LLMs in Action: How Norkart Leverages AI
summary
Norkart builds a RAG pipeline called PlanPrat that combines RAG retrieval with OCR multimodal checks and GPT-4 Turbo on Azure to answer property permit questions; they index ~70,000 zonal plans and augment embeddings with 150,000 synthetically generated questions for a domain-tuned model. Evaluation uses RAGAS with context precision/recall, faithfulness and answer correctness across a 17-question test set. PlanPrat runs in aisak, is in pilots with three architect firms and reduces manual search time while flagging hallucinations.
AI-written summary · about
chapters (10)
- 0:00Introduction and context
- 4:32The problem of zonal plans
- 12:04LLM basics and RAG concept
- 17:49Building the PlanPrat system
- 21:59Preprocessing and OCR challenges
- 27:28Implementation in eByggesøk
- 31:05Improving accuracy with RAG
- 37:19Geospatial data integration
- 40:11System evaluation and benchmarking
- 51:23Models, training, and conclusion
related talks
18mLLMs and GenAI: A Guide to Building Your Own RAG Model for Document Interaction
39mAgentic RAG Platform for AI-Assisted Workflows
22mThe Power of RAG at Scale: A Year of Explosive Growth and Observab...
41mFrom RAGs to riches: Build an AI document interrogation app in 30 mins
01h 34mHow to Construct and Assess Your First LLM Based RAG System
32mScaling Generative AI: Building Production-Ready LLM Applications
40mRolling out Enterprise AI: Tools, Insights, & Team Empowerment
48mGoing Beyond Prompt Engineering