A Gulf social-insurance authority

The problem

Users struggled to locate relevant retirement-law provisions through exact terms alone. Retirement law is written in statutory language that nobody uses to describe their own situation - someone asking when they can retire early has to guess the formal vocabulary before keyword search can help them. The cost of landing on the wrong provision is not a poor search experience; it is a member forming a wrong expectation about their own pension.

What we built

We combined multilingual embeddings, cosine similarity, structured legal content, and LLM relevance refinement to handle the complex legal language and the wide variation in how members phrase a query.

What it turned on

Refining relevance after retrieval, rather than trusting similarity alone, is what kept a semantically close but legally wrong provision from being presented as the answer. Relevant to any body whose entitlements are defined in statute that its members have to interpret for themselves.