Spectral Graph Pruning for Context Optimization in Retrieval-Augmented Generation

Abstract #

Spectral Graph Pruning (SGP) is a framework for efficient context optimization and compression in Retrieval-Augmented Generation (RAG). It models retrieved text segments as a heterogeneous semantic graph and applies query-biased spectral centrality analysis to identify and retain the most structurally important segments. SGP reduces token consumption by 40-50% while maintaining high reasoning accuracy on multi-hop benchmarks.

Methodology #

The SGP framework constructs a heterogeneous graph representing Chunk Nodes, Entity Nodes, and Structural document hierarchy.

Figure 1: SGP Pipeline: From heterogeneous graph construction to energy-mass pruning.

Key Contributions #

• Context Graph Construction: Models raw text chunks, extracted entities, and document hierarchy as nodes in a unified graph.
• Spectral Centrality: Uses the graph Laplacian to compute query-biased node importance.
• Adaptive Pruning: Detects the “elbow” in cumulative importance to retain only the minimal topological backbone required for reasoning.

Evaluation and Results #

SGP was evaluated on HotpotQA and MuSiQue benchmarks, demonstrating superior performance-to-compression ratios.

Performance Trade-offs #

Figure 2: Quality vs Compression trade-off. SGP maintains high F1 scores even at high compression rates.

Latency Analysis #

Figure 3: Latency breakdown showing minimal spectral scoring overhead compared to LLM inference savings.

Resources #

• Source Code
• View Paper (PDF)

Citation #

@article{sgp2026,
  title={Spectral Graph Pruning for Context Optimization in Retrieval-Augmented Generation},
  author={Gawade, Mayuri and Bavadekar, Aditya and Bhalerao, Shreyash and Bharat, Sarvesh and Bhandari, Chetan and Kokane, Chandrakant},
  journal={arXiv preprint},
  year={2026}
}