EntRAG Benchmark: Question Answering Dataset

Description

EntRAG is a specialized benchmark dataset designed for evaluating Retrieval-Augmented Generation (RAG) systems in enterprise contexts. The dataset addresses the unique challenges of business environments where information comes from heterogeneous sources including structured databases, documents, and dynamic mock APIs.

The dataset comprises 100 manually constructed question-answer pairs across six enterprise domains: Finance, Technical Documentation, Environment, Legal and Compliance, Human Resources, and Marketing and Sales. Questions are designed to evaluate both static document retrieval and dynamic API integration scenarios, reflecting realistic enterprise information needs.

Dataset Structure

Columns

• id: Unique identifier for each question-answer pair
• domain: The subject area or field of knowledge the question pertains to (e.g., "Technical Documentation", "Finance", "Healthcare")
• question_type: The category of reasoning required (e.g., "comparison", "factual", "analytical", "procedural")
• dynamism: Indicates whether the answer content changes over time ("static" for timeless information, "dynamic" for evolving content)
• question: A natural language question that requires information retrieval and reasoning to answer accurately
• reference_answer: The correct, comprehensive answer that serves as the ground truth for evaluation
• sources: Array of source documents that contain the information needed to answer the question, including:
  • id: Unique identifier for the source
  • filename: Name of the source document or API endpoint
  • pages: Array of specific page numbers where relevant information is found (empty for API sources)

Use Cases

This dataset is particularly valuable for:

• RAG System Evaluation: Testing RAG systems with realistic business scenarios and multi-source information integration
• Hybrid System Assessment: Evaluating systems that combine document retrieval with API-based data access
• Domain-Specific Analysis: Understanding RAG performance across different business domains
• Dynamic Information Handling: Assessing systems that work with both static documents and real-time data sources

Accessing the Dataset

You can load this dataset via the Hugging Face Datasets library using the following Python code:

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("fkapsahili/EntRAG")

# Access the data
for example in dataset['train']:
    print(f"Domain: {example['domain']}")
    print(f"Question Type: {example['question_type']}")
    print(f"Dynamism: {example['dynamism']}")
    print(f"Question: {example['question']}")
    print(f"Answer: {example['reference_answer']}")
    print(f"Sources: {len(example['sources'])} documents")
    print("---")

Alternative Loading Methods

For direct integration with evaluation frameworks:

import json
from datasets import load_dataset

# Load and convert to list format
dataset = load_dataset("fkapsahili/EntRAG", split="train")
qa_pairs = [dict(item) for item in dataset]

Integration with RAG Frameworks

This dataset supports evaluation of various RAG architectures and can be integrated with existing evaluation pipelines. The format is compatible with standard RAG evaluation frameworks and supports both document-based and API-integrated systems.

Dataset Statistics

• Total QA Pairs: 100 manually constructed questions
• Domains: 6 domains (Finance, Technical Documentation, Environment, Legal and Compliance, Human Resources, Marketing and Sales)
• Question Types: 7 reasoning patterns (simple queries, comparison, aggregation, multi-hop reasoning, simple with conditions, factual contradiction, post-processing)
• Dynamism Distribution:
  • Static questions: 28% (document-based retrieval)
  • Dynamic questions: 72% (requiring real-time API integration)
• Source Documents: 9,500+ pages from authentic enterprise documents across 10 major companies
• Company Sectors: Technology, healthcare, e-commerce, retail, automotive, and energy
• Mock APIs: 4 domain-specific APIs (finance, SEC filings, HR statistics, web search)

Citation

If you use this dataset in your research, please cite:

@dataset{entrag_2025,
  title={EntRAG: Enterprise RAG Benchmark},
  author={Fabio Kapsahili},
  year={2025},
  publisher={Hugging Face},
  url={https://huggingface.co/datasets/fkapsahili/EntRAG}
}

License

This dataset is released under Creative Commons Attribution 4.0. Please see the LICENSE file for full details.

Additional Resources

• Evaluation Code: https://github.com/fkapsahili/EntRAG

For questions, issues, please open an issue in the associated GitHub repository.