Enterprise data lakes accumulate columns faster than human stewards can document or classify them, leaving missing descriptions and unassigned governance labels. The authors frame this as documentation debt that undermines data discovery, access control, and regulatory compliance. They present Glyph, a production system that treats two coupled problems, column description generation and column type annotation, as cooperating LLM agents orchestrated as stateful graphs.

The Descriptor agent grounds generation in the pipeline source code that produces each column. Source code is retrieved on demand from an enterprise GitHub through a reasoning-acting tool loop, which the authors describe as active Retrieval-Augmented Generation. This design means descriptions are derived from the actual code that creates the data, rather than from column values alone. The Tagger agent assigns labels from a governed 275-leaf Data Classification Ontology. It runs three complementary strategies in parallel: a description tagger, a line-of-business regex tagger, and a metadata tagger backed by a fine-tuned contrastive encoder over a vector database. The ranked outputs from these strategies are fused with Reciprocal Rank Fusion (RRF).

The metadata encoder is a 6-layer MiniLM fine-tuned with an in-batch contrastive objective. On an in-distribution held-out split, same-tag retrieval improved from NDCG@10 0.55 to 0.92 and MAP@100 from 0.19 to 0.90 relative to the stock base encoder. End-to-end multi-label tagging quality is reported under a recall-weighted F2 objective across three evaluation groups. An ablation isolates each strategy and the RRF fusion, showing how the components contribute to the final result.

The engineering decisions that distinguish Glyph from prior column-type-annotation work and from commercial value/regex sensitivity scanners include a value-free and code-grounded design, per-tag provenance, and graceful degradation. These properties are what make multi-agent LLM cataloging auditable and operable as a production service.