Table of Contents
This glossary defines 40 terms that come up in industrial AI projects, from large language models and RAG to soft sensors and digital twins. Each entry gives the English, French and Arabic term and a short definition, so that engineers, managers and vendors writing in different languages mean the same thing.
Key takeaways
- The terms are grouped into six themes: core concepts, language models, document search, assistants and agents, deployment and security, and plant data and process models.
- Where a standard or official body defines a term (OECD, NIST, CISA, OWASP, the Open Source Initiative), the entry follows that definition and links to it.
- Acronyms such as LLM, RAG, OCR, GPU and DoE stay in Latin script in all three languages, and French practice keeps many English terms, such as token and fine-tuning.
- A table near the end separates terms that are often confused, such as open-weight and open source.
Core concepts
| English | French | Arabic |
|---|---|---|
| AI system | système d'IA | نظام ذكاء اصطناعي |
| Machine learning (ML) | apprentissage automatique | التعلم الآلي |
| Generative AI | IA générative | الذكاء الاصطناعي التوليدي |
| Large language model (LLM) | grand modèle de langage (LLM) | نموذج لغوي كبير |
| Small language model (SLM) | petit modèle de langage (SLM) | نموذج لغوي صغير |
- AI system. The OECD defines it as a machine-based system that infers, from the input it receives, how to generate outputs such as predictions, content, recommendations or decisions that can influence physical or virtual environments (OECD). Systems differ in how autonomous and adaptive they are after deployment.
- Machine learning (ML). Algorithms that learn from data to make predictions or decisions without being explicitly programmed for each case (CISA and partners). Most industrial models, from soft sensors to anomaly detectors, are ML models.
- Generative AI. Models that produce new content such as text, images or code. NIST's Generative AI Profile uses the definition "the class of AI models that emulate the structure and characteristics of input data in order to generate derived synthetic content" (NIST AI 600-1).
- Large language model (LLM). A machine learning model designed to understand a natural-language prompt and generate a response that humans can understand (CISA and partners). DALILE is an assistant of this kind that runs on your own servers.
- Small language model (SLM). A language model small enough to run on modest hardware, such as one GPU. There is no official size threshold; see small language models in industry.
Working with language models
| English | French | Arabic |
|---|---|---|
| Token | token (jeton) | رمز (token) |
| Context window | fenêtre de contexte | نافذة السياق |
| Inference | inférence | الاستدلال (inference) |
| Hallucination | hallucination | الهلوسة |
| Fine-tuning | fine-tuning (affinage) | الضبط الدقيق |
| Quantization | quantification | التكميم |
- Token. The unit of text a model reads and writes, often a word or part of a word. The same text can give very different token counts in Arabic, French and English, depending on the tokenizer (choosing an open-weight LLM).
- Context window. The maximum number of tokens a model can take into account at once: instructions, retrieved passages and its own answer included.
- Inference. Running a trained model to produce an output, as opposed to training it. On-premise, it is the daily workload that GPU servers are sized for (GPU sizing guide).
- Hallucination. Confidently stated but false content produced by a generative model. NIST calls it "confabulation" and describes it as a natural result of how these models generate text, for example by predicting the next token (NIST AI 600-1).
- Fine-tuning. Continuing the training of an existing model on your own examples, so its weights change. A common low-cost method, LoRA, freezes the original weights and trains small added matrices (Hu et al., 2021). It changes behaviour better than it adds knowledge; see fine-tuning vs RAG.
- Quantization. Storing a model's weights with fewer bits, for example 8 or 4 instead of 16, so it needs less GPU memory. It can reduce accuracy, so test the quantized model on your own tasks.
Searching documents and data
| English | French | Arabic |
|---|---|---|
| Retrieval-augmented generation (RAG) | génération augmentée par récupération (RAG) | التوليد المعزز بالاسترجاع (RAG) |
| Embedding | embedding (plongement vectoriel) | التضمين (embedding) |
| Chunking | découpage (chunking) | التقطيع (chunking) |
| Hybrid search | recherche hybride | البحث الهجين |
| Reranker | modèle de reclassement (reranker) | نموذج إعادة الترتيب (reranker) |
| Optical character recognition (OCR) | reconnaissance optique de caractères (OCR) | التعرف الضوئي على الحروف (OCR) |
| Text-to-SQL | text-to-SQL | تحويل النص إلى استعلام SQL |
- Retrieval-augmented generation (RAG). The system first retrieves relevant passages from your documents, then has the language model answer from them, with citations. The term comes from a 2020 paper by Lewis et al.; see our guide to RAG over maintenance manuals.
- Embedding. A vector of numbers that represents the meaning of a passage, so that passages with similar meaning sit close together. Embeddings are stored in a vector database and searched by similarity.
- Chunking. Splitting documents into passages before indexing them, ideally along their structure so a table or a warning is not cut from its context.
- Hybrid search. Combining vector search with keyword search, such as BM25, which finds exact part numbers and alarm codes that vector search tends to miss.
- Reranker. A second model that reorders the retrieved passages by relevance, so only the best few reach the language model.
- OCR. Converting scanned pages or images into text. Arabic script, numerals and tables remain hard, so check samples by hand (OCR for scanned technical documents).
- Text-to-SQL. Turning a plain-language question into a database query. DALILE handles the SQL so users can query databases in plain language; see text-to-SQL on production databases.
Assistants and agents
| English | French | Arabic |
|---|---|---|
| Chatbot | chatbot (agent conversationnel) | روبوت المحادثة (chatbot) |
| Copilot | copilote | مساعد مدمج (copilot) |
| AI agent | agent IA | وكيل ذكاء اصطناعي |
| Human in the loop | humain dans la boucle | الإنسان في الحلقة |
| Evaluation set | jeu d'évaluation | مجموعة التقييم |
- Chatbot. Software that answers in a conversation and changes nothing outside it.
- Copilot. An assistant built into the tool where the work is done. It drafts and suggests; the person decides and executes each action.
- AI agent. Software that can process data, make decisions and initiate autonomous actions using AI and ML models (CISA and partners). What sets it apart from a chatbot is permission to act on other systems; see chatbot vs copilot vs AI agent.
- Human in the loop. A person reviews or approves an AI output before it takes effect. Joint guidance from CISA and partner agencies advises limiting active control of OT by AI without a human in the loop (CISA and partners). More in human-in-the-loop AI agents.
- Evaluation set. A fixed set of real questions or cases with known correct answers, used to measure a system before go-live and after every change (evaluating an LLM on your documents).
Deployment, security and sovereignty
| English | French | Arabic |
|---|---|---|
| On-premise | on-premise (sur site) | داخل المنشأة (on-premise) |
| Air gap | air gap (isolement physique) | العزل عن الشبكة (air gap) |
| Sovereign AI | IA souveraine | الذكاء الاصطناعي السيادي |
| Open-weight model | modèle open-weight | نموذج مفتوح الأوزان |
| Prompt injection | injection de prompt | حقن التعليمات (prompt injection) |
| ANPDP | Autorité nationale de protection des données à caractère personnel (ANPDP) | السلطة الوطنية لحماية المعطيات ذات الطابع الشخصي |
- On-premise. Software and models run on servers the company controls, in its own facilities, rather than on a provider's cloud. DALILE runs this way: your data is processed on your own servers.
- Air gap. NIST's glossary defines it as an interface between two systems that are not physically connected, where any logical connection is not automated and data moves only manually, under human control (NIST CSRC). See air-gapped AI deployment.
- Sovereign AI. There is no standard definition. In industry it usually means that the company keeps control of its data, models, infrastructure and operations (sovereign AI guide; sovereign cloud vs private cloud vs on-premise).
- Open-weight model. A model whose trained weights you can download and run on your own hardware, under the publisher's licence. That is not the same as open source: the Open Source AI Definition also requires the training and inference code and enough information about the training data to build a substantially equivalent system.
- Prompt injection. An attack in which instructions hidden in a user's input, or in a file or web page the model reads, change its behaviour in unintended ways. OWASP ranks it first (LLM01) in its 2025 Top 10 for LLM applications (OWASP); see securing an on-premise LLM.
- ANPDP. Algeria's personal data protection authority, which oversees Law 18-07 of 10 June 2018, amended by Law 25-11 of 24 July 2025. According to the CMS guide, transferring personal data abroad requires its authorisation. Confirm with counsel how this applies to your project.
Plant data and process models
| English | French | Arabic |
|---|---|---|
| Operational technology (OT) | technologies opérationnelles (OT) | التكنولوجيا التشغيلية (OT) |
| Data historian | historian (base de données historique) | قاعدة البيانات التاريخية (historian) |
| Soft sensor | capteur logiciel | المستشعر البرمجي |
| Surrogate model | modèle de substitution | النموذج البديل |
| Digital twin | jumeau numérique | التوأم الرقمي |
| Golden batch | golden batch (lot de référence) | الدفعة الذهبية |
| Design of experiments (DoE) | plan d'expériences | تصميم التجارب (DoE) |
| Bayesian optimisation | optimisation bayésienne | التحسين البايزي |
| Anomaly detection | détection d'anomalies | كشف الشذوذ |
| Predictive maintenance | maintenance prédictive | الصيانة التنبؤية |
| Model drift | dérive du modèle | انجراف النموذج |
- Operational technology (OT). NIST describes it as the hardware, software and firmware used to detect or cause changes in physical processes through direct monitoring or control of devices (NIST CSRC). PLCs, DCS and SCADA are OT; see integrating AI with SCADA and OT.
- Data historian. "A centralized database supporting data analysis using statistical process control techniques" (NIST CSRC). It stores time-stamped process values and is the main data source for most plant models.
- Soft sensor. A model that estimates a variable you cannot measure online, such as product purity, from variables you already measure (soft sensors guide).
- Surrogate model. A fast, data-fitted approximation of a slow simulation, used for optimisation inside the range it was trained on (surrogate models guide).
- Digital twin. NIST describes digital twin technology as creating electronic representations of real-world entities and making it possible to view their states and the transitions between them (NIST IR 8356). See digital twins in manufacturing.
- Golden batch. One past batch, usually the best, used as the reference for new ones. A profile built from many good batches is stronger because it shows normal variation (golden batch analysis).
- Design of experiments (DoE). "An efficient procedure for planning experiments", with the plan laid out before any test is run, according to the NIST/SEMATECH e-Handbook.
- Bayesian optimisation. A sequential method that fits a model to the results so far and uses it to choose the next experiment, suited to expensive tests (DoE vs Bayesian optimisation).
- Anomaly detection. Flagging data that departs from normal behaviour, such as an unusual vibration pattern or a combination of process values never seen in normal operation (anomaly detection methods).
- Predictive maintenance. Planning maintenance from the measured condition of equipment and a forecast of its degradation, rather than at fixed intervals or after failure (predictive maintenance in oil and gas).
- Model drift. The loss of accuracy of a model in service because the process, raw materials or instruments have changed since its training data was collected. Compare predictions with lab results or actual outcomes over time to catch it.
Which terms are often confused?
| Often confused | The difference |
|---|---|
| Open-weight vs open source | Open-weight: downloadable weights under a licence. Open source (OSI): also the code and training-data information, with freedom to use, study, modify and share. |
| On-premise vs air-gapped | On-premise is about where the servers are; air-gapped is about having no network connection. An on-premise system is not necessarily air-gapped. |
| RAG vs fine-tuning | RAG changes what the model reads; fine-tuning changes its weights. |
| Chatbot vs copilot vs AI agent | The chatbot answers, the copilot drafts for a person, the agent acts on other systems. |
| Soft sensor vs surrogate vs digital twin | A soft sensor learns from plant and lab data, a surrogate from simulation runs; a digital twin represents an asset and follows its state. |
| Hallucination vs retrieval error | A retrieval error: the right passage was never found. A hallucination: the model states what the passages do not support. |
What we have seen at KABAS AI
Our two published projects, at ADFC (Sonatrach Group) and ENAD Shymeca, use Virtual Lab to test formulations virtually before physical trials, combining simulation, what-if scenarios and multi-criteria optimisation; the lab confirms what the model proposes. At ENAD Shymeca, the project delivered a 40% reduction in physical testing cycles.
A shared vocabulary costs little and prevents disputes later. Before a project starts:
- Put a short glossary in the specification. Define the terms your requirements depend on, such as on-premise, agent and accuracy, in every language the documents use (writing an AI project RFP).
- Ask vendors to answer in your terms: is the model open-weight, does the system need an internet connection, can it act on other systems?
- Write acceptance criteria against an evaluation set, not adjectives such as "accurate".
- Use the same terms in training material for operators and engineers.
Our consulting team offers readiness assessments and use-case identification, and can help fix these terms for your project.
Sources
- OECD.AI, Updates to the OECD's definition of an AI system explained (November 2023)
- CISA, ASD's ACSC and partners, Principles for the Secure Integration of Artificial Intelligence in Operational Technology
- NIST AI 600-1, Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile (July 2024)
- Hu et al., LoRA: Low-Rank Adaptation of Large Language Models (arXiv, 2021)
- Lewis et al., Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks (arXiv, 2020)
- NIST CSRC Glossary, Air gap
- Open Source Initiative, The Open Source AI Definition 1.0
- OWASP, LLM01:2025 Prompt Injection
- CMS Expert Guide to data protection and cyber security laws: Algeria
- NIST CSRC Glossary, Operational technology
- NIST CSRC Glossary, Data historian
- NIST IR 8356, Security and Trust Considerations for Digital Twin Technology (February 2025)
- NIST/SEMATECH e-Handbook of Statistical Methods, What is experimental design?
Frequently asked questions
Why use a glossary in three languages for an AI project?
In Algeria and the wider Maghreb, one project often mixes French specifications, English vendor documentation and Arabic conversations with operators. If each language uses a different word, or the same word for different things, requirements and acceptance criteria drift apart. Agreeing on the terms in all three languages at the start avoids that.
Is an open-weight model the same as an open-source model?
No. An open-weight model lets you download and run its trained weights under the publisher's licence. The Open Source Initiative's definition of open-source AI also requires the code used to train and run the model and enough information about the training data to build a substantially equivalent system, plus the freedom to use, study, modify and share it.
What is the difference between on-premise and air-gapped AI?
On-premise means the AI runs on servers you control in your own facility; those servers may still connect to other networks, for example for updates. Air-gapped means the system has no physical connection to other networks, and data moves in or out only manually, under human control.
Is a hallucination the same as a retrieval error?
No. In a RAG system, a retrieval error means the right passage was never found. A hallucination means the model states something that the retrieved passages do not support, or that is false. The two need different fixes, so measure them separately.
Should acronyms such as RAG, LLM and GPU be translated into French or Arabic?
In practice, industry professionals keep them in Latin script in all three languages. Give the full term once, in the reader's language, the first time the acronym appears, then use the acronym.



