How an Old Google Search Trick is Solving AI Hallucinations

If you’ve kept up with this newsletter, you’ve likely learned a lot about retrieval augmented generation, a commonly-used technique that reduces bad results from large language models by providing the models with access to data whose accuracy has been vetted.

Artificial intelligence developers have improved traditional RAG (otherwise known as vector-based RAG) by tailoring it to specific domains like law or medicine, but that doesn’t solve the chronic issue of LLMs being unable to tell the difference between requests and answers that sound similar but aren’t. In other words, they can confuse phrases like “my flight was canceled” with “I want to cancel my flight.”

Now, developers such as Contextual AI, Writer and Neo4j are experimenting with a newer approach, graph-based RAG, that looks quite good at fixing that limitation.