Going Meta - S02 Ep02: Using Ontologies to Guide Knowledge Graph Creation Part 2 

I feel incredibly fortunate to see the leading edge unfolding. An LLM can use an ontology to assign correct types of entities as it creates a knowledge graph from unstructured text. There are challenges ahead: entity resolution, using identifiers from existing graphs such as Wikidata, fine tuning an LLM on an ontology to avoid passing the types of thousands of entities in every prompt. I look forward to the next episode, and the next step towards the synergy between language models and graph+vector databases, and towards collective terrestrial intelligence.

Actually, the vector embeddings of statements in the user input would match with vector embeddings of triplets in the graph of ontologies, which could be stored as properties of the relationships, of the edges. Retrieved responses to prompts would be like fast thinking, and generated responses would be like slow thinking.

(Synergy of human intelligence with (Synergy of language models with (Synergy of vector with graph databases))). The whole is greater than the sum of the parts because the parts are different.

Could a vector embedding of the user input match with a cluster of nodes in a hybrid vector and graph database of ontologies, to retrieve the relevant entity and relationship types? If the retrieved entity and relationship types are used repeatedly, and kept in processor cache memory, latencies could nanoseconds.