Semantics, RAG & Data Meaning
10 pieces from Duczer East's practitioners on the layer under enterprise AI that nobody owns: what the data means, whether two systems agree on it, and why retrieval that returns the right documents can still produce the wrong answer.
What this topic is about
A data dictionary catalogues fields. A taxonomy classifies. A master-data programme reconciles records. None of them specifies what a record means when an automated system has to act on it, and that gap is where most agentic AI programmes fail. The agents are not badly built. The substrate they reason over never said whether a customer can also be a supplier, what “beneficial owner” includes, or which of two contradictory documents is authoritative.
The pieces in this collection work through that layer from several directions. Ontology and knowledge graphs as reasoning infrastructure rather than storage with edges. Document ingest as the moment provenance is either captured or lost. Retrieval systems that work perfectly and still answer wrong. And the executive-facing version of the same problem: a dashboard or a readiness survey that reports confidently on data whose definitions disagree three systems down.
The two adjacent topics have their own collections. Where the data lives and what the platform has to deliver is in Private AI & Data Architecture; what the agents consuming it are permitted to do is in Agentic AI Governance & Model Risk.
Start here
The thesis: agents do not retrieve data, they reason over it, and the reasoning is bounded by how precisely the data’s meaning has been specified.
-
Why Most Agentic AI Projects Fail Before They Start
The missing foundation is not compute or models. It is a scoped ontology, enforced as constraints, validated against real decisions.
9 min read -
Document Ingest Is a Semantic Event, Not a Parsing Problem
Signer, consent, version and provenance have to be captured as structured signals at the moment a document enters the institution.
8 min read -
Before You Trust an Agent's Decision, Ask What It Resolved
Orchestration vendors made agent behaviour inspectable. Few pushed that inspectability down to what the agent decided a term meant.
7 min read
Meaning at the data layer
Retrieval, dashboards and data-readiness surveys all sit on the same unexamined assumption: that the data underneath agrees with itself.
-
Structure-Aware RAG Is Right. The Benchmarks Aren't.
Embedding document structure into retrieval is the correct move. The evidence being cited for it measures the wrong thing.
8 min read -
ServiceNow Built the Dashboard. The Data Layer Decides If It's True.
An inventory of every AI system in production is only as reliable as the definitions the systems underneath disagree on.
4 min read -
The Cloudera Survey Found a Leadership Gap. A Semantic Gap Sits Beneath It.
Data readiness has an owner in most organizations. Data meaning usually does not.
4 min read
Notes from the field
Shorter reads on developments worth knowing about.
-
When Your RAG System Retrieves Perfectly and Still Lies to You
The failure mode nobody measures: the right documents come back and two of them contradict each other.
3 min read -
For CIOs Weighing "Context Graph" Pitches: Three Questions Before You Sign
Decision threads buried in chat and email hold the why of enterprise activity. What to ask a vendor who wants to mine them.
3 min read -
The Real AI Shift Isn't New Models. It's Control.
On the context layer, and what happens to implementations that stuffed it into oversized prompts.
3 min read -
NatWest's AI Story Is Really a Board-Level Story About Discipline
A bank that accepted the multi-year data work had to come before the AI announcement.
3 min read
Semantic foundations built before the first agent runs
Duczer East scopes and builds the ontology, knowledge-graph and ingest architecture that lets agents reason consistently over enterprise data — and the validation practice that keeps it true in production.