A technique that strengthens large language models’ outputs by integrating relevant information external to the model itself
Generative artificial intelligence has transformed how organizations process information and make decisions. Yet anyone working in industry knows that Large Language Models are trained on vast corpora of general knowledge and are therefore essentially blind to the vertical realities that characterize individual manufacturing companies. Having never read the manual of a specific plant, not knowing the bill of materials for a given product, and being unaware of which supplier is used for a certain component, a standard LLM faced with an operational question will either not answer or will generate a response that sounds plausible but is not reliable. In industry, where a wrong decision about a plant can translate into production downtime costing tens of thousands of euros, this limitation has very concrete—and often heavy—consequences.
The answer to this limitation is Retrieval-Augmented Generation (RAG), a technique within AI that enhances large language models’ outputs by integrating, at inference time, relevant information retrieved from knowledge sources external to the model itself.
Beyond document scanning
Adoption of RAG in manufacturing is growing, with more and more companies testing its value—especially in sectors like automotive, precision mechanics, energy, and chemicals/pharmaceuticals. However, its potential is still constrained not by a technological ceiling but by a data bottleneck. For this AI solution to work, the company’s knowledge base must be accessible, structured, and trustworthy. That is not easy to achieve in manufacturing, where information is dispersed across heterogeneous systems, only partially updated and formalized, and, in some cases, effectively “archived” exclusively in the heads of experienced personnel.
IDC research indicates that up to 90% of corporate information assets are unstructured data, such as PDF technical manuals, emails, maintenance logs, and notes. In manufacturing, this vast pool of “Dark Data” is a vital resource that RAG can unlock—provided a prior effort is made to reorganize the information.
A RAG program in the industrial domain should therefore start by building a robust knowledge foundation suited to this context, in three key phases:
1. Collection and digitization
This phase goes far beyond scanning documents: it identifies the information that truly has operational value, the context in which it is used, and the level of freshness required. The work combines Intelligent Document Processing techniques to automatically extract data from PDFs, images, emails, and scans with structured interviews of process owners and frontline staff to capture and formalize tacit knowledge. In parallel, legacy systems are integrated to make their information accessible within the RAG system.
2. Semantic structuring
Content is organized not by keywords, but by meaning and usage context. This involves semantic indexing of documents and data so that the RAG system can understand natural-language queries and return pertinent answers even amid heterogeneous terminology—turning these systems into true intelligent operational assistants.
3. Continuous governance
A RAG system without ongoing updates degrades quickly. If the knowledge base is not kept aligned as things change, the system will return obsolete information that drives operational errors and regulatory noncompliance. It is therefore essential to establish a structured workflow for review, validation, and change notification for each critical process.
High-impact use cases
Once the knowledge base is in place, RAG’s application potential in industry spans multiple operational areas, each with significant impact on efficiency and decision quality. The most mature use case is maintenance: a technician working on a plant typically faces fragmented documentation scattered across several systems; a RAG system, by contrast, lets them query the entire information base in natural language and receive contextualized answers with source traces.
On the procurement and quality control fronts, RAG compresses tasks that today demand hours of repetitive, manual effort and carry a high risk of error. The same holds for impact assessments of changes to a component, material, or manufacturing process: analyses that normally tie up R&D and production directors for days of document searching can be reduced to minutes thanks to RAG queries. Sales teams also benefit, as they can answer technical questions accurately during negotiations and markedly shorten quote preparation times.
RAG is also highly applicable in other contexts: from regulatory compliance to customer complaint management, and even accelerating new-hire onboarding. Across all these areas, the technology reduces lead times—starting with production downtime—improves productivity and performance quality as operating conditions change, limits errors, and, not least, promotes the formalization and systematic dissemination of company knowledge.
The economic impact of these optimizations is substantial. A McKinsey & Company report on generative AI finds that applying these technologies in supply chain and manufacturing could generate additional global value of between $200 and $310 billion per year. These gains stem largely from eliminating information search time and dramatically reducing operational bottlenecks.
Enabling conditions
When launching a RAG initiative in industry, it is critical to remember that success almost always hinges on organizational and cultural factors, not technology: lack of executive commitment, resistance to change, and difficulty in gathering information from in-house experts are the primary—and most recurrent—brakes. Without a clear mandate from leadership and organization-wide involvement from the project definition stage, a RAG system is likely to stall during the knowledge-capture phase.
As noted, this phase is non-negotiable. However complex and labor-intensive, building the corporate knowledge base determines the system’s real value—a value that cannot and should not be delegated to technology adoption alone, as is often assumed. It is equally important to proceed in stages, starting with a pilot that can demonstrate RAG’s value quickly. The optimal choice is to target a bounded process whose improvements are easy to measure—for example, reductions in maintenance time and machine downtime. This approach builds the internal trust needed to progressively scale RAG across the organization.
The ongoing evolution
Current RAG technology is evolving to overcome some of its inherent limitations. One key step is moving from purely statistical retrieval of text fragments to genuine logical understanding of the knowledge corpus, since data fragmentation and lack of correlation often prevent the model from reasoning or honoring dynamic or safety constraints. This is a particularly acute issue in manufacturing, where bills of materials, certifications, and failure histories, for instance, must be queried coherently and in an integrated way. The direction of travel is toward “intelligent RAG” that does more than find similar words: it behaves like an expert assistant by linking information across sources, checking consistency, and enforcing access permissions before answering.
At a statistical level, the so-called “retrieval noise” problem also persists, whereby the system identifies mathematical similarities between terms that lack genuine logical correlation. In manufacturing, this risk is amplified by the prevalence of similar acronyms, codes, and nomenclatures. In other words, beyond a certain threshold, the model struggles to distinguish signal from noise, raising error probability. For this reason, more advanced systems are adopting a different tack: not to search more, but to search better, using selection and validation mechanisms that filter relevant information before it ever reaches the model.
