Tensor GraphRAG

Differentiable Multi-Hop Reasoning for Retrieval-Augmented Generation

Official implementation of the paper:

Tensor GraphRAG: Differentiable Multi-Hop Reasoning for Retrieval-Augmented Generation

Tensor GraphRAG reformulates multi-hop retrieval-augmented generation as a hybrid continuous–discrete inference process over a learnable multi-relational document graph. It combines:

• Sparse tensor diffusion for query-conditioned multi-hop evidence propagation
• Spectral context modeling for global graph organization
• Confidence-controlled latent-link synthesis for bridging weakly connected evidence

---

Key Results

Setting	Tensor GraphRAG	Strongest Baseline (HippoRAG 2)
Overall EM	63.7	59.6
4-hop EM	54.8	45.9
Relative degradation (1→4 hop)	25.5%	36.2%
Reasoning latency	89 ms	105 ms

---

Repository Structure

Tensor-GraphRAG/
├── configs/                  # Dataset-specific hyperparameters
├── src/
│   ├── data/                 # Preprocessing & hop-split scripts
│   ├── models/               # Core modules (encoder, diffusion, spectral, latent-link)
│   ├── train.py
│   └── evaluate.py
├── scripts/                  # One-click reproduction scripts
├── notebooks/                # Qualitative analysis
└── requirements.txt