ESG data extraction using LLMs and RAG
Continued the Rystad collaboration with automated extraction of ESG data from oil and gas sustainability reports — comparing direct LLM extraction with a retrieval-augmented pipeline.
During the Spring 2026 semester, ReLU continued its collaboration with Rystad Energy, focusing on automated extraction of ESG (Environmental, Social, and Governance) data from sustainability reports published by oil and gas operators.
ESG data extraction using LLMs
These reports contain key indicators such as greenhouse gas (GHG) emissions, methane intensity and flaring rates, as well as social metrics like safety incidents and workforce diversity, and governance metrics such as board composition. However, they are highly unstructured and poorly standardized across companies.
ReLU NTNU developed and evaluated two approaches — a direct LLM-based extraction from PDFs, and a retrieval-augmented generation (RAG) pipeline that first identifies relevant sections before extraction.
The final deliverable consisted of the full source code for the extraction system, including pipelines, metric-specific prompt templates, evaluation functions, and audit logging for traceability and validation of the extracted data.