The "Why knowledge graph?" question i feel was not answered adequately. The question is implicitly comparing KG with NOSQL/SQL so only a thoughtful and specific comparison between the two is appropriate; I'll give it a shot: The answer is "the KG is able to express ambiguity in it's syntax that isn't possible in NOSQL/SQL, or is prohibitively expensive compared to KG's which is very cheap. In cypher, we easily express Patterns, that personify ambiguity/non-completeness. This ability to speak to your Database in patterns, allows us to instruct the KG to 'Give me whatever Entity type exists at this location; node or relationship, I'm not sure what it is, but i know it matches this Pattern: ' Another mental model: 'Ask a partial question to your KG, and ask it to give you a complete answer, and the complete question it answers, as an output.'. Ironically that's a very similar processes for which we use LLM's. Ambiguity in, Clarity out. KG + LLM are perfectly complementary tools. The best SQL can do as a comparison is the LIKE operator which is prohibitively expensive on a large corpus."
The "Why knowledge graph?" question i feel was not answered adequately. The question is implicitly comparing KG with NOSQL/SQL so only a thoughtful and specific comparison between the two is appropriate; I'll give it a shot:
The answer is "the KG is able to express ambiguity in it's syntax that isn't possible in NOSQL/SQL, or is prohibitively expensive compared to KG's which is very cheap. In cypher, we easily express Patterns, that personify ambiguity/non-completeness. This ability to speak to your Database in patterns, allows us to instruct the KG to 'Give me whatever Entity type exists at this location; node or relationship, I'm not sure what it is, but i know it matches this Pattern: ' Another mental model: 'Ask a partial question to your KG, and ask it to give you a complete answer, and the complete question it answers, as an output.'. Ironically that's a very similar processes for which we use LLM's. Ambiguity in, Clarity out. KG + LLM are perfectly complementary tools. The best SQL can do as a comparison is the LIKE operator which is prohibitively expensive on a large corpus."
cool! thank you very much!
is it possible to chunk the data and make graph purely based on the embedding of that data?
Blog link gives a 404. Has it moved to a new location?
sorry - here you go: jmhreif.com/blog/2024/rag-demo-retrieval/