com/jerryjliu/llama_index/main/docs/docs/examples/data/csv/titanic_train.csv' -O 'titanic_train.csv'

--2024-01-13 17:45:15--  https://raw.githubusercontent.com/jerryjliu/llama_index/main/docs/docs/examples/data/csv/titanic_train.csv
Resolving raw.githubusercontent.com (raw.githubusercontent.com)... 2606:50c0:8003::154, 2606:50c0:8002::154, 2606:50c0:8001::154, ...
Connecting to raw.githubusercontent.com (raw.githubusercontent.com)|2606:50c0:8003::154|:443... connected.
HTTP request sent, awaiting response... 200 OK
Length: 57726 (56K) [text/plain]
Saving to: ‘titanic_train.csv’

titanic_train.csv   100%[===================>]  56.37K  --.-KB/s    in 0.009s

2024-01-13 17:45:15 (6.45 MB/s) - ‘titanic_train.csv’ saved [57726/57726]

df = pd.read_csv("./titanic_train.csv")

query_engine = PandasQueryEngine(df=df, verbose=True)

response = query_engine.query(
    "What is the correlation between survival and age?",
)

INFO:httpx:HTTP Request: POST https://api.openai.com/v1/chat/completions "HTTP/1.1 200 OK"
HTTP Request: POST https://api.openai.com/v1/chat/completions "HTTP/1.1 200 OK"
> Pandas Instructions:
```
df['survived'].corr(df['age'])
```
> Pandas Output: -0.07722109457217755

display(Markdown(f"<b>{response}</b>"))

-0.07722109457217755

# get pandas python instructions
print(response.metadata["pandas_instruction_str"])

df['survived'].corr(df['age'])

Additional Steps

Analyzing / Modifying prompts

Let’s look at the prompts!

from llama_index.core import PromptTemplate

query_engine = PandasQueryEngine(df=df, verbose=True)
prompts = query_engine.get_prompts()
print(prompts["pandas_prompt"].template)

You are working with a pandas dataframe in Python.
The name of the dataframe is `df`.
This is the result of `print(df.head())`:
{df_str}

Follow these instructions:
{instruction_str}
Query: {query_str}

Expression:

print(prompts["response_synthesis_prompt"].template)

Given an input question, synthesize a response from the query results.
Query: {query_str}

Pandas Instructions (optional):
{pandas_instructions}

Pandas Output: {pandas_output}

Response:

You can update prompts as well:

new_prompt = PromptTemplate(
    """\
You are working with a pandas dataframe in Python.
The name of the dataframe is `df`.
This is the result of `print(df.head())`:
{df_str}

Follow these instructions:
{instruction_str}
Query: {query_str}

Expression: """
)

query_engine.update_prompts({"pandas_prompt": new_prompt})

This is the instruction string (that you can customize by passing in instruction_str on initialization)

instruction_str = """\
1. Convert the query to executable Python code using Pandas.
2. The final line of code should be a Python expression that can be called with the `eval()` function.
3. The code should represent a solution to the query.
4. PRINT ONLY THE EXPRESSION.
5. Do not quote the expression.
"""

Implementing Query Engine using Query Pipeline Syntax

If you want to learn to construct your own Pandas Query Engine using our Query Pipeline syntax and the prompt components above, check out our below tutorial.

Setting up a Pandas DataFrame query engine with Query Pipelines