Agent Context File Parsing Test Dataset

中文 | English

This repository provides test datasets for Agent Context products to evaluate file parsing across different document formats.

Datasets are managed by file type through separate Git branches. Each branch focuses on one format or format family, making it easier to expand data, run regression tests, and evaluate parsing performance independently. The main branch only keeps repository documentation and the branch index; concrete test datasets live in their corresponding branches.

Branches

Branch	Dataset Type	Description
markdown	Markdown	Markdown parsing datasets, including nested directories, long documents, chapter notes, and atomic knowledge-point cards.
pdf	PDF	PDF parsing datasets for validating layout, pagination, text extraction, and structure reconstruction.
csv	CSV	CSV parsing datasets for delimiters, headers, field types, missing values, and large-file handling.
image	Image	Image parsing datasets for OCR, image understanding, mixed image-text content, and visual information extraction.
ppt	PowerPoint	Presentation parsing datasets for slide structure, text boxes, images, tables, and reading order.
word	Word	Word document parsing datasets for paragraphs, headings, lists, tables, headers, footers, and style structure.
excel	Excel	Excel workbook parsing datasets for multiple sheets, formulas, merged cells, table regions, and data types.

Usage

Switch to the corresponding branch to view or maintain a dataset for a specific file format:

git switch markdown
git switch pdf
git switch csv
git switch image
git switch ppt
git switch word
git switch excel

When adding new test data, place it in the matching format branch and update that branch's README with dataset size, source notes, directory structure, and applicable test scenarios.

Dataset Goals

• Cover common file parsing input formats used by Agent Context products.
• Provide reusable samples for parsing quality, performance, stability, and regression tests.
• Allow each file type to evolve independently without mixing unrelated datasets.
• Preserve realistic structural complexity, such as long documents, nested directories, tables, images, pagination, multiple sheets, and multiple slides.

Contributors

License

This repository is licensed under the Apache License 2.0.