LLMs in Action: How Norkart Leverages AI - Mathilde Ørstavik & Malte Loller-Andersen

Learn how Norkart built an AI system to simplify building permit applications in Norway using LLMs, custom models, and RAG architecture to interpret complex regulations.

Key takeaways
  • Norwegian municipalities handle ~90,000 building permit applications yearly, with 40% containing errors due to complex zonal regulations and documentation

  • Norkart developed “Plandprat” - an LLM-based system to analyze zonal plans and help understand building regulations by:

    • Using property boundaries to identify relevant zonal plans
    • Converting documents into searchable chunks through preprocessing
    • Employing RAG (Retrieval Augmented Generation) architecture to find relevant information
    • Providing property-specific answers about building regulations
  • Key technical components include:

    • Python/Flask backend with Docker deployment
    • Azure OpenAI GPT-4 for text generation
    • Norwegian-specific LLM models (NURVAT) for better language understanding
    • Custom embedding model (Plambert) trained on 70,000 zonal plans
    • RagAS framework for evaluation metrics
  • System evaluation focuses on:

    • Context precision and recall
    • Faithfulness of LLM responses
    • Answer correctness
    • Hallucination detection
  • Challenges addressed:

    • Processing various document formats (scanned, tables, maps)
    • Handling multiple applicable zonal plans
    • Converting complex legal language into understandable responses
    • Maintaining accuracy while providing concise answers
    • Dealing with Norwegian language specifics in LLM processing
  • The system improves building permit applications by:

    • Reducing time spent searching through regulations
    • Making complex rules more accessible
    • Providing property-specific answers
    • Helping identify relevant regulations early in the process