Architecture deep dive
How an AI agent's questions are routed through a governed knowledge layer, and how that resolves into three components: Access Intel, AI Economics Routing Prioritization, and Certified Knowledge Blocks.
End-to-end footprint
Three verticals: unstructured data on the left, structured data on the right, and the Trusted AI Knowledge Layer in the center, with AI consumer agents on top. Grey chips show example technologies for each component. Select a route to trace it, or a component to read about it.
Three routes, one decision
Route 1 always runs first. The knowledge layer then recommends 2a, 2b, or 2c for the question at hand, and the agent follows that recommendation.
Architecture components
Each component maps to a part of the footprint. Together they make the routing recommendation possible.
-
UNSTRUCTURED DATA
-
U-1Enterprise Business Documents
The enterprise's unstructured corpus: policies, contracts, procedures and reports held in document repositories. It is the source for brute-force retrieval (route 2a) when a known set of documents answers a critical question, and Business Document Intake scans and chunks it.
Technology: MS SharePoint -
U-2RAG Pipelines
Pipelines that embed documents, retrieve relevant passages and augment the prompt for AI consumers. Every step carries cost, so the knowledge layer decides when RAG is the right route for a question.
-
U-3Vector DBs
Stores document embeddings for semantic search (route 2b), the economic middle path for very large corpora. Collection metadata such as embedding coverage is shared with the knowledge layer, and contextual metadata from Business Document Intake improves retrieval.
Technology: PostgreSQL
-
-
AI SERVICE LAYER
-
T-1Access Intel
Before any route is taken, the graph resolves who is asking, what the data is, and what they inherit, so the agent only sees what that persona is entitled to.
Enabled by: Collibra + Neo4j; Automated Persona Permissions Framework -
T-2AI Economics Routing Prioritization
The knowledge layer recommends 2a, 2b, or 2c at runtime by weighing criticality, document count, token economics, and certified-metric matches.
Enabled by: Knowledge-Layer Metadata + Token-Economics Model -
T-3Certified Knowledge Blocks
Business logic harvested from BI reports and document slices from enterprise content are certified through the stewardship framework. They become reusable blocks: Certified Context Packages and Certified Data Products, ready to plug into any AI consumer. Each block is built once and reused everywhere, which raises the speed and accuracy of AI responses.
Enabled by: Starburst, Deasy Labs + Collibra; Governance Certification Framework
-
-
INTELLIGENCE COLLECTION & GOVERNANCE LAYER
-
I-1Business Document Intake
Scans enterprise documents for metadata and chunks large documents into smaller slices. It produces Document Slices and Document Metadata, and passes docs metadata and an inventory of doc slices to AI and Data Governance.
Technology: Deasy Labs -
I-2AI and Data Governance
The governance system of record and the human-facing interface where stewards define, approve and certify. It holds the ontology, taxonomy and classifications that give AI consumers business meaning, and feeds governed metadata to the AI Service Layer.
Technology: Collibra, BigID -
I-3Data Assets Intake
Harvests metadata from structured sources: source systems and datalakes, certified data products, and BI reports with their KPIs, metrics, dimensions and business logic. It passes assets metadata to AI and Data Governance.
Technology: Collibra Edge
-
-
STRUCTURED DATA
-
S-1Source Data & Datalakes
Operational source databases and the enterprise datalake where structured data lands. Source DBs feed the datalake, and Data Assets Intake scans both for metadata.
Technology: Google Cloud Platform (GCP) -
S-2Certified Data Products
Reusable data products that calculate KPIs and metrics once, using business rules approved through stewardship. They answer KPI questions precisely through route 2c, without documents or extra context.
Technology: Starburst -
S-3BI & Analytics
Reports and dashboards that consume certified data products. Report scanning harvests the KPIs, metrics and business logic buried in reports so they can be certified and reused.
Technology: PowerBI, Tableau, Looker
-
Technology-neutral: Neo4j or similar (Graph DB) · Starburst or similar ("query-anything-where-it-lives") · Collibra or similar (Data Governance & Metadata Catalog) · Deasy Labs or similar (Docs chunking, cross-file slices) · BigID or similar (auto data classification).