Community guidelines
Be specific and constructive. No vendor spam — promoting your own product belongs in a listing. Anyone can read; posting needs a free account.
Our RAG system can answer what a document says. It struggles with questions like why a contract changed, who approved it and what happened afterward across email, CRM and shared drives. People keep suggesting GraphRAG. Is that solving a real problem or adding a graph because graphs are fashionable again?
It solves a real problem if relationships and time are part of the answer. A vector hit can find the approval sentence but not necessarily connect it to the original contract and later exception. The hard part is building trustworthy entities and edges
We got most of the value first by adding document lineage, timestamps and stable IDs. No graph database.
Our metadata is inconsistent, so that probably explains more than the embedding model. Would you let an LLM create the links?
As a candidate generator, yes. Not as unquestioned truth. Store confidence and evidence for every link, and let users open the source. A beautiful graph of guessed relationships is worse than plain search because it looks authoritative