Textile Factories Get a Truth-Telling AI: Deterministic RAG Turns Risk Scores into Trustworthy Factory Decisions

Artificial intelligence has quietly woven itself into the fabric of modern manufacturing, but in one of the world’s oldest industries it has long had a credibility problem. Textile mills generate torrents of data from looms, dyeing lines, and quality-control stations, yet most machine learning systems that promise to predict defects remain opaque black boxes that shop-floor engineers cannot interrogate and managers cannot safely act upon. A new study published in Knowledge and Information Systems by researchers at Bursa Technical University in Turkey tackles that trust deficit head-on, presenting a framework that does something deceptively simple and technically demanding: it predicts quality risk, explains the prediction, proposes what to do about it, and then answers questions about all of the above without ever drifting into fabrication.

The team, led by Ahmet Metin with Irem Erdas, Kenan Yilmaz, and supervisor Turgay Tugay Bilgin, worked with real production data from an industrial textile partner under a confidentiality agreement. Their starting point is a transparent metric called the CompositeRiskScore, a number from 0 to 100 that serves as a retrospective proxy for quality risk. Rather than inventing an abstract index, the researchers built it from four components that practitioners on a mill floor already understand: recorded quality outcomes, meter difference, machine downtime, and production duration. The score is deliberately defined as a retrospective measure, which means it summarizes what went wrong in a completed production run. The prediction task is then framed with unusual discipline: only information available before quality control, such as production planning parameters and process descriptors, may be used as input variables.

That restriction might sound like an accounting detail, but it addresses one of the most common and corrosive failure modes in industrial machine learning, known as data leakage. If a model is accidentally trained on features that encode the very outcome it is supposed to predict, it will appear almost clairvoyant in the lab and nearly useless in production. The Bursa team engineered their pipeline to be leakage-aware at every stage, and they went further in how they evaluated it. Instead of splitting data rows randomly into training, validation, and test sets, they used a roll-group-disjoint split, ensuring that analytical rows derived from the same physical production roll never straddled different partitions. This prevents the model from effectively memorizing a roll it has already seen in disguise, and it yields performance estimates that are more moderate but far more honest.

Under this stricter evaluation, the model still achieves meaningful risk-ranking performance, correctly ordering production runs by their likelihood of quality problems even when the absolute numbers are less spectacular than leaked models typically report. The researchers also stress-tested generalization with two deliberately harder scenarios: predicting for fabric patterns the model had never encountered, and predicting for looms it had never seen. Both remained challenging, a candid acknowledgment that pattern-specific and machine-specific idiosyncrasies are real barriers in textile analytics. This kind of transparency about where a model struggles is rare in industrial AI literature, where headline accuracy figures often obscure the gap between benchmark performance and deployment reality.

Prediction alone, however, tells a mill operator that something is risky; it does not say what to change. The framework’s second pillar converts model attributions into recommendation candidates, and here the authors make a conceptual move that deserves attention. Explanations from techniques such as SHAP, which quantifies each feature’s contribution to a prediction, are treated not as prescriptions but as raw material. The system organizes them into groups with explicit actionability levels: features that can be adjusted in real time during production, features that require a setup change, features that belong to planning or procurement decisions, and contextual features that cannot be changed at all. A loom speed might be tweakable within minutes, while a yarn specification might require renegotiating with a supplier, and the fabric pattern itself is simply a given.

Once candidate changes are identified, they are subjected to model-consistent what-if checks. The system re-runs the trained model with the proposed modification and measures whether the predicted CompositeRiskScore actually falls. This is a crucial guardrail against a subtle error that plagues counterfactual explanation methods: an attribution may suggest a feature matters, but the model’s actual response to changing that feature may be nonlinear, saturated, or overridden by interactions with other variables. By measuring the model’s internal response directly, the framework filters out recommendations that sound plausible but would not move the predicted risk. The authors are equally careful about what these checks do not prove. A reduction in model-predicted risk is not evidence that applying the change in the physical world will improve realized quality; it is a statement about the model, not about the loom.

The third pillar is the retrieval layer, and it is where the study engages most directly with the current enthusiasm for generative AI. Retrieval-augmented generation, or RAG, has become the dominant pattern for grounding large language models in factual documents, but in industrial settings an open-ended generative system can hallucinate process parameters, invent quality figures, or leak post-production information into answers that operators treat as authoritative. The Bursa team’s answer is what they call deterministic RAG, a deliberately constrained design. The retrieval interface operates over structured JSON artifacts rather than free text, uses a single primary context instead of blending multiple retrieved passages, and preserves locked prediction and explanation fields that the answering layer cannot alter.

The constraints extend to what the system is allowed to know. Post-QC results and ground-truth quality information are excluded from deployment-facing answers entirely, so the interface cannot accidentally reveal the outcome it is supposed to be helping prevent. And when a user asks about something absent from the selected context, the system returns an explicit fallback, stating that the requested information is not available in context, rather than improvising a plausible-sounding fabrication. The result is less a conversational oracle than a structured decision support interface: traceable, reproducible, and auditable, with every answer anchored to a specific retrieved artifact and a locked model output.

The significance of this work extends well beyond textiles. Factories across sectors are being sold generative AI copilots, and the central question regulators, auditors, and engineers ask is the same one this framework answers: can you trace every recommendation back to a verifiable model output and a verifiable data artifact? By chaining leakage-controlled prediction, actionability-tiered explanation, model-consistent what-if validation, and locked-context retrieval into a single workflow, the researchers demonstrate that the answer can be yes without sacrificing the usefulness of either machine learning or retrieval-based interfaces. The design choices, from the roll-group-disjoint evaluation to the not-available-in-context fallback, each close a specific pathway to silent failure.

Limitations remain, and the authors do not hide them. The dataset is proprietary and cannot be released publicly, available only upon reasonable request with the industrial partner’s approval, which means independent replication will require other factories to adopt the protocol on their own data. The what-if checks quantify model response, not physical causation, so the framework still depends on human judgment to translate a predicted risk reduction into an operational intervention. Unseen patterns and unseen looms continue to challenge generalization, suggesting that transfer across product lines is an open research problem. Yet the study’s core contribution stands: a working demonstration that industrial AI can be simultaneously predictive, prescriptive, and honest about the boundaries of its own knowledge. For an industry that literally measures quality thread by thread, that combination of precision and candor may prove to be the most valuable fabric of all.

Subject of Research: Explainable machine learning and deterministic retrieval-augmented decision support for quality risk prediction in industrial textile manufacturing

Article Title: Explainable risk prediction and prescriptive decision support with deterministic RAG in industrial textile manufacturing

Article References: Explainable risk prediction and prescriptive decision support with deterministic RAG in industrial textile manufacturing. (n.d.). https://doi.org/10.1007/s10115-026-02911-1

Image Credits: AI Generated

DOI: 10.1007/s10115-026-02911-1

Keywords: explainable artificial intelligence, textile manufacturing, CompositeRiskScore, retrieval-augmented generation, decision support systems, prescriptive analytics, SHAP, counterfactual explanations, data leakage, quality prediction, machine learning, industrial automation