DataFrame Structured Data Extraction

This demo shows how you can extract tabular DataFrames from raw text.

This was directly inspired by jxnl’s dataframe example here: https://github.com/jxnl/openai_function_call/blob/main/auto_dataframe.py.

We show this with different levels of complexity, all backed by the OpenAI Function API:

• (more code) How to build an extractor yourself using our OpenAIPydanticProgram

• (less code) Using our out-of-the-box DFFullProgram and DFRowsProgram objects

Build a DF Extractor Yourself (Using OpenAIPydanticProgram)

Our OpenAIPydanticProgram is a wrapper around an OpenAI LLM that supports function calling - it will return structured outputs in the form of a Pydantic object.

We import our DataFrame and DataFrameRowsOnly objects.

To create an output extractor, you just need to 1) specify the relevant Pydantic object, and 2) Add the right prompt

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

!pip install llama-index

from llama_index.program import (
    OpenAIPydanticProgram,
    DFFullProgram,
    DataFrame,
    DataFrameRowsOnly,
)
from llama_index.llms import OpenAI

program = OpenAIPydanticProgram.from_defaults(
    output_cls=DataFrame,
    llm=OpenAI(temperature=0, model="gpt-4-0613"),
    prompt_template_str=(
        "Please extract the following query into a structured data according"
        " to: {input_str}.Please extract both the set of column names and a"
        " set of rows."
    ),
    verbose=True,
)

# NOTE: the test example is taken from jxnl's repo

response_obj = program(
    input_str="""My name is John and I am 25 years old. I live in 
        New York and I like to play basketball. His name is 
        Mike and he is 30 years old. He lives in San Francisco 
        and he likes to play baseball. Sarah is 20 years old 
        and she lives in Los Angeles. She likes to play tennis.
        Her name is Mary and she is 35 years old. 
        She lives in Chicago."""
)
response_obj

Function call: DataFrame with args: {
  "columns": [
    {
      "column_name": "Name",
      "column_desc": "Name of the person"
    },
    {
      "column_name": "Age",
      "column_desc": "Age of the person"
    },
    {
      "column_name": "City",
      "column_desc": "City where the person lives"
    },
    {
      "column_name": "Hobby",
      "column_desc": "What the person likes to do"
    }
  ],
  "rows": [
    {
      "row_values": ["John", 25, "New York", "play basketball"]
    },
    {
      "row_values": ["Mike", 30, "San Francisco", "play baseball"]
    },
    {
      "row_values": ["Sarah", 20, "Los Angeles", "play tennis"]
    },
    {
      "row_values": ["Mary", 35, "Chicago", "play tennis"]
    }
  ]
}

DataFrame(description=None, columns=[DataFrameColumn(column_name='Name', column_desc='Name of the person'), DataFrameColumn(column_name='Age', column_desc='Age of the person'), DataFrameColumn(column_name='City', column_desc='City where the person lives'), DataFrameColumn(column_name='Hobby', column_desc='What the person likes to do')], rows=[DataFrameRow(row_values=['John', 25, 'New York', 'play basketball']), DataFrameRow(row_values=['Mike', 30, 'San Francisco', 'play baseball']), DataFrameRow(row_values=['Sarah', 20, 'Los Angeles', 'play tennis']), DataFrameRow(row_values=['Mary', 35, 'Chicago', 'play tennis'])])

program = OpenAIPydanticProgram.from_defaults(
    output_cls=DataFrameRowsOnly,
    llm=OpenAI(temperature=0, model="gpt-4-0613"),
    prompt_template_str=(
        "Please extract the following text into a structured data:"
        " {input_str}. The column names are the following: ['Name', 'Age',"
        " 'City', 'Favorite Sport']. Do not specify additional parameters that"
        " are not in the function schema. "
    ),
    verbose=True,
)

program(
    input_str="""My name is John and I am 25 years old. I live in 
        New York and I like to play basketball. His name is 
        Mike and he is 30 years old. He lives in San Francisco 
        and he likes to play baseball. Sarah is 20 years old 
        and she lives in Los Angeles. She likes to play tennis.
        Her name is Mary and she is 35 years old. 
        She lives in Chicago."""
)

Function call: DataFrameRowsOnly with args: {
  "rows": [
    {
      "row_values": ["John", 25, "New York", "basketball"]
    },
    {
      "row_values": ["Mike", 30, "San Francisco", "baseball"]
    },
    {
      "row_values": ["Sarah", 20, "Los Angeles", "tennis"]
    },
    {
      "row_values": ["Mary", 35, "Chicago", ""]
    }
  ]
}

DataFrameRowsOnly(rows=[DataFrameRow(row_values=['John', 25, 'New York', 'basketball']), DataFrameRow(row_values=['Mike', 30, 'San Francisco', 'baseball']), DataFrameRow(row_values=['Sarah', 20, 'Los Angeles', 'tennis']), DataFrameRow(row_values=['Mary', 35, 'Chicago', ''])])

Use our DataFrame Programs

We provide convenience wrappers for DFFullProgram and DFRowsProgram. This allows a simpler object creation interface than specifying all details through the OpenAIPydanticProgram.

from llama_index.program import (
    OpenAIPydanticProgram,
    DFFullProgram,
    DFRowsProgram,
)
import pandas as pd

# initialize empty df
df = pd.DataFrame(
    {
        "Name": pd.Series(dtype="str"),
        "Age": pd.Series(dtype="int"),
        "City": pd.Series(dtype="str"),
        "Favorite Sport": pd.Series(dtype="str"),
    }
)

# initialize program, using existing df as schema
df_rows_program = DFRowsProgram.from_defaults(
    pydantic_program_cls=OpenAIPydanticProgram, df=df
)

# parse text, using existing df as schema
result_obj = df_rows_program(
    input_str="""My name is John and I am 25 years old. I live in 
        New York and I like to play basketball. His name is 
        Mike and he is 30 years old. He lives in San Francisco 
        and he likes to play baseball. Sarah is 20 years old 
        and she lives in Los Angeles. She likes to play tennis.
        Her name is Mary and she is 35 years old. 
        She lives in Chicago."""
)

result_obj.to_df(existing_df=df)

/Users/jerryliu/Programming/gpt_index/llama_index/program/predefined/df.py:65: FutureWarning: The frame.append method is deprecated and will be removed from pandas in a future version. Use pandas.concat instead.
  return existing_df.append(new_df, ignore_index=True)

	Name	Age	City	Favorite Sport
0	John	25	New York	Basketball
1	Mike	30	San Francisco	Baseball
2	Sarah	20	Los Angeles	Tennis
3	Mary	35	Chicago

# initialize program that can do joint schema extraction and structured data extraction
df_full_program = DFFullProgram.from_defaults(
    pydantic_program_cls=OpenAIPydanticProgram,
)

result_obj = df_full_program(
    input_str="""My name is John and I am 25 years old. I live in 
        New York and I like to play basketball. His name is 
        Mike and he is 30 years old. He lives in San Francisco 
        and he likes to play baseball. Sarah is 20 years old 
        and she lives in Los Angeles. She likes to play tennis.
        Her name is Mary and she is 35 years old. 
        She lives in Chicago."""
)

result_obj.to_df()

	Name	Age	Location	Hobby
0	John	25	New York	Basketball
1	Mike	30	San Francisco	Baseball
2	Sarah	20	Los Angeles	Tennis
3	Mary	35	Chicago

# initialize empty df
df = pd.DataFrame(
    {
        "City": pd.Series(dtype="str"),
        "State": pd.Series(dtype="str"),
        "Population": pd.Series(dtype="int"),
    }
)

# initialize program, using existing df as schema
df_rows_program = DFRowsProgram.from_defaults(
    pydantic_program_cls=OpenAIPydanticProgram, df=df
)

input_text = """San Francisco is in California, has a population of 800,000. 
New York City is the most populous city in the United States. \
With a 2020 population of 8,804,190 distributed over 300.46 square miles (778.2 km2), \
New York City is the most densely populated major city in the United States.
New York City is in New York State.
Boston (US: /ˈbɔːstən/),[8] officially the City of Boston, is the capital and largest city of the Commonwealth of Massachusetts \
and the cultural and financial center of the New England region of the Northeastern United States. \
The city boundaries encompass an area of about 48.4 sq mi (125 km2)[9] and a population of 675,647 as of 2020.[4]
"""

# parse text, using existing df as schema
result_obj = df_rows_program(input_str=input_text)

new_df = result_obj.to_df(existing_df=df)
new_df

/Users/jerryliu/Programming/gpt_index/llama_index/program/predefined/df.py:65: FutureWarning: The frame.append method is deprecated and will be removed from pandas in a future version. Use pandas.concat instead.
  return existing_df.append(new_df, ignore_index=True)

	City	State	Population
0	San Francisco	California	800000
1	New York City	New York	8804190
2	Boston	Massachusetts	675647