Why Semantically Connected Standards Data Helps LLMs

This page explains the theory behind using structured accessibility standards data (graph links, IDs, schemas, evidence fields) to improve LLM reliability when giving accessibility implementation guidance.

Core Idea

• Instead of relying on model memory, use retrieval from curated, versioned standards artifacts.
• Represent standards relationships explicitly (for example, maps_to, supports_outcome_for, part_of).
• Require traceability in outputs (node IDs, edge/relation, confidence, and source URLs).

Why This Reduces Hallucinations

• Constraint: structured schemas and IDs limit free-form invention.
• Grounding: retrieval-first workflows anchor responses in current source artifacts.
• Auditability: claim-level traceability exposes unsupported recommendations.
• Consistency: shared graph relations reduce contradictory guidance across sessions/projects.

Practical Pattern

1. Retrieve only the relevant standards nodes/edges for the code under review.
2. Generate recommendations linked to those retrieved elements.
3. Return a traceability table for each recommendation.
4. Mark uncertain when evidence is missing.

Known Limits

• This reduces hallucinations; it does not eliminate them.
• Poor retrieval scope or stale sources still degrade output quality.
• Accessibility conformance still requires testing and human review.

Resources and Background Reading

• W3C Semantic Web Standards
• RDF Concepts and Abstract Syntax
• JSON-LD 1.1
• SKOS Simple Knowledge Organization System
• Retrieval-Augmented Generation for Knowledge-Intensive NLP (RAG)
• GraphRAG: Combining knowledge graphs and LLM retrieval (overview)
• WCAG 2.2
• WAI-ARIA 1.2

In This Repository

• Accessibility agent contract
• Canonical standards link graph
• Link graph playbook
• Interactive standards link viewer