Graph-free Retrieval with Graph-Modified Embeddings
What it is
This paper proposes a method to improve dense vector retrieval by using graph-modified embeddings to guide search while retaining original embeddings for semantic relevance. It demonstrates that this multi-stage approach can recover up to 25.9% more evidence than plain vector retrieval on synthetic benchmarks without requiring graph access at query time.
The problem it solves
Dense vector retrieval often misses evidence connected through intermediate entities, while direct graph retrieval requires expensive and often unavailable graph access at query time.
What you could use it for
- Improving retrieval accuracy for questions requiring linked facts or comparisons among candidates
- Reducing the number of returned nodes needed to match baseline recovery rates
- Enhancing RAG pipelines where graph structures exist offline but cannot be queried in real-time
Read the source ↗Back to the library →
Summary written by an AI model from the source; check the source before you rely on it.