LLMs, which are large language models, have revolutionized natural language processing in many ways, but still suffer from critical limitations, especially when they need to deal with recent facts, specific domain information or multi-step reasoning. The Retrieval Augmented Generation (RAG), a new approach, aims to fill these gaps. It allows language models to integrate and retrieve information from external sources. However, most existing graph-based RAG systems are optimized for static corpora and struggle with efficiency, accuracy, and scalability when the data is continually growing—such as in news feeds, research repositories, or user-generated online content.
Introducing EraRAG – Efficient updates for evolving data
In order to address these issues, Huawei researchers, Hong Kong University of Science and Technology and WeBank developed EraRAGThe novel retrieval enhanced generation framework is designed specifically for corpora that are dynamic and constantly expanding. EraRAG does not rebuild the whole retrieval structure when new data comes in. Instead, it uses selective localized updates which only touch the affected parts of the graph.
The Core Features
- LSH (Locality-Sensitive Hashing) based on Hyperplanes:
The corpus of each text is divided into vectors. EraRAG then uses randomly sampled hyperplanes to project these vectors into binary hash codes—a process that groups semantically similar chunks into the same “bucket.” The LSH approach is a good way to maintain both semantic coherence as well as efficient grouping. - Hierarchical and Multi-Layered Diagram Construction
EraRAG uses a multilayered graph as its core structure. Each layer summarizes segments of text using language models. Segments that are too large are split, while those too small are merged—ensuring both semantic consistency and balanced granularity. Summary representations on higher levels allow efficient retrieval of both fine-grained queries and abstract ones. - The Localized Incremental Update:
When new data arrives, its embedding is hashed using the original hyperplanes—ensuring consistency with the initial graph construction. The graph is not changed, but only the segments/buckets directly affected by the new data are updated. Updates propagate up the hierarchy of the graph, but they are always localized in the region affected, thereby saving computation costs and time. - Reproducibility and Determinism
EraRAG does not use the same set of hashes as standard LSH. It is therefore possible to assign buckets in a deterministic way, and reproduce them, which allows for efficient, consistent updates.
Performance and Impact
The EraRAG system has been tested on many benchmarks for question-answering.
- Save money on updates: GraphRAG reduces graph reconstruction times and token usage by up to 95% compared with leading graph-based RAG techniques (e.g. GraphRAG RAPTOR HippoRAG).
- Keeps high accuracy: EraRAG consistently outperforms other retrieval architectures in both accuracy and recall—across static, growing, and abstract question answering tasks—with minimal compromise in retrieval quality or multi-hop reasoning capabilities.
- Supports Versatile Queries Needs EraRAG’s graph multilayer design provides a way to quickly retrieve factual information or semantic summary at high levels. It can also be tailored to match the query.
The Practical Effects
EraRAG offers a scalable and robust retrieval framework ideal for real-world settings where data is continuously added—such as live news, scholarly archives, or user-driven platforms. EraRAG strikes the right balance between retrieval and adaptability. This makes LLM applications that are backed by LLM more accurate, reliable and trusted in environments where things change quickly.
Take a look at the Paper The following are some examples of how to get started: GitHub. | The AI Dev newsletter is here! Read by Over 40k developers There are researchers from NVIDIA and OpenAI. Meta, Microsoft and JP Morgan Chase. Amgen, Aflac and Wells Fargo. [SUBSCRIBE NOW]
Nikhil has been an intern consultant with Marktechpost. He has a dual integrated degree in Materials from the Indian Institute of Technology Kharagpur. Nikhil, an AI/ML fanatic, is constantly researching AI/ML applications for biomaterials and other biomedical fields. Material Science is his background. His passion for exploring and contributing new advances comes from this.
