Glossary: AI Terms Every Lighting Specifier Should Know

AI terminology travels fast and lands inconsistently. The same word means different things in a research paper, a vendor pitch, and a press release. For lighting professionals evaluating AI product tools — or trying to understand what vendors are actually offering — a grounded vocabulary makes the difference between a useful conversation and an expensive misunderstanding.

This glossary defines the terms that come up most often in AI product intelligence conversations, with plain-language explanations and examples drawn from lighting industry contexts.

Large Language Model (LLM)

A Large Language Model is the AI system that reads a question and produces a response in natural language. It has been trained on an enormous volume of text and has learned the patterns of how language works — including how to explain technical concepts, how to compare products, and how to answer questions in a conversational way.

What an LLM does not inherently know is your specific product catalog. Without additional configuration, an LLM answering a question about your luminaires is drawing on whatever product information it encountered during training — which may be outdated, incomplete, or from a different manufacturer entirely.

In practice: An LLM that is configured to work with your specific product documentation is what makes an AI product assistant possible. The LLM provides the language intelligence; your catalog provides the product knowledge.

Source grounding

Source grounding means that an AI answer is derived from a specific document or set of documents, and that the connection between the answer and the source can be traced and verified.

A grounded AI product assistant answering a question about a luminaire's CRI will point back to the specific spec sheet and section from which it retrieved that value. An ungrounded assistant may produce the same answer — or a plausible but incorrect one — without any traceable basis.

Why it matters: In product sales and specification, answers are acted upon. A wrong CRI value or a fabricated compatibility claim leads to real downstream consequences. Source grounding is what makes AI answers auditable rather than just convenient.

Hallucination

Hallucination is the term for when an AI produces a response that sounds fluent and confident but is factually incorrect or entirely fabricated. The name comes from the fact that the AI "sees" something that isn't there — it generates language that follows the patterns of a correct answer without having the factual basis for one.

Hallucinations are a well-documented behaviour of general-purpose AI systems and occur most often when the model is asked about something outside its training data, or when it has conflicting or sparse information about a topic.

In a lighting context: A hallucinating AI product assistant might confidently provide a part number that doesn't exist, cite a lumen output for the wrong CCT, or describe a compatibility between a housing and trim that has never been manufactured. The output reads exactly like a correct answer.

The primary defence against hallucination in product contexts is source grounding — restricting the AI to answer only from documents you have provided, and flagging when a question falls outside the scope of those documents.

Knowledge base

In the context of AI product tools, a knowledge base is the collection of documents that have been processed and made queryable by the AI. It typically includes spec sheets, product catalogs, ordering guides, compatibility tables, application notes, and any other structured or unstructured documentation that contains product information.

The quality and completeness of the knowledge base directly determines the quality of AI answers. An AI can only tell you what's in its knowledge base — gaps in the documentation become gaps in the answers.

Key consideration: A knowledge base is not static. When your catalog changes — new products, revised specs, updated compatibility rules — the knowledge base needs to be updated to reflect those changes. How quickly and easily an AI tool allows this update is an important evaluation criterion.

Inference

Inference is the act of the AI generating a response. When a user asks a question, the AI processes the question, retrieves relevant information, and produces an answer — this entire process is called inference. The time it takes is called inference latency, and it's what you experience as the delay between asking a question and seeing the response.

Why it matters: In a sales workflow, response speed affects whether an AI tool is practical to use during a live customer conversation. An AI that takes 8–12 seconds to respond to a simple compatibility question is disruptive to a conversation. Most production product intelligence tools are optimised to respond within 1–3 seconds for standard queries.

Context

In AI, context refers to the information the model has available when generating a response. This includes the current question, any relevant documents retrieved from the knowledge base, and the history of the current conversation.

Context is what allows a well-designed AI product assistant to carry project information across a conversation. If a rep mentions at the start of a conversation that they're specifying for a remodel project with a 3.5-inch ceiling hole, a context-aware system will apply that constraint to subsequent questions — flagging that a 4-inch housing won't fit — without the rep having to repeat it.

Practical implication: Context awareness is what separates an AI product assistant from a search bar. A search bar returns the same result regardless of what you told it two questions ago. A context-aware AI builds on the conversation.

Accuracy vs. confidence

These two concepts are frequently confused in AI conversations. Accuracy refers to whether an answer is factually correct. Confidence refers to how certain the AI appears when giving its answer.

The critical point is that AI systems can produce high-confidence wrong answers. A hallucinated part number is delivered with exactly the same tone as a correct one. There is no audible uncertainty, no qualifier, no "I'm not sure about this." The answer simply arrives, stated as fact.

This is why source grounding matters more than any measure of AI confidence. A traceable answer that you can verify is more useful than a confident answer that you cannot.

Token

A token is the basic unit of text that an AI model processes. Roughly speaking, one token is about three-quarters of a word in English. The reason tokens matter in an AI product context is that most AI systems have a limit on how many tokens they can process in a single interaction — this is called the context window.

For product knowledge applications, the context window determines how much documentation the AI can actively reference when answering a question. A larger context window means more documents, more product detail, and more complex queries can be handled in a single interaction.

Why it comes up: When vendors talk about handling large catalogs, they're often referring to how they manage the relationship between catalog size and context window constraints. The practical question is whether the tool performs well when your entire catalog needs to be queryable, not just the subset that fits in a single context window.

Fine-tuning vs. retrieval-based approaches

These two terms describe different methods for making an AI knowledgeable about your specific products. They have very different implications for data security, update speed, and answer reliability.

Fine-tuning involves additional training of an AI model on your product data, so the model "learns" your catalog as part of its weights. The result is a model that has your product knowledge baked in — but which cannot easily be updated when the catalog changes, cannot trace its answers to source documents, and carries data security risks related to information being encoded in the model itself.

Retrieval-based approaches keep your product documentation in a separate database and retrieve relevant sections at query time. The AI reads those sections to formulate its answer. Your data stays in your database — not inside the model. Updates are immediate. Every answer can be traced to its source document.

For product knowledge applications — where the catalog changes frequently and answer accuracy is critical — retrieval-based approaches are the standard for production-grade deployments.