Key Takeaways

1. AI needs standardized infrastructure: Croissant, MCP and cataloging standards can reduce fragmentation across datasets, tools and AI applications.
2. Metadata is becoming executable: Modern metadata can define extraction logic, tool interfaces, relationships and operational context rather than simply describing an asset.
3. Enterprise AI requires connected knowledge and governance: AI BOMs, DCAT-3 and GraphRAG can help connect data lineage, asset management, compliance and knowledge retrieval.

Artificial intelligence is becoming more capable, but the infrastructure connecting AI models to data, tools, enterprise systems and governance frameworks remains fragmented.

In this episode of The Deeep Dive, we explore the standards and architectures helping solve this problem across the AI lifecycle. The discussion moves from how AI consumes training data to how AI agents interact with tools, how enterprises govern AI assets, and how knowledge graphs can improve retrieval and reasoning.

The episode examines five important pieces of the emerging AI infrastructure stack: ML Commons Croissant, Model Context Protocol (MCP), W3C DCAT-3, AI Bill of Materials (AI BOM), and Microsoft GraphRAG.

Why AI Needs Better Infrastructure

Large language models can reason over complex problems, generate software, analyze information and perform sophisticated tasks. Yet AI systems can still struggle with something seemingly basic: understanding and connecting to the data and systems around them.

Traditional datasets often depend on custom loading scripts, inconsistent formats, isolated repositories and manually defined relationships. This creates a fragmented AI ecosystem where developers repeatedly build adapters between models, datasets, applications and tools.

The episode explores the shift from passive metadata to active, executable metadata, where metadata can describe data, define relationships, provide extraction instructions, expose tools and support automated AI workflows.

ML Commons Croissant: Making Datasets Machine-Readable

The discussion begins with ML Commons Croissant, a machine-readable metadata format built using JSON-LD and Schema.org.

Croissant adds structured metadata around existing datasets without requiring organizations to completely reorganize their underlying data storage.

The episode explores its four major layers:

• Metadata: Dataset identity, licensing, citation and provenance.
• Resources: Files, archives and storage resources, including integrity information.
• Structure: Instructions for extracting and connecting data from complex resources.
• ML semantics: Information about features, targets, splits and machine learning use.

One of the important ideas discussed is the move from metadata that simply describes a dataset to metadata that can provide executable extraction logic.

Instead of requiring a developer to repeatedly write custom scripts to unzip files, extract identifiers and join datasets, structured metadata can describe how those operations should be performed.

Model Context Protocol: Connecting AI Agents to the Real World

The episode then moves from data ingestion to AI agents and Model Context Protocol (MCP).

AI applications increasingly need access to databases, APIs, repositories, files and other external systems. Without a common protocol, every AI application and every tool can require a separate integration.

MCP addresses this integration problem through a standardized protocol for connecting AI applications with external context and capabilities.

The discussion explores three core MCP concepts:

Resources provide information that an AI application can read as context.

Tools allow an AI system to perform actions through defined interfaces.

Prompts provide reusable, parameterized instructions that can be discovered and used by clients.

The episode also examines tool schemas, human approval for state-changing actions and MCP’s approach to separating protocol communication errors from tool execution errors.

This creates the possibility of AI agents that can receive structured feedback when a tool call fails and use that feedback to correct their next action.

DCAT-3 and Enterprise AI Cataloging

As AI systems scale across enterprises, organizations need ways to discover, classify and manage thousands of datasets, APIs, models and other digital assets.

The episode examines W3C DCAT-3 and related semantic cataloging approaches for organizing these assets.

DCAT-3 expands data cataloging beyond static files by supporting concepts such as Data Services and Dataset Series.

This matters for AI because enterprise data is increasingly accessed through live APIs, services, databases and continuously updated datasets rather than only through downloadable files.

The discussion also looks at the distinction between an intellectual asset and its physical distributions, allowing multiple representations of the same underlying asset to be managed within a structured catalog.

AI Bill of Materials and AI Governance

AI governance becomes increasingly important as organizations need to understand what goes into their models, where their training data came from and how models were developed and deployed.

The episode explores the concept of an AI Bill of Materials (AI BOM) as a machine-readable record of an AI system’s components and characteristics.

Topics discussed include:

• Model architecture
• Training configurations
• Model weights
• Cryptographic checksums
• Training data lineage
• Data transformations
• Licensing information
• Intended use
• Known limitations and anti-patterns
• Fairness considerations
• Energy consumption and environmental information

Together with dataset metadata and enterprise catalogs, an AI BOM can form part of an auditable chain connecting data, models and operational information.

GraphRAG: Moving Beyond Flat Vector Search

The final major topic is GraphRAG, Microsoft’s approach to retrieval augmented generation using knowledge graphs.

Traditional RAG typically breaks documents into chunks, converts those chunks into vectors and retrieves content based on semantic similarity. This works well for many targeted questions, but it can struggle when a question requires relationships across a large corpus.

GraphRAG approaches the problem differently.

An LLM processes source material to identify entities, relationships and claims. These become nodes and edges within a knowledge graph. Community detection can then organize the graph into hierarchical groups, while summaries can be generated for those communities.

The episode explores three GraphRAG search approaches:

Global Search: Designed for questions requiring synthesis across a large corpus.

Local Search: Focused on a specific entity and the relationships surrounding it.

Drift Search: Combines broad thematic exploration with targeted traversal through the knowledge graph.

This allows an AI system to reason over relationships and larger structures rather than relying only on isolated text chunks.

The Bigger Shift: Metadata Becomes Infrastructure

The technologies discussed in this episode may appear to solve different problems.

Croissant addresses dataset ingestion and machine-readable metadata.

MCP addresses connections between AI applications, context and tools.

DCAT-3 addresses cataloging and discovery.

AI BOM approaches AI system transparency and governance.

GraphRAG structures relationships within enterprise knowledge.

The common thread is the transition from passive metadata to active metadata.

Metadata is increasingly becoming part of the operational layer of AI systems. It can define how data is interpreted, how tools are called, how assets are cataloged and how knowledge is retrieved.

For organizations building enterprise AI systems, this means reliable AI infrastructure requires more than a powerful LLM. Data, metadata, tools, governance and knowledge retrieval need to work together.

Listen to the Episode

Explore how the next generation of AI infrastructure is being built around standards, protocols, metadata and knowledge graphs.

From Croissant and MCP to DCAT-3, AI BOM and GraphRAG, this episode examines the infrastructure layer that could determine how reliably AI operates at enterprise scale.