Manufacturing companies face a critical knowledge management crisis where decades of operational expertise remains trapped in expert minds, leading to extended onboarding periods, ineficient decision-making and knowledge loss through retirement. This paper presents the development of a semantic AI assistant system that combines Knowledge Graphs with Large Language Models to transform how manufacturing engineers access and utilize institutional knowledge. The approach addresses the integration of scattered data across incompatible systems through a three-stage pipeline: multi-format document ingestion, knowledge graph construction, and LLM-enhanced natural language querying. Our implementation demonstrates how semantic technologies can address real-world industrial challenges, reducing information retrieval time and creating pathways for knowledge preservation at enterprise scale.
Modern manufacturing organizations face an unprecedented knowledge management crisis. As
industrial processes become increasingly complex and specialized, critical operational knowledge becomes
concentrated in the minds of expert engineers and technicians[
The scope of this challenge is exemplified by our case study organization: a manufacturing company where new engineering hires require over two years of training to become productive contributors often due to the fragmented and inaccessible nature of institutional knowledge. Critical information resides across incompatible systems including Customer Relationship Management (CRM) platforms, legacy Lotus Notes databases, scattered document repositories, and informal email communications.
This paper presents a novel approach that bridges this semantic gap through the integration of
Knowledge Graphs (KGs) and Large Language Models (LLMs)[
Our system is comprised of a three-stage pipeline that transforms heterogeneous documentation into a unified, accessible knowledge resource. This approach builds upon established manufacturing knowledge graph methodologies while addressing the specific challenges of multi-format enterprise document processing.
The pipeline processes diverse content formats prevalent in manufacturing, including PDFs, HTML, PowerPoint, and structured data exports. The export formats are raw chunked text and accompanying image formats that are linked to the corresponding text: • PDFs: Extracted while maintaining formatting context. • HTML: Analyzed for structural hierarchies.
• PowerPoint: Extracts both text and embedded diagrams critical to specifications.
In the first iteration of the pipeline, a schema is inferred from example chunks of the text. The
ingested documents are then extracted based on the schema inferred and uploaded into the KG. This
approach leverages recent advances in LLM-powered knowledge graph construction[
The schema creation resulted in 11 entity types with 38 unique relationship types. The inferred schema exhibits a three-tier hierarchy: core entities (Equipment, Document), operational entities (Specification, Material, Variant, Position), and parameter entities (SpeedRange, LoadRequirement, VentilationRequirement, ControlMechanism, OperatingCondition). Manufacturing engineers validated both the extracted schema and entity relationships, confirming that the resulting ontology is a valid approach to structuring the available specification documents. However, additional analysis remains necessary to expand the schema and examine the ontological hierarchies for this specific domain.
The knowledge graph construction phase extracted 982 entities from the manufacturing documentation corpus. The extraction process revealed varying completion rates across diferent entity attributes within manufacturing knowledge domains:
We provide a frontend for the user with a chat interface that communicates with a Claude-3.5 instance in the backend. The LLM instance can write Cypher queries to query the Knowledge Graph for specific information. The answer then verbalizes this information and refers to the section of the document that is attached to this entry in the KG. Refer to Appendix A for an architectural overview.
The integration of Large Language Models transforms knowledge graph access from a technical query
interface into a natural conversation platform that mimics expert consultation[
Natural language query translation represents a critical component of the LLM integration[
Manufacturing operations are inherently fragmented, highly interdependent, and often involve
latent knowledge, making traditional approaches inadequate. By extracting and structuring this complex
information before embedding it into a Knowledge Management system, we achieved a far more
intelligent and context-aware retrieval process. This approach aligns with recent advances in graph neural
retrieval frameworks that demonstrate state-of-the-art performance on complex reasoning tasks[
We piloted the software with Sales Engineers in their specialization. The underlying document base consisted of Application Instructions and Technical Specifications. Initial feedback was enthusiastic and scaling the system should result in higher-quality decision-making, less turnaround time and higher employee satisfaction.
Technical challenges center primarily on data quality and schema evolution requirements. The provided documentation often repeats but with diferent terminology and contains often outdated information. Schema evolution proves particularly challenging as manufacturing processes and equipment evolve, requiring knowledge graph structures that can accommodate change without breaking existing functionality.
Document format diversity presents ongoing challenges for content ingestion. Manufacturing
organizations typically maintain information in dozens of diferent formats, many of which require
specialized processing approaches. This challenge suggests the value of standardized documentation
approaches within manufacturing organizations, as advocated by recent Industry 4.0 ontology
standardization eforts[
Organizational challenges focus on user adoption and knowledge contribution patterns. Engineers
demonstrate enthusiasm for knowledge access capabilities but show resistance to knowledge
contribution requirements. This finding echoes challenges reported in other industrial knowledge graph
deployments[
Best practices emerged from the deployment experience that can guide similar initiatives.
Iterative schema development allows for continuous refinement based on user feedback and system
performance. This approach aligns with established methodologies for industrial ontology engineering
platforms[
The author(s) have not employed any Generative AI tools.