Apple Releases 'Glyph' to Automate Data Catalog Classification with LLM Agents

Apple has released Glyph, an LLM agent-based system designed to address documentation debt in enterprise data lakes. The system consists of cooperating stateful graphs for generating column descriptions and type annotations for data classification.

Descriptor generates descriptions by searching enterprise GitHub for the pipeline source code that produces each column, while Tagger assigns labels according to a data classification ontology with 275 leaf nodes. Tagger runs three strategies in parallel—description-based, business line regex-based, and metadata-based using a fine-tuned contrastive encoder—before fusing the results with Reciprocal Rank Fusion (RRF).

Notably, fine-tuning a 6-layer MiniLM metadata encoder with a contrastive learning objective improved same-tag retrieval performance (NDCG@10) from 0.55 for the base encoder to 0.92. Unlike value-based scanners, Glyph is designed as a production-ready service supporting code-based design and tag-level provenance tracking for auditability and operations.