Philip Rathle

@HowToAI_ Yep. Called this 18 mos ago here: neo4j.com/blog/genai/gra…

RAG is broken and nobody's talking about it. Stanford researchers exposed the fatal flaw killing every "AI that reads your docs" product in existence. It’s called "Semantic Collapse," and it happens the second your knowledge base hits critical mass. If you've noticed your AI getting "dumber" as you add more data, this is exactly why. Right now, companies are dumping thousands of documents into their AI, thinking it’s getting smarter. When you add a document to RAG, it converts it into a high-dimensional vector. Under 10,000 documents, this works perfectly. Similar concepts cluster together. But past 10,000 documents, the space fills up. The clusters overlap. The distances compress. Everything starts to look "relevant." It is a mathematical law called the Curse of Dimensionality. In a 1000-dimensional space, 99.9% of your data lives on the outer edge. All points become equidistant from each other. That perfect, relevant document you are looking for now has the exact same mathematical similarity as 50 completely irrelevant ones. The Stanford findings are brutal: At 50,000 documents, precision drops by 87%. Semantic search actually becomes worse than old-school keyword search. Adding more context doesn’t fix the AI. It makes the hallucinations worse. Your "nearest neighbor" search isn't finding the best answer anymore. It's finding everyone. We thought RAG solved hallucinations. It didn't. It just hid them behind math.

@alexxubyte + @neo4j graphrag.com

The AI Application Stack for Building RAG Apps

@sahnlam + Neo4j for knowledge graphs & GraphRAG

The Open Source AI Stack

Close to home for Latent Space, Pfizer was recently onstage in NYC at the AI Engineer Summit with their example: ai.engineer/summit/2025/sc… I have more examples linked at the end of the GraphRAG Manifesto including some (LinkedIn) who deployed into production, saw a massive improvement, and wrote a research paper about it. Lots more have come out since I wrote the post. You’ll find more at graphrag.com/appendices/res…

I do get your point and will let folks add in their experiences here. But it’s worth pointing out that part of the thread came from the CEO of Klarna, who considers what they built only to have been possible because of GraphRAG with a graph database (Neo4j in this case): x.com/klarnaseb/stat… He talks some more about it on this Sequoia Training Data episode starting at minute 30: sequoiacap.com/podcast/traini…

i do really like this chart I think people are rightfully wary about the Return of the Knowledge Graph, having been burned before. But when opinions differ I let facts help me decide, and this is a great collection of evidence on the performance of GraphRAG, including third parties like @get_writer and @microsoft!

@latentspacepod @dharmesh @emileifrem @DMRadioOnline @neo4j 11/ We are all nodes in a massive, massive graph! So true!!

@latentspacepod @dharmesh @emileifrem @DMRadioOnline @neo4j 10/ Besides PageRank, one of my favorite graph algorithms is neo4j.com/docs/graph-dat… It’s a great way to resolve anonymous user breadcrumbs into a pseudonymous single identity.

Today’s @latentspacepod with @dharmesh is *of course* an awesome listen: latent.space/p/dharmesh A thread expanding on the graph database portion the discussion 🧵:

@joaomdmoura A big warm welcome to all of you! We are excited to have you here.

Whole family united in the Bay now! Amazing new CrewAI merch btw

@yaakovepstein51 Neo4j Aura starts at $65 perGB per month, includes hosting/patches/backups/etc., and can be gotten through your cloud marketplace of choice. Otherwise Neo4j has an open source Community Edition that may or may not work depending on your scalability/availability/security needs.

@prathle Any ideas for free open source alternatives for those of us in the non profit sector?

Klarna’s AI journey is rooted in the power of graphs. I agree with Sebastian that the value of Neo4j/knowledge graphs/GraphRAG is in the top line. It’s not about replacing SaaS, but bringing data from the many silos into a graph, and using that for better AI decisions.