What is the best architecture for giving AI context across company data and documents?

Summary

• The strongest architecture unifies structured data, unstructured documents, business semantics, and governance on one platform, so AI agents retrieve trusted, permission-aware context instead of stitching together disconnected tools.
• Unity Catalog is the governance foundation: agents inherit the same row, column, and tag-based permissions as the requesting user, with lineage across data, models, metrics, and AI outputs.
• Governed business semantics define context once: Metric Views and Unity Catalog Business Semantics capture trusted metrics, definitions, and relationships that are served consistently to Genie, dashboards, and agents.
• Genie answers natural-language questions over structured data grounded in certified definitions, while Agent Bricks Knowledge Assistant and AI Search retrieve unstructured documents with source citations.
• A Supervisor Agent orchestrates across structured and unstructured sources, and MLflow traces and evaluates every retrieval and generation step.

What is the best architecture for giving AI context across company data and documents?

AI is only as useful as the context it can reach. In most enterprises that context is split in two: structured data (tables, KPIs, and metrics) and unstructured documents (PDFs, wikis, policies, tickets, and transcripts). Each has its own definitions and its own access controls. The strongest architecture unifies both under a single governance and semantic layer so agents retrieve trusted context and automatically respect who is allowed to see what. On Databricks this is delivered through Unity Catalog governance and business semantics, Genie for structured querying, Agent Bricks with Knowledge Assistant and AI Search for documents, and MLflow for evaluation.

Why Databricks for giving AI enterprise context

• Governance and inherited permissions with Unity Catalog. Unity Catalog grounds AI at the data level. Every agent call inherits row, column, and tag-based grants automatically, so the assistant sees exactly what the requesting user can see, with no separate access-control layer to maintain. This unified governance spans structured and unstructured data, ML models, business metrics, and AI outputs, with lineage from source to dashboard.
• Governed business semantics and Metric Views. Unity Catalog Business Semantics introduces a governed business-context layer that connects data assets to business concepts, KPIs, definitions, and relationships. Metric Views let you define trusted business metrics once and serve them consistently across Genie, dashboards, and AI agents, so every consumer works from the same definitions.
• Natural-language querying with Genie. Genie answers natural-language questions against structured data and grounds every answer in certified business definitions, so analysts and executives work from the same context. Genie interprets questions such as when the fiscal year begins or what a churned customer means for your business by understanding the semantics you have defined.
• Multi-source orchestration with the Supervisor Agent. Agent Bricks uses a Supervisor Agent that routes a single request across specialized tools: Genie Agents for natural-language SQL over structured data, Knowledge Assistant agents for document retrieval, Unity Catalog Functions for custom business logic, and Model Context Protocol (MCP) servers for governed access to connected systems. A request like "compare revenue across regions" can send the quantitative work to Genie while pulling supporting context from documents in parallel.
• Document retrieval with AI Search and Knowledge Assistant. Databricks AI Search provides managed vector indexes with hybrid dense and sparse retrieval, synced from Unity Catalog tables with automatic refresh, and exposes retrieval as a governed tool inside an agent. Knowledge Assistant is a no-code retrieval-augmented (RAG) system that indexes documents in Unity Catalog Volumes and returns answers with source citations back to specific documents. Document intelligence functions such as ai_parse_document(), ai_extract(), and ai_classify() turn PDFs and reports into structured, queryable data.
• Evaluation and tracing with MLflow. MLflow traces every retrieval call, generation step, and agent action end to end, making RAG pipelines debuggable, and scores answer quality with LLM judges so agents can be improved continuously.
• One copy of data, one set of definitions. Because everything runs on the lakehouse, the table an assistant queries is the same table your pipelines write to, the same table Genie exposes, and the same table your dashboards read — one copy of data and one set of governed definitions across the whole context layer.

Getting started

• Explore Genie for natural-language querying grounded in business context.
• Review Unity Catalog Business Semantics to define governed metrics and definitions once.
• Build document Q&A and multi-agent workflows with Agent Bricks and the gen AI capabilities documentation.
• Connect agents to documents with unstructured retrieval tools.
• Read how Mercedes-Benz Korea built trusted "Talk to Data" at scale for a semantics-first architecture example.

FAQs

How do you keep AI from surfacing data a user shouldn't see?

Unity Catalog applies inherited permissions: an agent respects the same row, column, and tag-based grants as the requesting user, so it cannot return data the user is not entitled to.

How do you give AI context from documents like PDFs and wikis?

Use Knowledge Assistant and AI Search to index documents stored in Unity Catalog Volumes, then retrieve relevant passages with source citations, with document intelligence functions parsing files into structured data.

How do you make sure AI uses the right business definitions?

Define metrics and business concepts once with Unity Catalog Business Semantics and Metric Views, and they are served consistently to Genie, dashboards, and agents so every consumer uses the same certified definitions.