Case Study
AI Legal RAG Agent + Document Automation | 89,600-Section Law Corpus
An architecture firm in Oslo spends a lot of its day answering questions that have exact answers somewhere in Norwegian building law. How close to the neighboring boundary can this extension sit. What does the zoning plan allow on this plot. The answers exist, they are just buried across 89,600 sections.
I built a retrieval agent over that whole corpus and put it where the team already works, inside Monday.com and Microsoft 365 SharePoint. Ask it a question in Norwegian on a project item and it answers on the item, grounded in the law, with the source quoted back.
How it works
Why every answer quotes the law
A legal answer that sounds right and is wrong is worse than no answer. So retrieval is hybrid, keyword and semantic together with reranking on top, and every answer is quote-verified against the real law before it goes out. Full provenance on each one. The corpus refreshes nightly.
Ask it about a wall 1.6 meters from the eastern boundary where the requirement is four meters, and it comes back citing the fire-classification rules and TEK17, then offers to draft the dispensation application.
The document half
Answering questions was only part of the work. The same agent handles the paperwork: one click builds the SharePoint project folder and fills 18 Word templates, pulling the address, the responsible party, the zoning reference and the rest straight from the project item.
It posts what it did back into the Monday.com item, listing which documents it created and which values it filled in, so nobody has to trust it blindly.
Built to move if the rules change
Embeddings and vector search sit close to the data rather than inside a vendor. That means the Claude layer can be swapped for a self-hosted model on vLLM, Qwen or Llama, if GDPR data residency ever forces the whole thing on-premises. Auditability was part of the design from the start.
More screenshots
