langchain-excel-loader 0.1.0

No project description provided

An Excel Loader for Langchain that Preserves Document Structure

Background

The current solution from langchain for loading .xlsx is by using the Unstructured document loader. This has two disadvantages:

1. No attempt is made to preserve the structure of the document. This is as opposed to the CSV loader for example which ingests by row with the column title for each cell on the row:

CSV loader example

csv:

Name,Age Harry,21 Mary,48

Output:

[Document(page_content='Name: Harry \n Age: 21', metadata={'source':'csv.csv', 'row:0'}),
 Document(page_content='Name: Mary \n Age: 48', metadata={'source':'csv.csv', 'row:1'})]

Documents like these give the LLM the context to understand the meaning behind data.

Instead of an approach like the above, the Unstructured Excel Loader will simply add all the text content contained in the xlsx in one string with no indication of columns or rows.

2. The second disadvantage is that the Unstructured package is large with multiple system dependencies and so not suitable for all environments and use cases.

Implementation of the StructuredExcelLoader

This package provides a StructuredExcelLoader, which uses openpyxl to read the .xlsx file. Since Excel spreadsheets have a less fixed structure than csv files, we opt to preserve the column and row number for each cell, giving the LLM a greater remit in inferring meaning from the document.

Example Output

Given an Excel file sample.xlsx with two sheets:

Sheet: "Employees"

Employee	Department	Salary
John Doe	Sales	50000
Jane Smith	Marketing	55000

Sheet: "Departments"

Department	Location	Manager
Sales	New York	Bob Wilson
Marketing	Chicago	Sarah Lee

The StructuredExcelLoader will create separate documents for each sheet:

[
    Document(
        page_content='''SHEET: "Employees"

ROW 1:
CELL A1: Employee
CELL B1: Department
CELL C1: Salary

ROW 2:
CELL A2: John Doe
CELL B2: Sales
CELL C2: 50000

ROW 3:
CELL A3: Jane Smith
CELL B3: Marketing
CELL C3: 55000''',
        metadata={'source': 'sample.xlsx', 'sheet_name': 'Employees'}
    ),
    
    Document(
        page_content='''SHEET: "Departments"

ROW 1:
CELL A1: Department
CELL B1: Location
CELL C1: Manager

ROW 2:
CELL A2: Sales
CELL B2: New York
CELL C2: Bob Wilson

ROW 3:
CELL A3: Marketing
CELL B3: Chicago
CELL C3: Sarah Lee''',
        metadata={'source': 'sample.xlsx', 'sheet_name': 'Departments'}
    )
]

Disadvantages

This approach is not as strong when the .xlsx is extremely complex as the LLM struggles to maintain understanding of the positioning of multiple Tables within a sheet. Although with the latest models (e.g. ChatGPT 4o, Gemini 2.5 at the time of writing) this limit has improved along with the LLMs' abilities to understand cell references accurately

Future Work

After the effectiveness of this approach is validated, it should be incorportaed into the langchain_community.document_loaders repository, alongside the existing UnstructuredExcelLoader, which still provides use in some cases.

Alternatively, an additional boolean argument could be provided called "preserve_structure", which would be set to true by default. If it is explicity set to false, the loader could produce documents as raw text strings without cell references.