Knowledge Layers for Analytics Agents: Architectures, Frameworks, and Open Challenges

Large language model (LLM) agents are increasingly asked to answer questions over enterprise data, yet their failures in that setting are rarely failures of language modeling. They are failures of grounding: the agent does not know what a metric means, which join is legitimate, which table is authoritative, or what it is permitted to see. Between the raw data estate and the agent sits a knowledge layer — the governed definitions, knowledge graphs, catalogs, and access structures that determine what the agent can know and how reliably it can know it. This survey takes the knowledge layer, rather than the agent, as its unit of analysis, and organizes the literature into five branches: the definition layer of semantic and metric models; the knowledge substrate of enterprise and virtual knowledge graphs, catalogs, and metadata graphs; augmentation frameworks that bring structured knowledge into generation, with the Knowledge Augmented Generation (KAG) framework as an anchor case; grounding and consumption mechanisms, centered on the relationship between literature and the Spider to Spider 2.0 benchmark progression; and governance and trust — access control, provenance, freshness, and curation. The synthesis is deliberately critical. Several branches rest on vendor documentation rather than scholarship, the anchor framework's primary record is proprietary, the specializing semantic-layer literature is months old, and the governance branch is thin enough that its absences are among the survey's principal findings. An eighteen-item gap register, each item backed by documented searches, closes the survey as a research agenda ordered by consequence.