Researchers presented the SAGE framework for automating corporate documentation

Researchers presented SAGE — a multi-stage pipeline based on LLM, designed to transform corporate guidelines into structured work artifacts. The system automates data extraction from documents containing text, tables, and images, while ensuring strict version control and data integrity verification.

The SAGE architecture includes a versioned rule repository, a contract validation scheme between processing stages, and a data provenance tracking system. During operation, the extracted rules undergo deterministic structural validation and LLM-based semantic evaluation. The consistency module automatically removes duplicates, identifies contradictions and gaps in specifications, and directs only disputed elements to manual verification.

Testing on a sample of 120 documents showed a reduction in processing time from several days to 20–100 minutes. The system achieved 96% success at the document level with a hallucination rate of 3.2%. For comparison, without using control and verification mechanisms, the error rate increases to 15.7%. During the experiment, 3,896 rules were extracted and 812 ready-made artifacts were formed.

The implementation of such systems allows companies to minimize manual labor when working with technical documentation. The use of Creative Commons license ensures the transparency of the methodology, and the automation of the verification stage reduces the risks associated with inaccuracies in automatic data extraction from complex corporate sources.