The GNN is trained on a heterogeneous graph. Using a MLP projector, the embeddings of the nodes are projected onto an llm's embedding space using a text description obtained for each node.
Once this is done, the embeddings are then stored to a custom HNSW vector store.
This store can then be queried to obtain relevant results.
This can be used in RAG based LLM systems that need to work on graph data.