Redfield AB reposted this
Standard-RAG often struggles with corpus-level understanding. Questions like: "How many times did John visit Sally?" aren't answerable if no specific chunk directly contains that information. GraphRAG solves this by transforming text-data into a structured graph ahead of query time. In other words, we first build up a representation of all entities (names, companies etc.) and their relationships, so that when a user asks a question about something structural in the dataset, the information is there, ready-to-go. I built a small tech-demo that does this for a portion of Epstein files, showing how a dataset can be parsed in advance and uploaded to a graph database (in our case Memgraph) complete with a vector index. With the right setup, this enables us to quickly find information via a chat interface, enabling us to find global-level insights over the full dataset. This experimental version also incorporates other features such as hierarchical community detection, automatically generated community summaries, structured search (cypher), multi-hop search with reranking and more. If you're interested check out the article about how it works in Goose via MCP: https://lnkd.in/dY2_3qj7 P.S. Part 2 with a more detailed explainer on the technical details will be coming soon 😁