Peer-reviewed · Energy and AI · 2025

LLM data extraction for emissions assessment.

A large-language-model framework for extracting oil and gas field data from technical literature, validated against expert-labeled ground truth. Published in Energy and AI (Elsevier), 2025.

Summary

The framework uses GPT-4 and GPT-4o to extract critical field-level parameters from dispersed literature, comparing model output against a dataset labeled by domain experts. Through iterative prompt engineering, it reached a true positive rate of about 84 percent, while cutting cost per extraction by an order of magnitude and processing each document in seconds rather than the hours manual extraction requires.

57Blocks contribution

57Blocks contributed engineering and cloud infrastructure to the study. The research was led by Stanford University with collaborators at the University of Pittsburgh and Aramco research centers, and was published open access under a Creative Commons license.

Why it matters to clients

The same capability that recovers structured data from messy documents, validated against expert ground truth, is close to the work Excavator does when it recovers requirements from a legacy system, and to the data labeling practice we run at scale. Applied AI research like this is where our delivery engine and our domain skill meet.

Related Where this work connects

Have a data
extraction problem?

We build extraction systems validated against expert ground truth. Tell us the corpus.