Project
annex
annex is an agent for the EU AI Act. You describe an AI system, and it returns the articles that may apply, cited to the text, and says so when the Act does not settle the question.
why i built it
I wanted to know whether retrieval still helps when a model can read the whole law. The Act fits in one model's context window, so the usual answer was no longer obvious.
how it works
I compared three ways of finding the text on the same twelve questions. One read the whole Act in context, one used vector search, and one followed vector search with the Act's own cross-references. It runs fully local on Ollama.
what it shows
Reading the whole Act found every article the answers needed. Search plus cross-references found 83% of them with about a seventh of the prompt tokens, and showed which text it used.