RAG-GFM: Overcoming In-Memory Bottlenecks in Graph Foundation Models via Retrieval-Augmented Generation

This repository contains the official implementation of RAG-GFM, proposed in the paper RAG-GFM: Overcoming In-Memory Bottlenecks in Graph Foundation Models via Retrieval-Augmented Generation (WWW 2026).

RAG-GFM addresses the fundamental in-memory bottleneck of Graph Foundation Models (GFMs) by externalizing graph knowledge into a unified retrieval system. Instead of compressing heterogeneous semantic and structural knowledge into model parameters, RAG-GFM leverages retrieval-augmented generation to enable scalable, interpretable, and efficient cross-domain graph learning.

---

🔍 Key Ideas

• Dual-Modal Knowledge Externalization

  • Semantic Store: Prefix-structured node texts stored in a vector database for controllable semantic retrieval.
  • Structural Store: Centrality-based graph motifs encoded via Walk-Spectrum Encoding (WSE) to capture higher-order structural patterns.

• Cross-View Knowledge Alignment

  • Self-supervised alignment between semantic and structural views during multi-domain pre-training to learn transferable priors.

• In-Context Retrieval Augmentation

  • Retrieved texts and motifs are injected as contextual evidence during few-shot adaptation, enabling efficient downstream learning without updating backbone parameters.

---

📦 Environment Setup

Install dependencies:

pip install -r requirements.txt

We recommend using Python ≥ 3.9 and a CUDA-enabled environment for efficient training and retrieval.

---

🚀 Running Pipeline

The overall workflow follows three stages:
(1) Knowledge Externalization → (2) Cross-View Pre-training → (3) Retrieval-Augmented Adaptation

1. Build Semantic (Text) Vector Database

python build_nano_db.py

Creates a semantic vector database from prefix-structured node texts using dense embeddings.
This database supports top-k semantic retrieval during both pre-training and fine-tuning.

---

2. Train Graph Motif Finders

python train_all_motif_finders.py

Trains motif encoders based on Walk-Spectrum Encoding (WSE) to identify structurally important nodes and subgraphs.
These motifs form the basis of the structural retrieval store.

---

3. Build Structural Motif Databases

python build_all_motif_dbs.py

Constructs motif-level vector databases from trained motif finders, enabling efficient retrieval of transferable structural patterns.

---

4. Run Node Classification (Few-Shot)

python execute_cora.py

Evaluates RAG-GFM on node classification under few-shot and cross-domain (LODO) settings using the Cora dataset.

---

5. Run Graph Classification (Few-Shot)

python execute_graph_cora.py

Evaluates graph-level classification by reformulating node-centered ego-graphs as graph instances.

---

📊 Supported Tasks & Settings

• Tasks

  • Few-shot Node Classification
  • Few-shot Graph Classification

• Evaluation Protocols

  • Leave-One-Dataset-Out (LODO-dataset)
  • Leave-One-Domain-Out (LODO-domain)

• Domains

  • Citation Networks (Cora, CiteSeer, PubMed)
  • E-Commerce Graphs (Ogbn-Products)
  • Web Link Graphs (Wiki-CS)

---

📂 Dataset Download

Download the required datasets from the following links:

• Dataset Link 1
• Dataset Link 2

After downloading, place all datasets in the project root directory before running experiments.

---

📖 Citation

If you find this work useful, please cite:

@inproceedings{yuan2026rag,
  author    = {Haonan Yuan and Qingyun Sun and Jiacheng Tao and Xingcheng Fu and Jianxin Li},
  title     = {RAG-GFM: Overcoming In-Memory Bottlenecks in Graph Foundation Models via Retrieval-Augmented Generation},
  booktitle = {Proceedings of the ACM Web Conference 2026 (WWW '26)},
  year      = {2026},
  publisher = {ACM},
  address   = {New York, NY, USA},
  doi       = {10.1145/3774904.3792139},
  url       = {https://doi.org/10.1145/3774904.3792139}
}

---

📬 Contact

For questions or discussions, please contact Haonan Yuan or open an issue in this repository.

Enjoy exploring retrieval-augmented graph foundation models! 🚀