Agenda of SKGi workshop

LLMs are increasingly applied to Knowledge Graph (KG) construction in manufacturing as an alternative to manual, symbolic or neuro-symbolic approaches. However, neural methods including LLMs are naturally opaque and, hence, pose challenges to safety-critical domains such as manufacturing. While verification and validation (V&V) frameworks are common in software engineering, there is no structured V&V approach for KG construction in safety-critical industrial domains. To close this gap, we propose a five-dimensional V&V taxonomy for KG construction in the manufacturing domain. We evaluate our taxonomy on 50 reports from this field and support future work in this field by identifying gaps in reported V&V evaluation steps. This taxonomy offers researchers and practitioners a taxonomy structuring this research field to compare V&V evaluations of neuro-symbolic approaches for manufacturing KG construction.

Meeting climate targets requires accelerating building energy renovations through tailored renovation suggestions. Large Language Models (LLMs) could support this process, reducing the time and effort required from practitioners, provided the LLM’s outputs are grounded in domain expertise. Knowledge Graphs (KGs) provide a structured means of representing such expertise and offer a promising approach for supplying it to LLMs. In this context, KGs are not used to support multi-hop reasoning, but rather to ground a generation task by providing expert knowledge. However, it remains unclear to what extent LLMs benefit from the incorporation of structured expert knowledge represented as KGs in the energetic renovation domain. This short paper investigates the impact of input representation on the quality of LLM-generated renovation paths. Domain knowledge is captured using the Renovation and ENergy Optimization (RENO) ontology and provided to the model as five structured KG serializations (JSON-LD, Turtle, N-Triples, RDF/XML, and linearized) and one unstructured textual representation. Performance is evaluated using an LLM-as-a-judge approach, assessing completeness and ordering of the paths. The results indicate that structured KG representations can improve the completeness of generated renovation paths compared to the textual baseline, suggesting that structured knowledge can support the grounding of LLM outputs. However, the effectiveness of serialization formats varies across models, and no consistent best- performing format is identified. This short paper provides initial insights into how structured knowledge representations influence LLM performance on domain-specific generative tasks and highlights the need for further research into the interactions between ontology design, input representation, and prompt engineering.

Graph retrieval-augmented generation places retrieved subgraphs into the model’s context window at query time, paying a recurring token cost and exposing source data on every call. We study an alternative: compiling a knowledge graph offline into a bank of LoRA adapters, one per entity, that serve as a parametric knowledge layer queried by injecting weights rather than text, at zero query-time context cost. On the MetaQA dataset, we find that subgraph-trained adapters encode context-free factual knowledge that generalizes to unseen questions: on single-valued relations the adapter gains +0.243 exact-match score over a base model that is nearly blind closed- book (0.007), and only the correct adapter recovers this knowledge (an oracle gap of +0.283 over the base model). However, the stored knowledge is not recoverable by similarity: given a query with no subgraph, embedding- based and weight-space geometry retrieval both perform at chance, because a semantically neighbouring entity’s adapter does not contain the answer — knowledge is stored locally and does not transfer. Weight geometry correlates with subgraph semantics (𝜌= +0.329) but not with functional retrievability. We quantify the byte and context-token costs against graph retrieval-augmented generation and discuss deployment implications. Our results establish that parametric knowledge graph memory is feasible for storing knowledge, and identify selecting and composing the right adapters by a mechanism other than semantic similarity as the central open problem— motivating a learned, query-conditioned composition mechanism.

Knowledge graphs for the built environment fail less for want of awareness than for want of construction pipelines that are at once structurally correct and able to scale. We argue for a division of labour: language-model agents read source records and propose classifications, ontology checks decide whether a proposal is admissible, and templates are the only mechanism that writes triples, validated before anything is committed. We report early evidence from one room-automation system in one building: 785 substantive records across 32 rooms. Records are classified from their own row evidence, and the classifier can propose only admissible Brick point classes by construction: a guarantee of admissibility, not of correctness. The templates expand a graph of 4,998 triples that conforms to the project’s shape library, and controlled faults, which also guided that library, show that structural conformance does not establish semantic truth. The prototype is incomplete: proposed classes do not yet parameterise the templates, and equipment identity still reads the standardised component identifier. The two reference artefacts, the classification table and the room fixture, are same-team work, not independent ground truth, and other buildings and input formats are unevaluated.