# Efficient Knowledge Graph Construction and Retrieval from Unstructured Text for Large-Scale RAG Systems

**Authors:** Congmin Min, Rhea Mathew, Joyce Pan, Sahil Bansal, Abbas Keshavarzi, Amar Viswanathan Kannan  
**arXiv:** [2507.03226](https://arxiv.org/abs/2507.03226)  
**Date:** 2025-07-04  
**Summary page:** https://www.summarizepaper.com/en/paper/2507.03226v1/

## Key points

- Proposed a scalable and cost-efficient framework for deploying GraphRAG in enterprise environments
- Dependency-based knowledge graph construction pipeline that eliminates reliance on large language models (LLMs)
- Lightweight graph retrieval strategy for high-recall, low-latency subgraph extraction
- Evaluated the framework on SAP datasets for legacy code migration with strong empirical performance improvements over traditional RAG baselines
- Dependency-based construction approach achieved comparable performance to LLM-generated knowledge graphs while reducing costs and improving scalability
- Highlighted scalability by eliminating dependence on large language models for knowledge graph construction
- Future investigations needed to address limitations such as missing context-dependent or implicit relations not directly expressed in surface syntax
- Plan to evaluate generalizability of the method to other settings beyond SAP-specific domains by testing on broader public benchmarks like HotpotQA
- Study presents a promising path for scaling GraphRAG systems in real-world enterprise applications without prohibitive resource requirements, enabling practical, explainable, and domain-adaptable retrieval-augmented reasoning systems.

## Summary

In this study, we propose a scalable and cost-efficient framework for deploying  in enterprise environments. The adoption of GraphRAG has been limited due to the high computational cost of constructing knowledge graphs using large language models (LLMs) and the latency of graph-based retrieval. To address these challenges, we introduce two core innovations: a dependency-based knowledge graph construction pipeline that eliminates reliance on LLMs by leveraging industrial-grade NLP libraries for entity and relation extraction from unstructured text, and a lightweight graph retrieval strategy that combines hybrid query node identification with efficient one-hop traversal for high-recall, low-latency subgraph extraction. We evaluate our framework on two SAP datasets focused on legacy code migration and demonstrate strong empirical performance. Our system achieves significant improvements over traditional RAG baselines based on LLM-as-Judge and RAGAS metrics. Additionally, our dependency-based construction approach attains comparable performance to LLM-generated knowledge graphs while reducing costs and improving scalability. Furthermore, we highlight the scalability of our approach by eliminating dependence on large language models for knowledge graph construction. However, future investigations are needed to address limitations such as missing context-dependent or implicit relations not directly expressed in surface syntax. We also plan to evaluate the generalizability of our method to other settings beyond SAP-specific domains by testing it on broader public benchmarks like HotpotQA. In conclusion, our study presents a promising path for scaling GraphRAG systems in real-world enterprise applications without prohibitive resource requirements. By combining efficient knowledge graph construction from unstructured text with lightweight subgraph retrieval strategies, we pave the way for practical, explainable, and domain-adaptable retrieval-augmented reasoning systems in large-scale enterprise environments.