About this essay
- Authors: Michael Rowe 1 (ORCID; mrowe@lincoln.ac.uk) and Wesley Lynch 2 (LinkedIn; wesley@snapplify.com)
- Affiliations: 1) University of Lincoln; 2) Snapplify
- Created: Nov 08, 2025
- Version: 0.9 (last updated: Feb 12, 2026)
- Keywords: AI-forward, artificial intelligence, context engineering, curriculum development, curriculum infrastructure, education technology, graph database, human-AI collaboration, model context protocol, vector database
- License: Creative Commons Attribution 4.0 International
Abstract
Professional curricula are comprehensively documented but not systematically queryable, creating artificial information scarcity. Regulatory compliance reporting consumes weeks of staff time, quality assurance requires exhaustive manual verification, and curriculum office teams cannot efficiently answer structural questions about the programmes they maintain. Current approaches — manual document review, VLE keyword search, curriculum mapping spreadsheets, and purpose-built curriculum management systems — fail to expose curriculum structure in queryable form. We propose an architecture where graph databases become the source of truth for curriculum structure, with vector databases for content retrieval and the Model Context Protocol providing accessible interfaces. This makes documented curriculum structure explicitly queryable — prerequisite chains, competency mappings, and assessment coverage — enabling compliance verification in hours rather than weeks. The architecture suits AI-forward institutions: those treating AI integration as ongoing strategic practice requiring active engagement with evolving technologies. Technology handles structural verification; educators retain essential authority over educational meaning-making. The proposal argues for removing technical barriers to interrogating curriculum complexity rather than eliminating that complexity through technological solution. By transforming documentation into operational architecture, institutions can build trust and auditability into AI-mediated processes, ensuring that curriculum design and improvement are driven by analytical insight rather than administrative necessity.
Key takeaways
- Curricula are documented but not queryable. The structural knowledge — prerequisite chains, competency mappings, assessment coverage — exists in prose and staff heads, not as relationships that can be traversed. This creates artificial information scarcity where staff must manually extract information from documents they themselves created.
- A three-layer architecture addresses this. Graph databases as source of truth for curriculum structure, vector databases for semantic retrieval, and a Model Context Protocol layer for role-specific access can reduce compliance reporting from weeks to hours.
- The graph inverts the document-centric model. Changes happen once in the graph and all derived artefacts — module specifications, programme handbooks, VLE structures — are generated from it, eliminating the synchronisation failures that currently plague accreditation cycles.
- The same infrastructure supports internal and external quality assurance. Not only regulatory compliance but also systematic checks against an institution’s own design principles, run regularly rather than only at accreditation points.
- Technology supports professional judgement rather than replacing it. The architecture removes technical barriers to interrogating what curricula actually contain; educators retain full authority over what that structure should be and how to interpret it, supporting the competencies required of teachers to design and enact effective feedback and assessment processes (Carless & Winstone, 2023).
Artificial scarcity in curriculum structure
Professional education curricula — in medicine, nursing, allied health, and other regulated disciplines — are comprehensively documented across hundreds of module specifications, teaching plans, and assessments. Yet when curriculum teams must demonstrate competency coverage for regulatory audits, answer questions about prerequisite chains, or identify assessment gaps, current systems require manual document review consuming days or weeks of staff time. This reflects a lack of curriculum agility—the capacity for institutions to act responsibly and rapidly on change and expectations (Brink et al., 2021 Annotation).
The core issue is a mismatch between what exists and what is accessible. Curricula have explicit structure: hierarchical organisation (programmes → modules → sessions → learning outcomes), prerequisite relationships, competency mappings, comprehensive content documentation. This structure exists in documents and staff knowledge, but not as queryable data. Staff who created this structure cannot efficiently access it for verification or compliance reporting. A companion essay examines this pattern in broader institutional terms, arguing that documentation undergoes a category shift — from reference material to operational architecture — when AI agents become consumers of institutional knowledge (Rowe, 2026). This essay focuses on the specific architectural response for professional curricula.
Professional education in regulated disciplines presents this problem in its sharpest form. Regulatory bodies require institutions to demonstrate — with specific evidence — that every mandated competency receives adequate teaching and assessment coverage. The hierarchical structure, typed relationships (prerequisite dependencies, competency mappings, assessment coverage), and compliance requirements make professional curricula naturally suited to graph-based representation where relationships are first-class entities rather than implicit connections buried in prose. Knowledge graphs are increasingly recognised as tools for curriculum design, enabling institutions to map learning objectives to competencies and identify structural gaps (Abu-Salih & Alotaibi, 2024 Annotation), and have been characterised as “pivotal tools for connecting knowledge points across courses” in higher education (Wang et al., 2025).
Current systems make this structural information inaccessible through technical limitations. Document-based approaches cannot expose relationships between curriculum elements. Even purpose-built curriculum management systems focus on documentation workflow rather than enabling structural queries. Staff cannot ask “show me all assessment touchpoints for prescribing safety competency across prerequisite chains” because infrastructure does not support relationship traversal. The structure exists — module specifications list prerequisites, learning outcomes map to competencies — but exists in prose requiring human interpretation rather than computational querying.
This reflects a broader need to move from ad hoc tool use toward curriculum models that leverage learning technologies as a foundational component of higher education infrastructure (Gosper & Ifenthaler, 2013 Annotation). This architecture particularly suits what we term AI-forward institutions — those treating AI integration as ongoing strategic practice requiring active engagement with evolving technologies, rather than fixed deployment of finished solutions. The infrastructure proposed here enables such institutions to control their technological direction while acknowledging the ongoing engagement this requires.
Current approaches and their limitations
Institutions employ various approaches to curriculum management, from commercial systems to ad hoc arrangements across general-purpose tools. All share a fundamental limitation: they capture curriculum content without making curriculum structure queryable.
Commercial curriculum management systems (WorkTribe, Akari, Courseleaf) provide structured storage with defined fields for learning outcomes, assessment methods, prerequisites, and credit values. They support approval workflows, version control, and change management. But even purpose-built systems focus on documentation workflow rather than queryable relationships. They capture that a module has prerequisites without enabling traversal queries like “show me all assessment touchpoints across prerequisite chains for competency X.” Data exists in structured fields, but relationships between entities remain trapped within individual records rather than being queryable across the system. Integration across curriculum management systems, VLEs, and student record systems remains limited, requiring manual cross-referencing.
Where purpose-built systems are absent or insufficient, institutions rely on arrangements across multiple general-purpose tools: SharePoint for document storage, VLEs for teaching delivery, Excel spreadsheets for competency mapping. These spawn shadow systems — unofficial workarounds created by staff adapting to technological inadequacy. Experienced module leads maintain personal spreadsheets tracking prerequisites. Quality assurance teams develop institution-specific search strategies for finding information across repositories. Different programme teams develop incompatible naming conventions — one calls it “clinical competency,” another “professional capability” — making cross-programme quality assurance difficult. This tacit knowledge concentrates in long-serving staff, creating organisational risk when they leave.
Perhaps the most telling example is “the Matrix” — the detailed Excel spreadsheets many institutions maintain mapping curriculum elements to regulatory competencies. Programme directors compile these for accreditation reviews, with rows representing learning outcomes and columns representing competencies. These exist separately from actual curriculum documentation. When modules change or regulatory frameworks update, spreadsheets require manual updating — often by a single staff member who understands the complex cross-referencing. Accreditation reviews reveal outdated mappings, prompting emergency updates consuming weeks under deadline pressure. Spreadsheets cannot represent complex relationships — hierarchical structures, prerequisite chains, conditional dependencies — forcing staff to accept that some relationships simply cannot be represented.
The common thread across all these approaches is that rich structural knowledge exists in human understanding and scattered across incompatible systems, but not in forms enabling systematic verification or efficient compliance reporting. Current systems force staff to become experts in navigating institutional technology rather than focusing on curriculum design and enhancement. Empirical research confirms this misalignment: Hilliger et al. (2024) found that institutional quality assurance processes are often structurally misaligned with continuous curriculum improvement, with existing tools failing to support the analytical capabilities that programme teams actually need.
Proposed architecture
Why graph databases
The solution requires infrastructure supporting queries over typed relationships, not merely content search. To understand why, consider what a quality assurance officer actually needs to do when verifying that a prescribing safety competency has adequate assessment coverage across a programme. They need to identify which learning outcomes address that competency, then trace which assessments test those outcomes, then check whether the modules containing those assessments have appropriate prerequisites. This is a traversal problem — following a chain of typed relationships from competency through outcomes through assessments through modules — and it is fundamentally different from a search problem. Finding content “about prescribing” through keyword or semantic search does not verify structural coverage. It tells you where prescribing is mentioned; it does not tell you whether the structural chain from competency to assessment is complete.
Graph databases make relationships first-class entities that can be traversed, filtered, and aggregated, enabling precisely these structural queries. By integrating diverse information sources and making relationships explicit, knowledge graphs provide a common vision for institutional data (Ehrlinger & Wöß, 2016 Annotation). As Xiao (2023) notes, graph databases “enable complex relationships between entities to be processed effectively” in educational contexts, offering substantial advantages over relational databases as the number and depth of relationship traversals increases.
Three-layer infrastructure
Graph database layer. This layer stores curriculum structure as typed, directional relationships between entities. Programmes CONTAIN modules, modules CONTAIN sessions, sessions ADDRESS learning outcomes, assessments TEST outcomes, outcomes MAP_TO regulatory competencies. Prerequisites are explicit REQUIRES relationships enabling traversal of dependency chains. Each entity and relationship is explicitly typed, making precise queries possible: “show all modules required to reach clinical prescribing competency” traverses REQUIRES relationships; “identify learning outcomes lacking assessment coverage” queries for outcomes without ASSESSES relationships.
Critically, the graph database becomes the source of truth where educators work directly on curriculum structure. When module leads modify prerequisites, add learning outcomes, or adjust competency mappings, they work in graph-native tools. Documents (module specifications, programme handbooks), VLE structures (course hierarchies, content organisation), and compliance reports (accreditation evidence, regulator submissions) are generated from the graph rather than existing as separate artefacts requiring manual synchronisation. This architectural decision inverts current systems where documents are primary and structure must be extracted. Changes happen once in the graph, with derived artefacts automatically updating — addressing the maintenance burden that makes current approaches so fragile.
Temporal versioning extends the graph’s value as an institutional record. By timestamping nodes and relationships — recording when a prerequisite was added, when a competency mapping changed, when a learning outcome was revised — the graph captures curriculum history as well as current state. This supports accreditation evidence (demonstrating how the curriculum evolved in response to regulatory changes), enables identification of stale elements that have not been reviewed for several years, and provides an audit trail that integrates naturally with the governance processes described later in this essay.
Vector database layer. This layer stores curriculum content (lecture materials, reading lists, teaching notes, module specifications) as embeddings enabling semantic search. When staff need to find “all teaching materials related to prescribing errors,” vector databases retrieve semantically similar content across the entire curriculum regardless of which programme or module contains it. This complements rather than replaces graph-based queries. Semantic similarity finds content about similar topics; typed relationships enable structural traversal. A quality assurance officer might use semantic search to find all prescribing-related content, then use graph queries to verify whether that content adequately maps to regulatory competencies and has sufficient assessment coverage.
Model Context Protocol layer. This provides the accessibility layer enabling natural language queries and stakeholder-specific interfaces. Staff can ask “show me competency coverage for Year 2” in plain language rather than learning database query syntax. System prompts provide appropriate context — curriculum office staff receive compliance-focused responses with audit trails, module leads receive pedagogical guidance about prerequisite assumptions, quality assurance teams receive gap analysis highlighting structural issues.
MCP is relatively new (Anthropic, 2024) but addresses a genuine standardisation need: providing consistent tool interfaces across different AI systems. Even as specific protocols evolve, the principle — standardised semantic interfaces between AI models and institutional data — remains architecturally valuable. This layer is crucial for adoption: without it, queryable infrastructure would only serve database administrators rather than the staff who actually need structural access.
Multi-stakeholder access requires a permissions layer commensurate with the sensitivity and purpose of different access modes. Educators modifying prerequisite structures need different permissions from quality assurance staff running read-only compliance queries, from students browsing learning pathway information, or from external regulators reviewing accreditation evidence. Standard OAuth and single sign-on protocols — already deployed across most institutional identity management systems — provide the authentication foundation. Role-based access controls then define what each actor can query, modify, or export. This is not a novel architectural problem; it is a solved one in enterprise software. Acknowledging it here ensures that access governance is treated as a design requirement from the outset, not retrofitted after deployment.
Ontology considerations
Representing curriculum structure requires an ontology defining entity types (Programme, Module, Session, Learning Outcome, Assessment, Competency) and relationship types (CONTAINS, REQUIRES, ADDRESSES, TESTS, MAP_TO). The Higher Education Reference Model (HERM) provides a pragmatic starting point with standardised vocabulary enabling interoperability across institutions and systems (CAUDIT, 2025; UCISA, 2025).
Ontology choice involves epistemological commitments about knowledge categorisation. HERM embodies particular assumptions about how curricula should be structured — assumptions that may not adequately represent all institutional contexts or disciplinary approaches. A nursing curriculum’s relationship between “clinical placement” and “practice competency” might not map cleanly to HERM’s categories. An integrated medical curriculum where content does not organise neatly into discrete modules might struggle with HERM’s hierarchical assumptions.
What matters, though, is internal consistency rather than adherence to external standards. An institution could use HERM, develop institution-specific ontologies, or adopt programme-specific frameworks. The crucial requirement is that the chosen ontology enables consistent relationship encoding within the institution’s context. Graph databases are schema-flexible — ontologies can evolve as curricula develop and new relationship types become important to represent. The key architectural principle is making relationships explicit and queryable, not imposing rigid structures.
Inverting technological constraints
This architecture inverts the relationship between staff and technology. Current systems force staff to adapt working practices to technological constraints — translating relational thinking into document management, folder hierarchies, and keyword strategies. Graph infrastructure supports how educators naturally conceptualise curricula: as entities in relationship. “This module builds on that one” becomes a queryable REQUIRES relationship. “This learning outcome addresses that competency” becomes a traversable MAP_TO relationship.
Staff can interact with curriculum structure directly rather than through the mediation of document search and manual compilation. A module lead asking “what prerequisite knowledge can I assume?” receives a direct answer from traversing graph relationships rather than requiring them to manually examine multiple module specifications. Quality assurance processes asking “does every clinical module adequately assume pharmacology prerequisites?” become straightforward graph queries rather than days of manual verification. The technology enables staff work rather than dictating it.
This does not eliminate professional judgement — educators still decide what prerequisites should exist, which competencies learning outcomes should address, whether assessment coverage is adequate. The infrastructure makes those decisions explicit, queryable, and verifiable, supporting professional judgement rather than replacing it.
What this enables
Regulatory compliance and quality assurance
The most immediate value proposition addresses the compliance burden. Institutions can generate regulatory compliance reports in hours rather than weeks through straightforward graph queries. A query like “demonstrate that every GMC prescribing safety competency has adequate assessment coverage” traverses ADDRESSES relationships from competencies to learning outcomes, then ASSESSES relationships from outcomes to assessments, returning complete audit trails showing exactly which modules, sessions, and assessments address each competency. The graph structure provides the verifiable evidence regulators require — not semantic similarity suggesting coverage might exist, but explicit typed relationships confirming it does.
Quality assurance processes benefit from systematic gap identification that current approaches cannot achieve. Knowledge graphs enable institutions to “identify gaps in their curriculum” and support “more informed decision-making” through surfacing connections and patterns not apparent through traditional methods (Abu-Salih & Alotaibi, 2024). Wang et al. (2025) demonstrate that graph-based curriculum representations can be evaluated for “accuracy, consistency, completeness and timeliness” — precisely the quality dimensions that manual review struggles to verify systematically. Missing prerequisite chains become visible when REQUIRES relationships show isolated modules with no inbound connections. Learning outcomes lacking assessments appear when querying for outcomes without ASSESSES relationships. Competencies with insufficient coverage emerge when aggregating assessments across MAP_TO relationships reveals imbalances. These structural issues remain invisible in document-based systems — staff might eventually discover them through exhaustive manual review, but systematic identification requires queryable relationships.
The maintenance advantage is substantial. When regulatory frameworks update — GMC adds new prescribing competencies, HCPC revises standards of proficiency — institutions can immediately query which learning outcomes and assessments map to affected competencies, identifying exactly where curriculum updates are required. Current approaches require staff to manually search documents hoping terminology matches, inevitably missing connections. Graph queries definitively answer “what curriculum elements are affected by this regulatory change?”
The same infrastructure supports internal quality assurance against an institution’s own policies and design principles — not only external regulatory requirements. If an institution has programme design guidelines specifying, for example, that clinical assessments should not rely exclusively on summative end-of-year examinations, the graph can be queried to surface modules that diverge from this principle. This functions as a form of internal peer review: a systematic, queryable check against the institution’s own standards that can be run on a regular basis rather than only at accreditation points. It can also flag elements that may warrant review — learning outcomes that have not been updated in several years, or assessment distributions that look atypical across a programme. Importantly, these are signals for educator attention, not automated determinations: the graph surfaces patterns; curriculum teams decide what, if anything, to do about them.
Curriculum design support for educators
Educators gain comprehensive curriculum visibility enabling more effective design and collaboration. A module lead can query “where else is anticoagulation taught?” and receive definitive answers across all programmes and years, revealing integration opportunities or unnecessary duplication. They can ask “what prerequisite knowledge about cardiovascular physiology can I assume?” and see exactly which prior modules addressed relevant learning outcomes and whether assessments verified student attainment.
This supports constructive alignment — ensuring learning outcomes, teaching activities, and assessments genuinely cohere (Biggs & Tang, 2011). An educator designing a clinical prescribing module can verify that learning outcomes align with regulatory competencies, that teaching sessions adequately address those outcomes, and that assessments appropriately test them. When alignment gaps appear — an outcome without corresponding teaching, an assessment testing content not in the learning outcomes — the graph makes them explicit rather than requiring retrospective discovery during curriculum reviews.
Just-in-time access to prerequisite information enables appropriately pitched teaching. Understanding what students have actually studied (and what they have been assessed on) helps educators calibrate expectations. The distinction matters: a topic appearing in Year 1 lectures but not assessed differs from a topic with multiple assessment touchpoints. Graph queries distinguish these, providing educators with actionable intelligence about prior knowledge they can genuinely assume versus content requiring review.
Student access and beyond
While the primary value proposition focuses on staff efficiency and compliance, the same infrastructure enables student-facing benefits. Students could query prerequisite chains directly (“what should I study before cardiology?”), understand competency coverage across their programme, and navigate learning pathways informed by explicit structure. This does not eliminate the need for educator guidance — interpreting why sequences exist, when flexibility is appropriate — but removes artificial barriers to basic structural information that students currently reverse-engineer from scattered documents.
The infrastructure could potentially extend to direct regulator access, enabling bodies like the GMC or NMC to query institutional curriculum structure during accreditation reviews rather than requiring institutions to compile evidence. This would shift compliance from reactive documentation to transparent structural verification. Such access raises questions about institutional autonomy and gaming behaviours that merit separate analysis — we mention it here as a logical extension of queryable infrastructure rather than a primary design goal.
Curriculum as designed versus curriculum as enacted
The graph captures the curriculum as designed: official prerequisites, competency mappings, learning outcomes, and assessment blueprints as documented and submitted for regulatory approval. It does not capture the curriculum as enacted — real-time pedagogical adjustments, informal connections made through teaching, or the tacit judgements about when documented sequences prove flexible in practice. This distinction matters for appropriate expectations. The infrastructure makes formal structure queryable and verifiable; lived pedagogical experience remains in educator-student relationships. These are complementary, not competing: making formal structure explicit frees educators to focus professional judgement on interpretation and flexibility rather than manual information extraction.
Implementation considerations
Technical feasibility and institutional control
The proposed architecture combines technologies at different maturity levels. Graph databases (Neo4j, RDF triple stores) have decades of production use in relationship-intensive domains — this component is genuinely mature and proven. Vector databases and semantic search infrastructure are newer, mainstream only since the large language model surge of the past two to three years, with rapidly evolving implementations. The Model Context Protocol is very recent (released 2024), still developing, with protocol refinements expected. The underlying AI models — whether used for semantic search embeddings or natural language interfaces — are changing rapidly, with new capabilities and architectures emerging regularly.
This technological heterogeneity is an opportunity for AI-forward institutions to control their own direction by choosing technology stacks suited to their contexts and values. An institution could deploy:
- Small open-source models running locally on institutional infrastructure, maintaining complete data sovereignty
- Large frontier models via API for enhanced capabilities, trading some control for performance
- Hybrid approaches using local models for sensitive data and API models for general queries
- Different embedding models optimised for biomedical text versus general educational content
The graph database core remains stable regardless of these choices. When better embedding models emerge, institutions can re-embed content without restructuring the graph. When new AI providers offer improved capabilities, interfaces can switch providers without rebuilding curriculum structure. When MCP evolves or alternative protocols emerge, the abstraction layer can adapt while core infrastructure persists. This architectural separation — stable structure in graphs, evolving intelligence in models — provides genuine flexibility.
This flexibility carries responsibilities. Institutions adopting this architecture commit to active engagement with evolving infrastructure rather than implementing a finished solution. They will need capability to evaluate new models, update embedding approaches, and adapt interfaces as technologies develop. This is not deploy-and-forget infrastructure. Some institutions will embrace this approach, viewing technology choice as strategic capability worth investing in. Others will prefer stable vendor solutions even with less flexibility, outsourcing technology decisions to established providers. Both positions are legitimate; institutions should assess their technical capacity, strategic orientation, and resource availability before committing to infrastructure requiring active stewardship.
Transition pathway
Institutions do not start from nothing. They have existing curriculum management systems, documents in repositories, VLE structures, and spreadsheet-based mappings. The transition involves extracting structured data from current systems, transforming it to graph-compatible format, validating extracted relationships through faculty review, and gradually shifting workflows to graph-native tools.
This requires phased implementation over 18–24 months, following a defined process for knowledge graph development that includes extraction, transformation, and validation (Tamašauskaitė & Groth, 2023 Annotation). Early phases focus on data extraction and populating the graph database while existing systems continue operating normally. Middle phases introduce query interfaces for staff to explore curriculum structure without changing their working practices. Later phases shift curriculum modification workflows to graph-native tools, making the graph the source of truth with documents generated from it. Parallel operation during transition is essential — staff will not abandon familiar systems until new infrastructure demonstrably improves their work.
The phases and timelines for transition are necessarily institution-specific, depending on existing systems, data quality, and technical capacity; the 18–24 month arc described here is indicative rather than prescriptive.
Change management realities
The most significant implementation challenges are human, not technical. Faculty must see clear efficiency gains justifying workflow changes — “this makes my work easier” rather than “the institution requires this new system.” Research on curriculum analytics adoption confirms this: successful implementation requires engaging stakeholders across different phases of design, with demonstrated value at each stage rather than top-down mandates (Hilliger et al., 2024). Curriculum office staff need demonstrated time savings for compliance reporting before committing to learning new tools. Quality assurance teams require confidence that graph queries actually identify issues more reliably than current manual processes.
Success requires strong executive sponsorship ensuring resource allocation, faculty involvement from the beginning (not just as end users), realistic timeline expectations (transformation takes years, not months), and clear communication about benefits at each phase. Pilot implementations demonstrating value to specific stakeholder groups — generating one compliance report in hours that previously took weeks — build momentum more effectively than comprehensive system rollouts.
The organisational change is substantial: shifting from document-centric to structure-centric curriculum management alters how staff conceptualise their work. This is not merely learning new software; it is changing mental models about what curriculum management means.
Maintenance and governance
A key architectural advantage: as the graph becomes the working system, maintenance integrates into normal curriculum development workflow rather than creating additional burden. When educators modify prerequisites, they are updating the graph directly — the source of truth. Derived artefacts (module specifications, programme handbooks, VLE course structures) regenerate automatically. This contrasts with current approaches where changes require updating multiple separate systems and documents, inevitably creating synchronisation failures.
Governance policies must define who can modify curriculum structure, establish approval workflows for significant changes, implement audit logging for all modifications, version the graph state so that changes can be traced and where necessary reversed, and ensure regular validation checks for data quality. Rather than creating parallel governance structures, this approach makes existing governance processes explicit and enforceable through technical infrastructure. The technology supports governance rather than circumventing it.
Data quality becomes an organisational concern rather than a technical one. If staff enter unclear prerequisite relationships or map learning outcomes to incorrect competencies, the graph faithfully represents those decisions — making errors more visible but not preventing them. Clear policies, training, and quality assurance processes ensure the queryable infrastructure contains accurate structural information.
Conclusion
Professional education curricula contain explicit, documented structure — prerequisite dependencies, competency mappings, assessment coverage — that current systems make artificially inaccessible. The consequence is not just inefficiency but a category error: institutions treat compliance and quality assurance as administrative burdens requiring manual labour, when they are fundamentally structural queries that should take hours rather than weeks.
The architecture proposed here — graph databases for curriculum structure, vector databases for content retrieval, MCP for stakeholder access — does not impose new structure on curricula. It makes the structure that already exists visible and queryable. The distinction matters. Educators designed these prerequisite chains, wrote these competency mappings, built these assessment frameworks. The problem has never been that the structure is absent. It is that the infrastructure renders it inaccessible, forcing the people who created it to extract it manually from the documents they themselves produced.
This requires the kind of sustained institutional commitment that goes beyond deploying a new system. It means treating curriculum infrastructure as a strategic investment, developing internal capability to steward evolving technology, and accepting that the transition will take years rather than months. The proposal does not suit every institution, and the choice between building internal capability and relying on vendor solutions is legitimate either way. What is less tenable is the current default — leaving curriculum structure buried in documents and staff knowledge, forcing compliance and quality assurance to depend on manual extraction of information that could be queried directly. For institutions prepared to make the investment, graph-based curriculum infrastructure offers a path from administrative burden to institutional capability.
References
Abu-Salih, B., & Alotaibi, S. (2024). A systematic literature review of knowledge graph construction and application in education. Heliyon, 10(3), e25383. https://doi.org/10.1016/j.heliyon.2024.e25383 Annotation
Anthropic. (2024, November 25). Introducing the Model Context Protocol. https://www.anthropic.com/news/model-context-protocol
Biggs, J. B., & Tang, C. (2011). Teaching for quality learning at university (4th ed.). Society for Research into Higher Education & Open University Press.
Brink, S., Ehrlinger, L., & Wöß, W. (2021). Curriculum Agility: Responsive Organization, Dynamic Content, and Flexible Education. IEEE Frontiers in Education Conference (FIE). https://doi.org/10.1109/FIE49875.2021.9637255 Annotation
Carless, D., & Winstone, N. (2023). What feedback literate teachers do: an empirically-derived competency framework. Higher Education, 86(6), 1433–1451. https://doi.org/10.1007/s10734-022-00979-5
CAUDIT. (2025). Higher Education Reference Models (HERM), version 3.2.0. https://www.caudit.edu.au/communities/caudit-higher-education-reference-models/
Ehrlinger, L., & Wöß, W. (2016). Towards a Definition of Knowledge Graphs. SEMANTiCS. Annotation
Gosper, M., & Ifenthaler, D. (Eds.). (2013). Curriculum models for the 21st century: using learning technologies in higher education. Springer Science & Business Media. Annotation
Hilliger, I., Miranda, C., Celis, S., & Perez-Sanagustin, M. (2024). Curriculum analytics adoption in higher education: A multiple case study engaging stakeholders in different phases of design. British Journal of Educational Technology, 55(3), 785–801. https://doi.org/10.1111/bjet.13374
Rowe, M. (2026). Documentation becomes infrastructure when AI agents are the readers. https://michael-rowe.github.io/home-michael/Essays/documentation-as-infrastructure/
Tamašauskaitė, R., & Groth, P. (2023). Defining a Knowledge Graph Development Process Through a Systematic Review. ACM Computing Surveys. https://doi.org/10.1145/3592624 Annotation
UCISA. (2025). Higher Education Reference Models. https://www.ucisa.ac.uk/groups/enterprise-architecture/herm
Wang, Q., Hou, S., Wan, S., Feng, X., & Feng, H. (2025). Applying Knowledge Graph to Interdisciplinary Higher Education. European Journal of Education, 60(2), e70078. https://doi.org/10.1111/ejed.70078
Xiao, G. (2023). A Personalized Learning Path for French Study in Colleges Based on a Big Data Knowledge Map. Scientific Programming, 2023(1), 4359133. https://doi.org/10.1155/2023/4359133