Generating Data Dictionary for Excel Files Using OpenPyxl and AI Agents

Each firm I labored for till at the moment, there it was: the resilient MS Excel.

Excel was first launched in 1985 and has remained robust till at the moment. It has survived the rise of relational databases, the evolution of many programming languages, the Web with its infinite variety of on-line purposes, and at last, additionally it is surviving the period of the AI.

Phew!

Do you could have any doubts about how resilient Excel is? I don’t.

I believe the explanation for that’s its practicality to begin and manipulate a doc shortly. Take into consideration this case: we’re at work, in a gathering, and out of the blue the management shares a CSV file and asks for a fast calculation or just a few calculated numbers. Now, the choices are:

1. Open an IDE (or a pocket book) and begin coding like loopy to generate a easy matplotlib graphic;

2. Open Energy BI, import the information, and begin making a report with dynamic graphics.

3. Open the CSV in Excel, write a few formulation, and create a graphic.

I can’t converse for you, however many instances I am going for choice 3. Particularly as a result of Excel recordsdata are suitable with every thing, simply shareable, and beginner-friendly.

I’m saying all of this as an Introduction to make my level that I don’t assume that Excel recordsdata are going away anytime quickly, even with the quick growth of AI. Many will love that, many will hate that.

So, my motion right here was to leverage AI to make Excel recordsdata higher documented. One of many important complaints of information groups about Excel is the dearth of greatest practices and reproducibility, provided that the names of the columns can have any names and information varieties, however zero documentation.

So, I’ve created an AI Agent that reads the Excel file and creates this small documentation. Right here is the way it works:

The Excel file is transformed to CSV and fed into the Massive Language Mannequin (LLM).

The AI Agent generates the information dictionary with column data (variable identify, information sort, description).

The information dictionary will get added as feedback to the Excel file’s header.

Output file saved with feedback.

Okay. Fingers-on now. Let’s get that completed on this tutorial.

Code

Let’s code! | Picture generated by AI. Meta Llama, 2025. https://meta.ai

We are going to start by establishing a digital setting. Create a venv with the instrument of your selection, similar to Poetry, Python Venv, Anaconda, or UV. I actually like UV, as it’s the quickest and the only, in my view. When you’ve got UV put in [5], open a terminal and create your venv.

uv init data-docs
cd data-docs
uv venv
uv add streamlit openpyxl pandas agno mcp google-genai

Now, allow us to import the required modules. This mission was created with Python 3.12.1, however I consider Python 3.9 or greater may do the trick already. We are going to use:

Agno: for the AI Agent administration
OpenPyxl: for the manipulation of Excel recordsdata
Streamlit: for the front-end interface.
Pandas, OS, JSON, Dedent and Google Genai as assist modules.

# Imports
import os
import json
import streamlit as st
from textwrap import dedent

from agno.agent import Agent
from agno.fashions.google import Gemini
from agno.instruments.file import FileTools

from openpyxl import load_workbook
from openpyxl.feedback import Remark
import pandas as pd

Nice. The following step is creating the capabilities we’ll must deal with the Excel recordsdata and to create the AI Agent.

Discover that every one the capabilities have detailed docstrings. That is intentional as a result of LLMs use docstrings to know what a given perform does and determine whether or not to make use of it or not as a instrument.

So, if you happen to’re utilizing Python capabilities as Instruments for an AI Agent, ensure that to make use of detailed docstrings. These days, with free copilots similar to Windsurf [6] it’s even simpler to create them.

Changing the file to CSV

This perform will:

Take the Excel file and browse solely the primary 10 rows. That is sufficient for us to ship to the LLM. Doing that, we’re additionally stopping sending too many tokens as enter and making this agent too costly.
Save the file as CSV to make use of as enter for the AI Agent. The CSV format is less complicated for the mannequin to soak up, as it’s a bunch of textual content separated by commas. And we all know LLMs shine working with textual content.

Right here is the perform.

def convert_to_csv(file_path:str):
   """
    Use this instrument to transform the excel file to CSV.

    * file_path: Path to the Excel file to be transformed
    """
   # Load the file  
   df = pd.read_excel(file_path).head(10)

   # Convert to CSV
   st.write("Changing to CSV... :leftwards_arrow_with_hook:")
   return df.to_csv('temp.csv', index=False)

Let’s transfer on.

Creating the Agent

The following perform creates the AI agent. I’m utilizing Agno [1], as it is rather versatile and simple to make use of. I additionally selected the mannequin Gemini 2.0 Flash. Throughout the take a look at section, this was the best-performing mannequin producing the information docs. To make use of it, you have to an API Key from Google. Don’t neglect to get one right here [7].

The perform:

Receives the CSV output from the earlier perform.
Passes by means of the AI Agent, which generates the information dictionary with column identify, description, and information sort.
Discover that the description argument is the immediate for the agent. Make it detailed and exact.
The information dictionary will likely be saved as a JSON file utilizing a instrument referred to as FileTools that may learn and write recordsdata.
I’ve arrange retries=2 so we will work round any error on a primary strive.

def create_agent(apy_key):
    agent = Agent(
        mannequin=Gemini(id="gemini-2.0-flash", api_key=apy_key),
        description= dedent("""
                            You might be an agent that reads the temp.csv dataset offered to you and 
                            based mostly on the identify and information sort of every column header, decide the next data:
                            - The information sorts of every column
                            - The outline of every column
                            - The primary column numer is 0

                            Utilizing the FileTools offered, create an information dictionary in JSON format that features the beneath data:
                            {<ColNumber>: {ColName: <ColName>, DataType: <DataType>, Description: <Description>}}

                            In case you are unable to find out the information sort or description of a column, return 'N/A' for that column for the lacking values.
                            
                            """),
        instruments=[ FileTools(read_files=True, save_files=True) ],
        retries=2,
        show_tool_calls=True
        )

    return agent

Okay. Now we’d like one other perform to avoid wasting the information dictionary to the file.

Including Knowledge Dictionary to the File’s Header

That is the final perform to be created. It’s going to:

Get the information dictionary json from the earlier step and the unique Excel file.
Add the information dictionary to the file’s header as feedback.
Save the output file.
As soon as the file is saved, it shows a obtain button for the consumer to get the modified file.

def add_comments_to_header(file_path:str, data_dict:dict="data_dict.json"):
    """
    Use this instrument so as to add the information dictionary {data_dict.json} as feedback to the header of an Excel file and save the output file.

    The perform takes the Excel file path as argument and provides the {data_dict.json} as feedback to every cell
    Begin counting from column 0
    within the first row of the Excel file, utilizing the next format:    
        * Column Quantity: <column_number>
        * Column Identify: <column_name>
        * Knowledge Kind: <data_type>
        * Description: <description>

    Parameters
    ----------
    * file_path : str
        The trail to the Excel file to be processed
    * data_dict : dict
        The information dictionary containing the column quantity, column identify, information sort, description, and variety of null values

    """
    
    # Load the information dictionary
    data_dict = json.load(open(data_dict))

    # Load the workbook
    wb = load_workbook(file_path)

    # Get the lively worksheet
    ws = wb.lively

    # Iterate over every column within the first row (header)
    for n, col in enumerate(ws.iter_cols(min_row=1, max_row=1)):
        for header_cell in col:
            header_cell.remark = Remark(dedent(f"""
                              ColName: {data_dict[str(n)]['ColName']}, 
                              DataType: {data_dict[str(n)]['DataType']},
                              Description: {data_dict[str(n)]['Description']}
    """),'AI Agent')

    # Save the workbook
    st.write("Saving File... :floppy_disk:")
    wb.save('output.xlsx')

    # Create a obtain button
    with open('output.xlsx', 'rb') as f:
        st.download_button(
            label="Obtain output.xlsx",
            information=f,
            file_name='output.xlsx',
            mime='software/vnd.openxmlformats-officedocument.spreadsheetml.sheet'
        )

Okay. The following step is to attach all of this collectively on a Streamlit front-end script.

Streamlit Entrance-Finish

On this step, I might have created a distinct file for the front-end and imported the capabilities in there. However I made a decision to make use of the identical file, so let’s begin with the well-known:

if __name__ == "__main__":

First, a few traces to configure the web page and messages displayed within the Net Software. We are going to use the content material centered on the web page, and there may be some details about how the App works.

# Config web page Streamlit
    st.set_page_config(format="centered", 
                       page_title="Knowledge Docs", 
                       page_icon=":paperclip:",
                       initial_sidebar_state="expanded")
    
    # Title
    st.title("Knowledge Docs :paperclip:")
    st.subheader("Generate an information dictionary in your Excel file.")
    st.caption("1. Enter your Gemini API key and the trail of the Excel file on the sidebar.")
    st.caption("2. Run the agent.")
    st.caption("3. The agent will generate an information dictionary and add it as feedback to the header of the Excel file.")
    st.caption("ColName: <ColName> | DataType: <DataType> | Description: <Description>")
    
    st.divider()

Subsequent, we’ll arrange the sidebar, the place the consumer can enter their API Key from Google and choose a .xlsx file to be modified.

There’s a button to run the appliance, one other to reset the app state, and a progress bar. Nothing too fancy.

with st.sidebar:
        # Enter your API key
        st.caption("Enter your API key and the trail of the Excel file.")
        api_key = st.text_input("API key: ", placeholder="Google Gemini API key", sort="password")
        
        # Add file
        input_file = st.file_uploader("File add", 
                                       sort='xlsx')
        

        # Run the agent
        agent_run = st.button("Run")

        # progress bar
        progress_bar = st.empty()
        progress_bar.progress(0, textual content="Initializing...")

        st.divider()

        # Reset session state
        if st.button("Reset Session"):
            st.session_state.clear()
            st.rerun()

As soon as the run button is clicked, it triggers the remainder of the code to run the Agent. Right here is the sequence of steps carried out:

The primary perform is named to rework the file to CSV
The progress is registered on the progress bar.
The Agent is created.
Progress bar up to date.
A immediate is fed into the agent to learn the temp.csv file, create the information dictionary, and save the output to data_dictionary.json.
The information dictionary is printed on the display, so the consumer can see what was generated whereas it’s being saved to the Excel file.
The Excel file is modified and saved.

# Create the agent
    if agent_run:
        # Convert Excel file to CSV
        convert_to_csv(input_file)

        # Register progress
        progress_bar.progress(15, textual content="Processing CSV...")

        # Create the agent
        agent = create_agent(api_key)

        # Begin the script
        st.write("Operating Agent... :runner:")

        # Register progress
        progress_bar.progress(50, textual content="AI Agent is operating...")

        # Run the agent    
        agent.print_response(dedent(f"""
                                1. Use FileTools to learn the temp.csv as enter to create the information dictionary for the columns within the dataset. 
                                2. Utilizing the FileTools instrument, save the information dictionary to a file named 'data_dict.json'.
                                
                                """),
                        markdown=True)

        # Print the information dictionary
        st.write("Producing Knowledge Dictionary... :page_facing_up:")
        with open('data_dict.json', 'r') as f:
            data_dict = json.load(f)
            st.json(data_dict, expanded=False)

        # Add feedback to header
        add_comments_to_header(input_file, 'data_dict.json')

        # Take away non permanent recordsdata
        st.write("Eradicating non permanent recordsdata... :wastebasket:")
        os.take away('temp.csv')
        os.take away('data_dict.json')    
    
    # If file exists, present success message
    if os.path.exists('output.xlsx'):
        st.success("Completed! :white_check_mark:")
        os.take away('output.xlsx')

    # Progress bar finish
    progress_bar.progress(100, textual content="Completed!")

That’s it. Here’s a demonstration of the agent in motion.

Knowledge Docs added to your Excel File. Picture by the writer.

Stunning outcome!

Attempt It

You’ll be able to strive the deployed app right here: https://excel-datadocs.streamlit.app/

Earlier than You Go

In my humble opinion, Excel recordsdata should not going away anytime quickly. Loving or hating them, we’ll have to stay with them for some time.

Excel recordsdata are versatile, simple to deal with and share, thus they’re nonetheless very helpful for the routine ad-hoc duties at work.

Nevertheless, now we will leverage AI to assist us deal with these recordsdata and make them higher. Artificial Intelligence is touching so many factors of our lives. The routine and instruments at work are solely one other one.

Let’s benefit from AI and work smarter daily!

If you happen to preferred this content material, discover extra of my work in my web site and GitHub, shared beneath.