Don’t generate the graph. Decide it.
Extract a knowledge graph with a Jev-style decision model on Ollama
Most tools ask an AI to write a knowledge graph and hope it is right. This one asks small yes-or-no questions, one link at a time, and keeps only the answers the model is sure about. The rest wait for a person.
A chat model writes the answer
~250 tokenswritten for each sentence (a token is about ¾ of a word)
A decision model answers a question
3 tokensfor two complete answers
Does EdgeQuake use PostgreSQL? 0.99 Did Acme Inc found PostgreSQL? 0.01
A real call to tev1 on one computer, 2.9 seconds. 0.99 means “very sure: yes”.
2.3×
fewer wrong links than Mistral Small
On a public news test of 288 sentences it kept 218 wrong links against 504. It found 154 right links to Mistral’s 183.
10×
faster than a chat model on the same computer
One seven-sentence note: 7.9 seconds for tev1 against 77.8 for gemma4. Later, with tev1 already loaded: 4.3 seconds.
0
invented kinds of name or link
Everything must come from the list you wrote. A link your list forbids is never even asked.
What it does not do. A model trained on that same test still scores higher (0.89 on names and 0.71 on links, against 0.69 and 0.39 here). A chat model on a hosted service can be faster: 3.9 seconds on that note. And a high number can still be wrong: it kept both “Ollama uses Nimble” (right, 0.94) and the reversed “Nimble uses Ollama” (wrong, 0.85, still above the 0.80 cutoff). This article shows all of it.