HyperGraphCAG (Hypergraph-based Retrieval Augmented Generation with Semantic Cache) is an advanced RAG framework that combines structured knowledge representation using hypergraphs with efficient response generation via semantic caching.
Domain-specific documents are segmented into text chunks. A language model extracts entities and relations to build a knowledge hypergraph, where each hyperedge connects two or more entities through a descriptive relation.
At query time, a semantic search retrieves the most relevant entities and hyperedges from the hypergraph based on cosine similarity. These are expanded into a subgraph that captures the relevant context for the query.
The selected subgraph and its associated textual chunks are preprocessed and stored in a semantic cache. During inference, this cache is directly loaded into the model’s memory, enabling low-latency and cost-free response generation without external API calls.
- Hypergraph-based structured knowledge representation
- Semantic search with graph-guided retrieval
- Efficient, cache-based generation with minimal latency
- Fully open-source and offline-compatible setup
Planned improvements include:
- Extending cache memory capacity to overcome LLM context limitations
- Enhancing entity and relation extraction accuracy
- Optimizing the cost and scalability of the cache construction phase
Francesco Della Valle
Master's Thesis in Big Data Engineering
University of Naples "Federico II"
Academic Year 2024/2025