From RAG to Industrial World Models: Why “LLM + RAG” Is Not Enough to Comprehend a Factory

Over the past two years, the rapid advancement of large language models has made industrial AI one of the most closely watched directions in the digital transformation of manufacturing. From equipment maintenance Q&A and process knowledge retrieval, to production decision-making support and intelligent O&M assistants, a growing number of enterprises are beginning to explore the integration of LLM capabilities into industrial scenarios. In actual deployment, the integration of “Large Language Models + RAG (Retrieval-Augmented Generation)” has become the de facto standard approach across the industry.

The technical approach is not particularly complicated:

Enterprises vectorize documents such as equipment manuals, process specifications, operation and maintenance experience, and quality standards, enabling semantic retrieval through vector databases and then generating final answers in conjunction with large language models. This model has indeed improved the efficiency of industrial knowledge retrieval. But as more companies look to AI for deeper involvement in fault diagnosis, process optimization, production scheduling, and even intelligent decision-making, a question is becoming increasingly evident:

AI may be able to read documents, but that doesn’t mean it can read a factory.

This marks a new phase in the evolution of industrial AI.

The industry’s focus is shifting from “knowledge retrieval” to “industrial world modeling.”

The first threshold for industrial AI: Industrial knowledge is not the same as industrial documents.

The reason RAG has succeeded in the internet domain comes down to one fundamental fact: most internet knowledge exists in textual form. Whether it is encyclopedic knowledge, technical materials, or industry reports, the core of all such content is descriptive information.

But industrial knowledge is different. In a real factory, what drives production and operations is not the documents themselves, but the equipment relationships, process logic, control rules, and operational constraints that lie hidden behind those documents.

Take a typical production scenario:

• Reactor temperature rises;
• Cooling system load increases as a result;
• Heat exchange efficiency is subsequently affected;
• Ultimately, the stability of downstream processes is compromised.

This is a classic industrial causal chain.

It is not a textual description, but rather a set of dynamic interconnections that span across equipment, systems, and process stages.

Likewise, a single production decision is often subject to multiple constraints across various dimensions:

• Process constraints;
• Safety constraints;
• Quality constraints;
• Energy consumption constraints;
• Equipment operating limits.

Together, these constraints define what can and cannot be done.

And this type of knowledge is precisely what traditional document retrieval systems find hardest to capture.

So the first challenge facing industrial AI is not a lack of model capability; it is a fundamental limitation in how knowledge itself is represented.

Why RAG Cannot Solve the Core Problem of Industrial Intelligence

Drawing on extensive project experience, three types of challenges have consistently emerged as common pain points in industrial AI deployments today.

Category 1: Missing Relationship Knowledge

An industrial system is, at its core, a complex network.

• Equipment is interconnected with equipment;
• Processes are interdependent with processes;
• Business workflows are coordinated with one another.

For example: Raw material silo → Batching system → Mixing equipment → Sintering process → Finished product silo

The entire process forms a complete operational chain. Traditional vector retrieval may be able to find relevant documents, but it cannot fully capture these relational networks.

Category 2: Missing Rule Knowledge

Industrial production is not an open-ended generative scenario; a large amount of knowledge consists of hard constraint rules. For example:

• If temperature exceeds the safety threshold, an alarm must be triggered;
• If pressure exceeds the limit, interlock shutdown must be activated;
• Specific process parameters must stay within their operating ranges;
• Certain operations must be performed in a prescribed sequence.

This knowledge is not probabilistic—it is deterministic. It cannot be inferred through semantic similarity.

Category 3: Missing Temporal Knowledge

Industrial processes are inherently temporal in nature.

• Different valve opening sequences;
• Different equipment startup orders;
• Different timing of process execution;
• Can lead to completely different outcomes.

Industrial operations are, at their core, a continuously evolving dynamic system. Traditional RAG, however, focuses largely on static knowledge retrieval and struggles to capture this dynamic evolution.

Ontology + LLM: Industrial AI Is Moving Toward a New Technical Architecture

If large language models excel at understanding language, then industrial ontologies excel at understanding the industrial world.

In recent years, a growing number of international industrial AI research institutions have begun integrating ontologies with LLMs, exploring a development path that combines neural networks with symbolic systems—the Neuro-Symbolic AI approach.

The reason is straightforward: LLMs know how to organize knowledge, while industrial ontologies know how knowledge is connected.

An industrial ontology, at its core, is designed to answer four fundamental questions:

• What exists in the factory;
• How they relate to one another;
• What business rules must be followed;
• What operational constraints apply.

For example:

Equipment includes pumps, valves, motors, and sensors;

Processes include feeding, reacting, cooling, and conveying;

Business functions include planning, production, quality control, and O&M;

Rules include safety rules, process rules, and quality rules.

Once these objects, relationships, and rules are defined in a standardized way, the factory has a unified industrial cognitive framework at its disposal. The LLM can then reason and interact based on this framework.

From a technical perspective:

• LLM handles language understanding;
• Ontology handles industrial knowledge organization;
• Rule engine handles constraint management;
• Causal model handles logical reasoning.

Only when these four factors work in synergy can true industrial intelligence capabilities be realized.

MOM Systems: The Engineered Implementation of Industrial Cognitive Frameworks

Industrial ontology is not just a theoretical framework—its real value lies in how it is implemented in actual production management processes. Drawing on years of industrial digitalization practice, Redcoast has independently developed its MOM production and operation management system.

Unlike traditional MES system, which focuses primarily on production execution, MOM system covers a much broader spectrum, including production planning, on-site scheduling, equipment O&M, quality management, energy management, materials management, and other business areas.

More importantly, the system is built on a unified industrial object model, enabling standardized modeling of equipment, processes, personnel, materials, and business workflows.

This means:

• Enterprises gain not just data;
• But also the relationships between that data.
• Not just business records;
• But also the underlying logic of how operations actually run.

This is also an important prerequisite for subsequent industrial intelligence development.

Digital Twin: Giving the Factory a Real-Time Digital Mirror

If the industrial ontology answers the question: “What is the factory?”

Then the digital twin answers: “What is happening in the factory right now?”

Many companies today equate 3D visualization with digital twins. Indeed, 3D visualization is just one of the ways a digital twin can be presented. The true core value of a digital twin lies in real-time mapping. Through continuous connectivity with on-site equipment, control systems, and production data, the digital space can reflect the factory’s operational status in real time. Key data such as temperature, pressure, flow rate, rotational speed, energy consumption, material status are continuously synchronized and updated. Therefore, the digital twin is not just a display of what’s happening on the factory floor; it serves as a real-time vehicle for the state of the industrial world.

And this is precisely a critical component of the Industrial World Model.

The Industrial World Model: From “Knowing What Happened” to “Knowing Why It Happened”

Currently, World Models are emerging as a significant research direction in the AI field. For the industrial domain, the Industrial World Model is essentially a digital representation of how a factory operates.

It does not just describe the factory—it seeks to understand it.

A complete Industrial World Model typically comprises four core capabilities:

• World Knowledge

Building an industrial knowledge system with ontology.

• World State

Achieving real-time perception of on-site status through digital twins.

• World Laws

Describing operating laws through mechanistic models, process models, and causal models.

• World Simulation

Predicting future state evolution through simulation and algorithms.

Once these capabilities are integrated, AI evolves from simple Q&A to operational simulation and decision analysis.

This enables true understanding of:

• Why it happened;
• What will happen next;
• How to prevent it from happening.

This marks a significant milestone in the evolution of industrial AI—from assistive tools to intelligent systems.

Why Industrial Agents Cannot Simply Copy General-purpose Agents?

As Agent technology continues to gain momentum, more and more enterprises are beginning to explore the development of industrial Agents. But industrial Agents and office Agents are fundamentally different.

In office settings, the cost of trial and error is low—an email can be resent, a report can be revised. But industrial production is a different story. A single wrong decision can lead to:

• Equipment damage;
• Product scrap;
• Production line shutdowns;
• Safety incidents.

Therefore, an industrial Agent cannot be built on unbounded generative logic; it must be subject to strict constraints. To this end, Redcoast adheres to the Neuro-Symbolic AI development path by integrating neural networks with symbolic systems.

• Leveraging LLMs for natural interaction and knowledge understanding;
• Leveraging industrial ontologies to establish a unified industrial cognition;
• Leveraging rule engines to manage operational boundaries;
• Leveraging causal models to enhance decision credibility;
• Leveraging approval mechanisms to ensure critical operations remain safe and controllable.

Under this framework,, the industrial Agent is no longer just a Q&A assistant; rather, it becomes an intelligent execution unit capable of perception, analysis, reasoning, and decision-making. At the current stage, it assists human decision-making in a Copilot mode, while gradually evolving toward autonomous operations in an Autopilot model in the future.

The core of industrial AI competition is shifting from model capability to industrial understanding.

Over the past few years, industrial AI has focused primarily on model scale, parameter count, and knowledge retrieval capabilities. But as industrial practice deepens, a new consensus is beginning to take shape across the industry: the real challenge for industrial AI is not how to make models read documents—it’s how to make them understand the factory.

• From Ontology to Digital Twin;
• From Causal Reasoning to the Industrial World Model;
• From Knowledge Q&A to Industrial Agents.

Industrial intelligence is moving from “information access” to “operational understanding”, and from “cognitive augmentation” to “decision-making support.”

In the future, the most critical digital asset for manufacturing enterprises will no longer be data itself, but the Industrial World Model that can continuously describe, understand, and simulate the operational dynamics of the factory.

This is also the direction in which Redcoast will continue to deepen its efforts in the future, because the ultimate competition in industrial AI is not about who has the larger model, but rather who understands the industrial world better.