Descriptions:
Emil Eifrem, CEO and co-founder of Neo4j, sits down with the Latent Space podcast to discuss how knowledge graphs are reshaping AI retrieval pipelines — from GraphRAG architecture to the practical limits of vector-only search. Eifrem argues that the key advantages of graph-based retrieval are not speed per se, but accuracy and explainability: unlike cosine-similarity scores in vector space, graph relationships are explicit and visually inspectable, making it possible to audit why the system retrieved specific documents.
The conversation covers the maturation of GQL (Graph Query Language) as an ISO standard — the first sibling language to SQL — and how its growing presence in LLM training data is improving out-of-the-box text-to-Cypher quality. Eifrem also discloses that Neo4j internally fine-tunes a derived version of a Gemini model for its Cypher co-pilot feature, adding a post-processing regex layer to correct directional arrow issues that the base model still gets wrong. He is candid that even fine-tuned models fall short of 99% reliability for all Cypher generation tasks today.
Other topics include the evolution of vector databases (Eifrem frames them as complementary rather than competitive), YouTube’s LLM-based recommendation system as a case study in tokenizing non-language entities, and fraud detection as a graph-native use case that remains difficult to replicate in flat vector stores. Developers building hybrid RAG pipelines or evaluating graph databases for enterprise search will find specific implementation insights not available in standard documentation.
📺 Source: Latent Space · Published April 18, 2026
🏷️ Format: Interview
