Pathway map
From knowledge graphs to adaptive semantics: the pathway map
How does a DITA and iiRDS knowledge architecture get from governed semantics to a model that learns from AI agent use — and who does what on the way?
A pathway map reworks a whole work into the logic of a pathway card: lanes for its domains, stations with before, after and maturity, and one route per reader role. Where do you start?
The short answer
This pathway map reworks Michael Iantosca’s white paper “From knowledge graphs to adaptive semantics” into stations and routes. It lays out seven lanes, from DITA content and iiRDS delivery to ontology evolution and governance, and one route each for technical writers, technical marketers, technical project managers and documentation systems developers. Every route runs through governance: only validated, approved changes reach the enterprise model.
For: Technical writers, technical marketers, technical project managers and documentation systems developers
Key points
- Inference extends what the knowledge graph knows; ontology evolution extends how the graph can represent knowledge.
- The adaptive layer is a second cycle around the governed semantic core, not a replacement for it.
- Most stations are building blocks available today; learning from agent use and proposing typed changes is research as of October 2026.
- Every route passes the shared trunk: governance approval and the enterprise model.
- No loop, view or agent adapts away safety-related or required information for use.
Where do you start?
Pick the sentence closest to where your organisation stands; it leads to the first station that matters for you.
- Our documentation is still documents and chapters.Station: Typed topics and maps
- We write DITA and want delivery metadata.Station: Information types and use cases
- We have a knowledge graph and want AI agents on it.Station: Effective semantic view
- Our agents run, and we want the model to learn from them.Station: Agent experience as evidence
Beyond the paper · Course
Pathway map practice course
This is not one of the AI TechDoc mini courses. It belongs to Michael Iantosca’s white paper and walks you through the structure we built around it — the start points, what never changes, the evidence, the shared trunk, the reader routes and the paper’s loop — with tasks and their solutions.
- The white paperMichael Iantosca’s text, unchanged
- Beyond the paperOur map around it
- Your courseStart points and chapters
- TasksEach with its solution
Opens in a learning window with the map, explanations, terms and context.
The map
Lanes run left to right; each route is numbered and drawn with its own line pattern. Stations of the shared trunk are framed.
Lanes
Content (DITA)
Delivery context (iiRDS)
Product world
Knowledge graph and inference
Retrieval and AI agents
Ontology evolution
Governance
- Content (DITA)
- Topics, maps, information types, metadata and product applicability: the structured content itself.
- Delivery context (iiRDS)
- Information types, use cases, user roles and delivery metadata that tell a system what a piece of content is for.
- Product world
- Products, systems, components, procedures, safety requirements and engineering relationships.
- Knowledge graph and inference
- Digital twins per ontology, joined into one graph; inference derives facts inside the known vocabulary.
- Retrieval and AI agents
- The effective semantic view, the vector index and the agents that answer and act on it.
- Ontology evolution
- The adaptive semantic layer: agent experience as evidence, semantic gaps, candidate changes and their evaluation.
- Governance
- The shared trunk: approval, the enterprise model and the versioned semantic core.
Routes
One route per reader role. Each stop says what the role does there; every route runs through the shared trunk.
Route 1
Technical writers
Write and structure the content the graph is built from
Your route starts where your work already is: typed topics and their metadata. It shows what changes when that content becomes part of a graph that agents use — and what stays your decision.
- Typed topics and maps
Write topics as self-contained units of one information type, so each can be found and linked on its own.
- Metadata and product applicability
Declare the products, variants and audiences each topic applies to; this is what connects your content to the product world.
- Information types and use cases
Classify the units with iiRDS information types and use cases, so a consumer knows what each one is for.
- Safety requirements and engineering relationships
Link safety messages to the safety requirements they come from, so an agent can follow the link instead of guessing.
- The ontology as semantic contract
Learn the terms of the contract — which concepts and relationships the ontology defines — so your metadata uses them instead of inventing synonyms.
- Seven signals of a semantic gap
Report what readers keep asking that the content can’t answer, and say whether the content or the vocabulary is missing.
- Governance approvalShared trunk
Bring the content view into the decision: what a proposed concept would mean for topics, metadata and translations.
- Only approved changes reach the enterprise modelShared trunk
Update topics and metadata only against approved changes, never against candidates.
Route 2
Technical marketers
Explain the value of structured semantics without overselling it
Your route goes from what a shared vocabulary buys an organisation to what adaptive semantics can and cannot promise. You explain; you don’t promise what the stations marked research can’t deliver yet.
- The ontology as semantic contract
Explain the ontology as a common language that lets content, products, metadata, AI and business processes understand one another — not as a taxonomy.
- Delivery metadata and user roles
Show what delivery by role and situation means for customers: the right unit for the right person, not more content.
- Unified knowledge graph
Show the difference a connected twin makes: from keyword search to following relationships from a user’s situation to the applicable content.
- AI agents on the graph
Describe what agents do on the graph in concrete tasks — troubleshooting, assembling procedures, retrieving evidence — not as general intelligence.
- Agent experience as evidence
Present agent use as a new source of insight for modelling, and say plainly that this part is research.
- Governance approvalShared trunk
Make governance part of the message: adaptive doesn’t mean uncontrolled.
- Only approved changes reach the enterprise modelShared trunk
Promise only what reached the enterprise model; label everything else as a candidate or an experiment.
Route 3
Technical project managers
Plan the two cycles as one governed process
Your route goes through the decisions and dependencies: who supplies evidence, who proposes, who evaluates and who approves — and in which order. PLM, R&D and purchasing managers take this route too.
- Products, components and procedures
Clarify ownership: engineering owns the product ontology, documentation reads it, and both know who changes what.
- Unified knowledge graph
Map which systems feed the graph — CCMS, PLM, delivery — and who owns each pipeline.
- Agent experience as evidence
Agree which agent and user evidence is recorded, for how long, and who may analyse it.
- Candidate change (typed edit)
Set up how candidates are raised, typed and documented, and who may raise them.
- Evaluation: does agent performance improve?
Plan evaluation and validation as separate steps with their own criteria and owners.
- Governance approvalShared trunk
Establish the governance decision: who approves, on which evidence, and how a rejection is recorded.
- Only approved changes reach the enterprise modelShared trunk
Keep the promotion path fixed and auditable from discovery to the enterprise model.
- Versioned semantic core
Make version numbers visible in releases, so every downstream system can say which model it uses.
Route 4
Documentation systems developers
Build the information system from start to finish, with its review and update loops
Your route goes through the system: graphs, inference, the view, the index and the loop that improves them. You configure, check and hand over; the route says what, not how to code it.
- Digital twins per ontology
Set up one graph per ontology and the pipelines that fill it from the CCMS, the iiRDS packages and the product data.
- Inference within the known vocabulary
Configure the rules and decide which inferred facts are materialized and which are computed on demand.
- Effective semantic view
Define the view the vectorization receives — governed graph, validated additions, selected inferences — and record its version.
- Enriched vector index
Build the index from that view and log which version each build used.
- Candidate change (typed edit)
Give candidate changes a place outside the official model, for example separate named graphs, so they can be tested without touching it.
- Governance approvalShared trunk
Hand candidates over to governance with their evidence and evaluation results; the system never promotes anything by itself.
- Only approved changes reach the enterprise modelShared trunk
Promote approved changes into the official model through the fixed path.
- Incremental re-indexing
Re-embed only the material an approved change affects and keep the index in step with the model version.
Stations
Each station is a small pathway: where practice starts, where it can get to, what does not change, how mature it is as of the review date, and the evidence it leaves.
What the maturity labels mean
- Building block available
- The building block exists today — an established method, a standard practice or a common feature of the tools. You still have to set it up and run it in your organisation; it is not a finished product and not proof of compliance.
- Product-specific
- Some products or platforms offer it, others don’t. Check what yours does before you rely on it.
- Draft or proposal
- Described in a draft standard, a proposal or an individual draft. It may still change or never be adopted.
- Research
- Proposed in research or shown in prototypes. Not ready for everyday operation.
Content (DITA)
Content (DITA)
Building block availableTyped topics and maps
- Before
- Content lives in documents and chapters; a system can only retrieve whole files.
- After
- Content is split into typed DITA topics (task, concept, reference) assembled by maps, so each unit can be found and reused on its own.
- What does not change
- Writers still decide what a topic says; typing structures content, it doesn’t write it.
- Evidence
- Topic and map inventory with their information types
On routes:1 · Technical writers
Content (DITA)
Building block availableMetadata and product applicability
- Before
- Which product or variant a topic applies to is implied by the manual it sits in.
- After
- Every topic carries metadata on the products, variants and audiences it applies to, so a graph can connect it to the product world.
- What does not change
- Applicability is declared by people who know the product; the graph only carries it.
- Evidence
- Metadata scheme and its coverage per topic set
On routes:1 · Technical writers
Delivery context (iiRDS)
Delivery context (iiRDS)
Building block availableInformation types and use cases
- Before
- A delivery system knows file names, not whether something is a maintenance task or a safety notice.
- After
- Content is classified with iiRDS information types and use cases, so a consumer knows what each unit is for.
- What does not change
- The iiRDS standard stays curated by its consortium; local extensions don’t change its core.
- Evidence
- iiRDS package with classified information units
Background reading: “EvoOntology vs. iiRDS: understanding two divergent ontology paradigms” (Saina Veigel, AI TechDoc Press, September 23, 2026) — why a curated standard and an adaptive semantic layer play different roles.
On routes:1 · Technical writers
Delivery context (iiRDS)
Building block availableDelivery metadata and user roles
- Before
- Content is delivered the same way to everyone.
- After
- iiRDS delivery metadata and user roles say who needs which unit in which context, so delivery and agents can filter by situation.
- What does not change
- User roles describe information needs; they don’t grant access rights.
- Evidence
- Delivery metadata in the iiRDS package
On routes:2 · Technical marketers
Product world
Product world
Building block availableProducts, components and procedures
- Before
- Product structure lives in PLM and ERP; documentation refers to it by name.
- After
- A product or domain ontology describes systems, components and procedures, so content and product instances can be linked.
- What does not change
- Engineering owns the product ontology; documentation reads it and doesn’t redefine it.
- Evidence
- Product ontology with its classes and the instance data it is filled from
On routes:3 · Technical project managers
Product world
Building block availableSafety requirements and engineering relationships
- Before
- Safety requirements are cited in documents; their link to components and procedures is implicit.
- After
- Safety requirements and engineering relationships are part of the governed product and safety ontologies, so an agent can follow them.
- What does not change
- Safety requirements come from the risk assessment and the applicable law; no ontology and no agent creates or relaxes them.
- Evidence
- Traceable links between safety requirements, components and procedures
On routes:1 · Technical writers
Knowledge graph and inference
Knowledge graph and inference
Building block availableThe ontology as semantic contract
- Before
- Each system interprets the same data its own way.
- After
- The governed ontology, expressed in standards such as RDF, RDFS, OWL, SHACL and SKOS, is the contract that tells every system what the data means — independent of Python, Java or any other tool.
- What does not change
- The model defines what may exist; the graph records what does exist. Neither replaces the other.
- Evidence
- Versioned ontology files and their validation rules
Knowledge graph and inference
Building block availableDigital twins per ontology
- Before
- DITA, iiRDS and product data sit in separate stores.
- After
- Each ontology supports its own knowledge graph — a digital twin of content, delivery context or product — filled by data pipelines, content processing and APIs.
- What does not change
- A digital twin represents the environment; it doesn’t decide what is correct in it.
- Evidence
- Populated graphs with their source pipelines
On routes:4 · Documentation systems developers
Knowledge graph and inference
Building block availableUnified knowledge graph
- Before
- Keyword search retrieves documents.
- After
- The twins are connected into one graph in which a procedure, a component and a user context are one hop apart, so an agent can follow relationships from a situation to the applicable content.
- What does not change
- The unified graph is only as good as the links that people and pipelines put into it.
- Evidence
- Cross-links between the content, delivery and product graphs
On routes:2 · Technical marketers3 · Technical project managers
Knowledge graph and inference
Building block availableInference within the known vocabulary
- Before
- Every fact has to be stored explicitly.
- After
- An inference engine derives facts from relationships and rules — Pump P410 is part of the engine because it is part of the cooling system — and can materialize them in the graph.
- What does not change
- Inference doesn’t invent concepts; it applies the vocabulary that is already defined.
- Evidence
- Rules and materialized inferences, each with the rule that produced it
On routes:4 · Documentation systems developers
Retrieval and AI agents
Retrieval and AI agents
Product-specificEffective semantic view
- Before
- The vector index is built from whatever the pipeline happens to receive.
- After
- A defined view combines the governed digital twin graph, validated additions from ontology evolution and selected materialized inferences; it is the source of the embeddings.
- What does not change
- The governed core stays stable; the view is what is presented, not what is approved.
- Evidence
- Definition of the view and the graph versions it draws on
On routes:4 · Documentation systems developers
Retrieval and AI agents
Building block availableEnriched vector index
- Before
- Embeddings carry only the text they were made from.
- After
- Concepts, relationships, entity facts and annotations from the effective view enrich the embeddings, so semantic retrieval finds more of what users mean.
- What does not change
- Embeddings approximate meaning; they don’t preserve formal relationships, constraints or rules — reasoning stays with the graph.
- Evidence
- Index build log with the view version each build used
On routes:4 · Documentation systems developers
Retrieval and AI agents
Product-specificIncremental re-indexing
- Before
- Every semantic change means rebuilding the whole index.
- After
- An approved change is versioned, the affected entities and graph neighborhoods are identified, and only that material is re-embedded.
- What does not change
- Re-indexing follows approval; it never runs ahead of it.
- Evidence
- Change version, affected neighborhood and re-embedding record
On routes:4 · Documentation systems developers
Retrieval and AI agents
Product-specificAI agents on the graph
- Before
- Agents search documents and paraphrase what they find.
- After
- Agents use the graph, the inference engine and the vector index to answer questions, troubleshoot equipment, assemble procedures and retrieve supporting evidence.
- What does not change
- An agent’s answer serves the source; it never replaces the required information for use.
- Evidence
- Agent answers with the graph paths and sources they used
On routes:2 · Technical marketers
Ontology evolution
Ontology evolution
ResearchAgent experience as evidence
- Before
- The ontology changes only through modelling workshops.
- After
- How agents and users actually use the model is recorded as evidence of where it is weak or incomplete.
- What does not change
- Evidence informs modelling decisions; it doesn’t make them.
- Evidence
- Logged questions, answers and agent behaviour, kept for analysis as agreed
Background reading: “EvoOntology vs. iiRDS: understanding two divergent ontology paradigms” (Saina Veigel, AI TechDoc Press, September 23, 2026) — why a curated standard and an adaptive semantic layer play different roles.
On routes:2 · Technical marketers3 · Technical project managers
Ontology evolution
ResearchSeven signals of a semantic gap
- Before
- A missing answer is treated as missing data.
- After
- Recurring patterns — incomplete answers, unmodeled concepts, relationships rebuilt again and again, query workarounds, missing classifications, tooling gaps, consistently weak areas — point to a vocabulary that can’t express what users ask.
- What does not change
- Not every gap is semantic; some answers are missing because the content is.
- Evidence
- Gap reports with the signals that raised them
On routes:1 · Technical writers
Ontology evolution
ResearchCandidate change (typed edit)
- Before
- Technicians keep asking “Which procedure after replacement?”, and the model has no way to express it.
- After
- The adaptive layer proposes a typed edit — a concept, relationship, rule, annotation or tool — such as recommendedAfterReplacement or Operational Criticality, without changing the official model.
- What does not change
- A candidate is a proposal; the governed core stays untouched until approval.
- Evidence
- Candidate change with its evidence and the proposed definition
On routes:3 · Technical project managers4 · Documentation systems developers
Ontology evolution
ResearchEvaluation: does agent performance improve?
- Before
- Changes are judged by whether they look right.
- After
- Each candidate is tested against the question whether agent performance improves with it, and validated before it goes to governance.
- What does not change
- A better score on agent tasks is neither a safety argument nor proof of correctness.
- Evidence
- Evaluation and validation results per candidate
On routes:3 · Technical project managers
Governance
Governance · Shared trunk
Building block availableGovernance approval
- Before
- Changes reach the model when someone with access makes them.
- After
- Only validated candidates go to a governance decision; DITA, iiRDS, product, safety and regulatory ontologies change only through it.
- What does not change
- AI can help discover what the ontology may be missing; governance determines what becomes part of the enterprise ontology.
- Evidence
- Decision record: approved, rejected or deferred, by whom and on which evidence
On routes:1 · Technical writers2 · Technical marketers3 · Technical project managers4 · Documentation systems developers
Governance · Shared trunk
Building block availableOnly approved changes reach the enterprise model
- Before
- Experimental additions leak into production.
- After
- The promotion path is fixed: adaptive discovery → candidate change → evaluation → validation → governance approval → enterprise model.
- What does not change
- The enterprise model stays the trusted foundation; the adaptive layer around it can be switched off without losing it.
- Evidence
- Promotion log per change
On routes:1 · Technical writers2 · Technical marketers3 · Technical project managers4 · Documentation systems developers
Governance
Building block availableVersioned semantic core
- Before
- Nobody can say which model version an answer was based on.
- After
- The governed core is controlled, stable and versioned; graphs, the view and the index record which version they use.
- What does not change
- Versioning documents a change; it doesn’t approve it.
- Evidence
- Version history of the ontologies and of the graphs built on them
On routes:3 · Technical project managers
Interchanges
Where one lane depends on another.
Metadata and product applicabilityDelivery metadata and user roles
DITA applicability and iiRDS delivery metadata describe the same unit from two sides: what it applies to, and who needs it when.
Metadata and product applicabilityProducts, components and procedures
Applicability values point to products and components of the product ontology; that link is what the unified graph follows.
Safety requirements and engineering relationshipsGovernance approval
Changes that touch safety or regulatory ontologies always go through governance; the adaptive layer never edits them directly.
Inference within the known vocabularyEffective semantic view
Selected materialized inferences become part of the view the embeddings are made from.
AI agents on the graphAgent experience as evidence
What agents do on the graph becomes the evidence the evolution cycle learns from.
Evaluation: does agent performance improve?Governance approval
Only evaluated and validated candidates reach a governance decision.
Only approved changes reach the enterprise modelIncremental re-indexing
An approved change triggers the re-indexing of the material it affects.
Only approved changes reach the enterprise modelThe ontology as semantic contract
An approved change becomes part of the contract every system works with.
The loop in five moves
- Learn — how agents use the model
- Detect — gaps the ontology cannot express
- Propose — a concept, relationship or rule
- Evaluate — does agent performance improve?
- Govern — promote validated changes
The eight-step improvement cycle
- Users interact with the AI agent.
- The agent uses the knowledge graph and inference engine.
- The agent encounters semantic limitations.
- The adaptive layer identifies and analyses those limitations.
- New concepts, relationships, rules, annotations, or tools are proposed.
- Changes are evaluated and governed.
- Approved improvements become available to the knowledge graph and AI agents.
- Agent performance improves, while new usage creates additional learning opportunities.
The loop in five moves and the eight-step improvement cycle are Michael Iantosca’s, quoted from his white paper “From knowledge graphs to adaptive semantics”.
Limits of this map
This map reworks one white paper into stations and routes; it is neither an implementation guide nor a product comparison.
The stations marked research — agent experience as evidence, the signals of a semantic gap, candidate changes and their evaluation — describe an approach (EvoOntology) that is not an established building block as of October 2026. Product-specific means it depends on the graph database, the embedding pipeline or the agent platform in use. The maturity labels are the editors’ assessment as of the review date, not statements of the white paper.
The context
Why a map and not a single path
A pathway card shows one way from a before to an after. The white paper describes a whole architecture: DITA content, iiRDS delivery metadata and a product ontology feed digital twins, the twins join into one knowledge graph, inference and a vector index serve AI agents, and a second cycle around the governed core learns from what the agents need. No single before and after covers that. The map keeps the logic of the pathway card — before, after, what does not change, maturity, evidence — and applies it station by station.
How to read it
Lanes are the domains of the paper, from content to governance. Stations are states and building blocks; each says where practice usually starts, where it can get to and what does not change. The routes are the same landscape seen by four reader roles: each stop says what that role does there. Interchanges show where one lane depends on another. The loop in five moves and the eight-step improvement cycle are quoted from the paper.
Start with “Where do you start?”, follow your route, and open the stations on the way. The role cards on this map explain the functional roles the paper relies on — what has to be done, independent of job titles.
What the map does not claim
The map describes; it doesn’t prescribe a product or an order of purchase. The stations marked research are not established practice yet. And whatever an agent learns, the risk assessment and the required safety information stay where they are: with the people responsible for them.
Questions readers ask next
- Who wrote the content this map is based on?
- The architecture, the loop and the improvement cycle are Michael Iantosca’s, from his white paper “From knowledge graphs to adaptive semantics”. Saina Veigel’s article on EvoOntology vs. iiRDS partly inspired the paper. The map, its stations, routes and maturity labels were drawn by the editors of knowledge.aitechdoc.world.
- Why are there exactly four routes?
- One route per reader role keeps the map readable: technical writers, technical marketers, technical project managers and documentation systems developers. What every role must pass is drawn once, as the shared trunk.
- Does EvoOntology replace the knowledge graph or the vector index?
- No. In the paper it replaces neither the digital twin graph, the inference engine nor the vector index; it adds an adaptive semantic layer that can improve all three, and only approved changes reach the enterprise model.
Sources
- From knowledge graphs to adaptive semantics: evolving DITA and iiRDS digital twins for AI (white paper) — Michael Iantosca, 25 September 2026
- EvoOntology vs. iiRDS: understanding two divergent ontology paradigms — AI TechDoc Press (Saina Veigel), 23 September 2026
Review log and changes
Every context card is checked against its sources before it is published, and again whenever it changes; the date under the byline is the last review. Corrections (something was wrong) and additions (something was missing) are logged below with date and time (Berlin time). Typos, formatting and link fixes are not listed.
Reviewed
No corrections or additions since publication.