Enterprise AI glossary
37 terms that come up in AI projects: search over documents, models and agents, security, operations, governance and standards, architecture. Each term has a short explanation of what it means in practice.
In short
Terms are grouped by topic. Each one has its own anchor (for example #rag), so a single term can be linked from an article, a document or a chat.
Search and RAG
RAG (Retrieval-Augmented Generation)
An architecture in which a language model, before answering, is given passages from your documents from a search index and answers from them rather than from what it memorised during training. Documents can be updated without retraining the model, the answer cites its source, and access rights are enforced at the search step. For enterprise assistants, this is the default architecture. More: RAG systems.
Agentic RAG
A variant of RAG in which the model itself decides which searches to run and in what order: several queries to the index, calls to external systems, an answer assembled over several steps. It handles tasks where one search is not enough, but it is noticeably more expensive and slower than plain RAG, and harder to verify.
Embedding
A numeric vector a model uses to represent the meaning of a text, or of an image or code. Texts with similar meaning get nearby vectors, and semantic search is built on that. RAG accuracy often depends more on the embedding model than on the model that writes the answer.
Chunking
Splitting documents into small passages before indexing. Passage size, overlap and respect for the document structure (sections, clauses, table headers) affect search quality more than anything else. Passages that are too small lose context; passages that are too large bring the model noise.
Reranker
A model that, after the initial search, reorders the passages found by how well they answer the question. It adds latency but usually improves the accuracy of what reaches the model noticeably. How much it helps is measured on your own questions.
Hybrid Search
Vector search by meaning and full-text search by words run together, and their results are merged. Vector search finds paraphrases; full-text search finds exact names, codes and numbers. For Hebrew and Russian, the full-text part has to handle morphology: prefixes, attached prepositions, case endings.
Vector Database
A store for embeddings with fast nearest-vector search. PostgreSQL with pgvector, Qdrant, Weaviate, Milvus, cloud services such as Pinecone. The choice depends on corpus size, requirements on where data may sit, and what your infrastructure already runs.
Cosine Similarity
A measure of how close two vectors are: the cosine of the angle between them, from −1 to 1. The closer to 1, the closer the meaning. It is the standard measure in semantic search because it does not depend on vector length.
Models and agents
LLM (Large Language Model)
A neural network trained on large volumes of text that produces text by predicting the next token. Common families: GPT, Claude, Gemini, Llama, Mistral, Qwen. Small open models can run on your own servers; the largest are available only through a vendor API.
Token
The unit of text a model works with: part of a word, a word or a symbol. The cost of a call and the size of the context window are counted in tokens. The same meaning in Hebrew or Russian usually takes more tokens than in English, so cost has to be calculated on your own texts.
Context Window
The maximum a model processes in one call: instructions, retrieved passages, conversation history and the answer, in tokens. A large window does not replace search: every call gets more expensive and slower, and the model pays less attention to the middle of a long context. It is an architectural choice, not just a question of capacity.
Fine-tuning
Further training of an existing model on your examples, which changes its weights. It suits a stable format, style or narrow terminology. For answers from current documents, RAG is usually better. Fine-tuning is added after RAG, when accuracy on a narrow task stops improving.
AI Agent
A system in which the model does not only answer but also acts: it calls APIs, queries databases, runs code. That changes the risk profile: every action needs authorisation, logging, a way to undo it, and a person in the loop where a mistake is costly.
Function Calling / Tool Use
The ability of a model to call an external function (search, calculation, database query, API) with structured parameters and receive the result. Agents are built on it. Every tool available to the model widens the attack surface, so the tool set is designed on the principle of least privilege.
MCP (Model Context Protocol)
An open protocol, proposed by Anthropic in 2024, for connecting data and tools to AI assistants through MCP servers. A system is connected once and different assistants can use it, instead of one integration per assistant. More: MCP integration.
Vertical AI
An AI system built for a specific industry: its own data model, industry integrations and the rules a regulator requires. Unlike a general assistant, it knows which questions to ask and which answers must not be given without a check.
Security
Hallucination
A model answer that is fluent and confident but wrong or invented. Hallucinations cannot be eliminated completely. They are limited by grounding in documents through RAG, a required source citation, refusing to answer when data is insufficient, and human review where a mistake is costly.
Prompt Injection
An attack in which text that reaches the model (a user message, a document, an email, a web page) contains instructions, and the model carries them out as commands. Defence: treat retrieved content as data, not commands; limit the model's access to tools; check structured answers separately; and never let a model answer trigger a privileged action without review.
Guardrails
Checks before and after the model call. They block unwanted input (injections, personal data that must not leave) and unwanted answers (policy violations, an answer without a source, a wrong format). They are necessary but not sufficient: they catch the simple cases, and the architecture catches the hard ones.
Zero Data Retention (ZDR)
A model vendor's commitment not to store prompts and answers and not to use them for training. It matters when sensitive data goes through an external API. Check the contract: "ZDR" in marketing copy without a contractual commitment is not protection.
Operations
Observability
Being able to answer what happened in a specific conversation, why, and how often it recurs. It covers logs of requests, retrieved passages and answers, latency and cost metrics, error rates and user feedback. Most failures show up in this data before complaints arrive.
Evaluation (LLM eval)
Measuring whether the system does what it should. In production, that means three layers: regression on a fixed set of questions at every release, sampled human review of real answers, and incident tracking, meaning escalations and refusals. Without all three, you cannot say whether an update made the system better or worse.
Latency
The time from request to answer. Users feel the time to the first word above all, so conversational interfaces stream the answer as it is generated. Every extra step (reranking, a tool call, a second model pass) adds latency, and it is measured together with accuracy.
Model Drift
An unnoticed change in a system's behaviour over time. It comes in two forms: the vendor updated the model and it answers the same questions differently, or the questions and data changed while the model stayed the same. It is caught by running a fixed set of questions regularly and watching the answers.
Rollback / Kill Switch
A mechanism that quickly returns the system to a previous version of the model, prompt or rules, or switches it off entirely. The key word is "tested": a rollback that has never been run is not a rollback, it is hope.
Governance and standards
Human in the Loop (HITL)
A design in which a person checks a high-stakes system decision before it is carried out. The EU AI Act requires human oversight for high-risk systems. In practice, a person should see any action that cannot be cheaply undone, any decision that affects someone's rights or money, and any document going to a regulator.
Audit Trail
A log that cannot be changed after the fact. For each system decision, it records the input, the answer, the model and rules version, the sources retrieved and any human intervention. The EU AI Act requires automatic logging for high-risk systems, and GDPR and Israeli law require that a decision affecting a person can be explained. Without the log, last year's decision cannot be justified.
Data Lineage
The recorded history of a system's data: where each passage came from, when it was loaded, how it was transformed and who has access to it. Without it, you cannot tell an auditor why the model answered as it did, and you cannot delete a person's data on request.
AI Governance Maturity
Checked with five questions. Is there a register: is every AI system in production known? Does each one have a risk level? Are the controls written down and followed? Are drift, hallucinations and incidents tracked? Does each system have a named owner? A "no" to any of them is a gap, not a task for later. More: AI strategy.
NIST AI RMF
The AI Risk Management Framework of the US National Institute of Standards and Technology, published in 2023. A voluntary methodology and a convenient basis for an internal governance system. It is built around four functions: Govern, Map, Measure, Manage.
ISO/IEC 42001
The international standard for an AI management system, published in December 2023. It allows certification with an external audit and is aligned in structure with ISO/IEC 27001. It is used when a client or regulator needs external confirmation. SLAtech does not hold ISO/IEC 42001 certification; the description is for reference.
EU AI Act
The EU regulation on artificial intelligence, whose requirements take effect in stages. It sorts systems into four risk levels: prohibited, high, limited and minimal. High-risk systems require risk and data management, technical documentation, logging, human oversight and post-launch monitoring. It applies extraterritorially: if your users are in the EU, it applies to you too.
Amendment 13 to the Israeli Privacy Protection Law
A change to Israeli law that widened its definitions, including that of especially sensitive information (among other things salary, financial activity and location). The amendment almost abolished database registration for the private sector, introduced a duty to notify the regulator of large databases with sensitive data and, for some organisations, a duty to appoint a privacy protection officer. The regulator gained the power to impose administrative fines. For an AI system this is a question of logs, access rights and which data goes to an external model vendor. More: Amendment 13 for systems.
Architecture and integrations
Multi-tenant
A SaaS architecture in which one system serves many clients from a shared codebase and infrastructure. Isolation comes at three levels: a shared database filtered by client, a separate schema per client, separate infrastructure per client. The choice depends on cost, regulatory requirements and availability commitments.
Tenant Isolation
The guarantee that one SaaS client cannot see, change or affect another client's data. It is enforced by a filter at the data access layer rather than in each query, and checked by automated tests that try to reach another client's data. In an AI system, isolation extends to the search index: another client's passage must not reach an answer.
WhatsApp Business API (WABA)
Meta's official interface through which an organisation sends and receives WhatsApp messages at volume. It requires business verification and pre-approved templates for outgoing messages. It is the basis of any business bot on WhatsApp. We connect to the official API, not to unofficial gateways.
HL7 FHIR
A standard for exchanging medical data over REST and JSON, gradually replacing the older HL7 v2. The most widely used version is R4. Without FHIR, integrating an AI system with an electronic medical record becomes a separate project.
Where to start
If a term from this glossary has come up in your project and it is not clear how it applies to your system, write to us. We will tell you what it means for you and where to start.
Talk to us
Related: RAG systems, AI strategy
SLAtech LTD