📚 DistComp: A Scalability and Reasoning Corpus for GraphRAG

Dataset Summary

The DistComp Corpus is a rigorously curated dataset designed specifically to evaluate the performance, scalability, and reasoning capabilities of Graph Retrieval-Augmented Generation (GraphRAG) systems. It was introduced alongside the LiteRAG architecture (EMNLP 2026 Submission).

While standard multi-hop QA datasets (e.g., HotpotQA) are excellent for evaluating basic entity linking, they predominantly test superficial fact-hopping. DistComp evaluates deep, abstract conceptual synthesis, complex semantic drift, and graph-based reasoning across disjoint knowledge clusters.

A unique feature of this dataset is that it provides fully pre-indexed Microsoft GraphRAG artifacts (Parquet files, LanceDB vector stores, GraphML) for varying corpus sizes. This allows researchers to bypass the prohibitively expensive and time-consuming indexing phase, enabling immediate "plug-and-play" evaluation of query engines.

Dataset Structure & Scalability Splits

To evaluate the $O(N)$ vs. $O(1)$ scaling characteristics of retrieval architectures, the corpus is stratified into geometrically increasing subsets. Each subset contains the raw text, the pre-computed GraphRAG outputs, and the prompt settings:

• doc40/ (40 documents)
• doc80/ (80 documents)
• doc160/ (160 documents)
• doc320/ (320 documents)
• doc640/ (640 documents)
• doc1280/ (1280 documents - approx. 500,000 tokens)

For each split, a corresponding JSON file (e.g., 40docs.json, 1280docs.json) containing the ground-truth QA pairs is provided.

Query Taxonomy

The dataset contains 160 expert-curated queries categorized into four distinct levels of graph traversal complexity:

1. Literal Citation (LC): Single-hop fact retrieval requiring high precision on specific entities.
   • Example: "What is the primary function of the tool NAMB?"
2. Local Reasoning (LR): Requires traversing explicit edges between two or more named entities within a localized subgraph.
   • Example: "Compare the approaches of 'NAMB' and 'Merge, Split, and Cluster' regarding Stream Processing."
3. Global Themes (G): High-level corpus summarization requiring the aggregation of information across multiple communities.
   • Example: "Summarize the major challenges associated with Edge and Fog computing discussed across the entire corpus."
4. Drift Queries (D): Complex, multi-hop exploration across disjoint community clusters, demanding the system to bridge entirely distinct topics.
   • Example: "Focus on the evolution of MPI. How do the papers on InfiniBand and Arm SVE illustrate the shift in where MPI operations are executed?"

Data Instances

An example of a query pair from the JSON evaluation files:

{
  "d_007": {
    "question": "Data Management is critical in Big Data. Contrast the approach of 'Brindexer' for file systems with 'DeepFreeze' for model training states.",
    "ground_truth": "Brindexer manages 'metadata' (file attributes, location) using partitioning and in-tree structures to speed up finding data in large namespaces. DeepFreeze manages 'payload data' (model states) using asynchronous I/O to speed up saving data during computation. Both optimize data operations but target different bottlenecks (lookup latency vs. write throughput)."
  }
}

Source Data & Provenance

The raw text data aggregates high-impact academic literature in the domain of Distributed Computing. The papers were sourced from nine tier-one IEEE and ACM conferences (including CCGRID, EuroSys, ICDCS, and SIGCOMM) published between 2020 and 2022. This domain was deliberately selected for its high semantic complexity and reliance on abstract conceptual dependencies (e.g., "Byzantine Fault Tolerance" vs. "Federated Learning").

Usage / Quick Start

You can use this dataset directly with LiteRAG, Microsoft GraphRAG, LightRAG, or HiRAG using the benchmarking system used in the paper

Using with LiteRAG Benchmark Suite:

# Clone the dataset
git clone https://huggingface.co/datasets/macarronesc/DistComp

# Run the benchmark comparing LiteRAG and GraphRAG on the 40-doc split
python run_benchmark.py \
    --prompts DistComp/40docs.json \
    --graphrag-dir DistComp/doc40/ \
    --literag-config config/literag_config.yaml \
    --engines literag graphrag_local

License

This dataset is released under the Apache 2.0 License. You are free to use, modify, and distribute this dataset for both commercial and non-commercial purposes, provided you include the original copyright notice and state any changes made.

Citation

If you use the DistComp corpus in your research, please cite our paper:

@article{literag2025,
  title={LiteRAG: Cost-Efficient Graph Retrieval-Augmented Generation},
  author={Anonymous},
  journal={EMNLP Submission},
  year={2026}
}