Technology / Knowledge Structures
Knowledge Structures
Your RAG pipeline finds the right documents. It cannot tell your AI which ones are still true.
We research how organisations should structure knowledge before it enters an AI system — and how that structure affects what the AI can reliably do with it.
RESEARCH
Active collaboration with MIT on knowledge ingestion and inference optimisation
The problem
Unstructured knowledge produces unreliable AI.
Most enterprise AI projects ingest documents as undifferentiated text. The AI retrieves whatever is most similar — regardless of whether it is current, verified, or relevant in the right way.
We treat knowledge structure as infrastructure. Before retrieval, before generation, before deployment.
Epistemic typing at ingestion
Every document chunk is classified by its epistemic role before storage. A hypothesis and a decision require different retrieval logic — conflating them degrades output quality.
Deterministic decay
Knowledge expires. We assign decay rates at ingestion based on content type and domain. Stale knowledge is flagged before it reaches retrieval, not after it produces a wrong answer.
Contradiction as a first-class signal
Most systems silently average conflicting information. We surface contradictions explicitly so the AI — and the human — can reason about them rather than ignore them.
Inference optimisation
We research how the structure of ingested knowledge affects inference speed and accuracy. Better structure means faster, more precise answers at lower cost.
Taxonomy
Nine epistemic types
We classify every knowledge unit at ingestion. Each type carries different retrieval weight, decay rate, and contradiction sensitivity.
Decision
A resolved choice with documented rationale
Evidence
Verified data that supports or refutes a claim
Hypothesis
An untested proposition under investigation
Observation
A recorded fact without causal interpretation
Open Question
A problem without a current answer
Contradiction
Two pieces of information in conflict
Assumption
A premise accepted without verification
Claim
An assertion awaiting validation
Signal
A weak indicator requiring accumulation before action