Exactly, thank you. Still in LLM-based extraction.

Actually, the name definitely came after noticing RAM prices. Though the idea where the graph-in-memory only for ephemeral RAG sessions came first, we won't pretend the naming wasn't influenced by RAM being in the spotlight.

There are two steps:

Vector search (HNSW): Find top-k similar entities/text units from the query embedding

Graph traversal (BFS): From those seed entities, traverse relationships (up to 2 hops by default) to find connected entities

This catches both semantically similar entities AND structurally related ones that might not match the query text.

Implementation: https://github.com/gibram-io/gibram/blob/main/pkg/engine/eng...

This is how I did it a few years back while working for a set store company. It works well.

I've just released a casual personal project for Ephemeral GraphRAG. It's still experimental and open source: https://github.com/gibram-io/gibram