Lexi AI

The challenge at Lexi AI.

Collective labour agreements are long, legalistic and scattered across hundreds of articles. An employee or HR adviser with a concrete question, say about leave, allowances or notice periods, has to wade through dozens of pages and hope to land on the right clause. For Lexi the brief was clear: build an assistant that answers such a question in plain language and grounds every answer in the correct source article. The hard part is the interdependence between clauses, because a leave rule references a definition elsewhere, which in turn points to an exception.

Our approach.

We built a RAG pipeline on a vector database for semantic retrieval, paired with a knowledge graph in Memgraph that makes the references between CAO articles, definitions and exceptions explicit. A question retrieves not just the most semantically related passage, but also the clauses it depends on. Claude composes the answer from that assembled context and cites the source articles.

The graph keeps the model from reading isolated fragments without their legal connections, so each answer stays traceable back to the agreement itself.

What it delivered.

Lexi answers CAO questions in plain language with an accuracy of 95 percent, measured against a curated set of reference questions. By combining vector retrieval with the Memgraph knowledge graph, answers are not only correct but backed by the right source articles, letting a user verify the reasoning.

The result is an assistant that removes the manual searching through dense CAO documents and gets HR professionals and employees to a well-supported answer faster.