From the archive: language models beyond chat
Three threads from ReLU's project archive: extracting ESG data, classifying documents, and building agents that query data and run analyses.
ReLU’s project archive already contains several semesters of work on language models and agents. Here are three projects relevant to the NLP & Agents lab, brought together as a starting point for this log.
Extracting ESG data from sustainability reports
In Spring 2026, the Rystad Energy team developed and evaluated two approaches to extracting ESG data from oil and gas operators’ sustainability reports: direct LLM extraction from PDFs, and a retrieval-augmented pipeline that first finds relevant sections.
The handoff included source code, metric-specific prompts, evaluation functions, and audit logging to trace extracted data back through the pipeline.
Read the Rystad project writeup →
Classifying documents in a large label hierarchy
The Triplan collaboration tackled document classification across a tree with thousands of leaf labels, including many with similar names. The aim was to narrow down candidate classes and suggest labels to the person doing the sorting.
The Fall 2025 research compared LLMs, classical machine learning, and encoder fine-tuning. The Spring 2026 team continued testing and refining the most promising approaches.
Read the Triplan project writeup →
Agents for data retrieval and analysis
With Solution Seeker, teams worked on a chat-based data exploration tool across Fall 2024 and Spring 2025. The agent could generate code and database queries, retrieve data, and produce visualizations or statistical analyses through Plotly Dash. The first phase handed over a proof-of-concept app and documentation; the second continued the work through partner systems.
Find both phases in the project archive →
For more work on extraction, summarization, entity matching, and evaluation, browse the full project archive.