Glossary of AI terms for eCommerce teams

An LLM reads a request written in plain language, keeps track of the conversation and answers in coherent natural language.

A shopper describes the job in their own words, and the assistant turns it into concrete product requirements: type, specs, price range and constraints.

A prompt sets the model's role, the rules and the answer format; prompt engineering is the work of making that answer controllable and useful in a given scenario.

On a product page the prompt can require answers grounded only in store data, comparisons across key specs, and a closing line that points to the next step toward purchase.

Tokens are the units of text a model measures input and output in: they drive context limits, response speed and the cost of processing.

Blending too many product descriptions into an answer makes the conversation slower and more expensive, so catalog context has to be assembled selectively.

The context window is how much information a model can take into account at once while it generates an answer.

In a TV consultation the context holds the shopper's request, the clarifying answers so far, a handful of relevant SKUs and the merchandising rules.

Alignment is the set of methods that make model behavior more useful, safe and predictable, including instruction tuning and reinforcement learning from human feedback.

In eCommerce that means the assistant does not invent facts, does not argue with catalog data and keeps its tone inside brand guidelines.

RAG combines retrieval over store data with generation, so the model answers from the live catalog, prices and stock instead of from memory.

A shopper asks which refrigerators fit under $800: the system first retrieves the models that qualify and only then writes the consultation.

Embeddings turn text, attributes and intent signals into vectors that can be compared by how close they are in meaning.

The phrase "fragrance-free cream for sensitive skin" can surface the right products even when the product page never uses those words.

Vector search finds items by closeness in meaning, and a vector database stores those representations and returns similar results fast.

That is how open-ended queries such as "a gift for a man who runs in winter" still surface products and catalog fragments.

Keyword search matches words, while semantic search reads the intent behind a query and maps it onto product attributes and descriptions.

When a shopper types "quiet robot vacuum for an apartment with pets", the semantic layer finds the right models even though no SKU name carries that wording.

Hybrid search blends lexical and semantic signals so that results respect both exact matches and the meaning behind the query.

It pays off in electronics and fashion, where shoppers mix precise specs with looser descriptions of the job or the style they are after.

Reranking is the second selection pass, where the system lifts the best-fitting options to the top of an already retrieved set.

After the first retrieval step the assistant can reorder results so the products that match both the request and the business rules come first.

Tool calling lets a model invoke external functions and data, and an agentic setup splits the task into several steps with a check on every result.

In the cart the assistant can check stock, confirm that two items are compatible and set up the next step toward checkout.

An AI shopping assistant helps the shopper pin down the need, make sense of the assortment and move to purchase without fighting through filters.

It works in search, on the product page and in the cart, where explaining differences, offering alternatives and settling doubts matter most.

Conversational commerce moves part of the shopping journey into a chat interface, where the shopper states the goal in their own words and gets help in return.

Instead of clicking through filters, the shopper describes the job and the system asks follow-up questions that lead to a fitting option.

Conversational recommenders work in rounds: the system gathers preferences, rebuilds the result set and explains why it picked what it picked.

On a laptop page the assistant can ask about the use case, the budget and the must-have specs, then offer alternatives without sending the shopper back to search.

Product discovery is the set of mechanics that help a shopper find the right product even when they do not know the exact model name or attribute.

A query such as "a lightweight stroller for the city and travel" should return a real shortlist, not a formal match on words.

Personalization adapts results and recommendations to the individual shopper, their interaction history and the context of the current session.

If a shopper usually picks certain brands and a certain price band, the assistant can shift recommendations into that zone right away.

These are the mechanics for offering a companion product, a better-fitting model or a ready-made bundle without pressuring the shopper.

On a smartphone page the assistant can pick compatible accessories and explain why those particular ones fit this shopper's scenario.

Catalog enrichment brings product data to a single shape: it aligns attributes, brands and units of measure and fills in the critical specs.

If one product page says USB-C and another says Type-C, the system has to treat both as the same signal when it matches products.

AI search quality is judged both on the relevance of results and on online behavior: clicks, add-to-cart, conversion and revenue.

Before a production launch you first check how well the system lifts the right products, then look at the effect on real business scenarios.

A/B testing compares a control and a test version of a scenario to measure the real effect of a change on the metrics that matter.

It is how you check whether the AI assistant adds to search conversion or add-to-cart against the shopper's current path.

LLMOps covers running an LLM application in production: quality monitoring, cost control, logging, security and shipping updates without losing control.

For an AI assistant that means tracking answers, errors, latency, cost, relevance regressions and data incidents.

A customer-facing AI scenario depends on response speed and unit economics: the slower and pricier the request, the harder it is to hold service quality at scale.

In search and guided selection, answering accurately is not enough: the answer also has to arrive within the shopper's patience window.

Hallucinations happen when a model produces a convincing but false answer. Grounded answers cut that risk by tying every claim to verifiable data.

In eCommerce this matters most for specs, compatibility, stock and product restrictions.

Prompt injection is an attempt to use shopper input or external content to make the model break its rules, reveal too much or take an action it should not take.

Once a system can call tools or pull in external data, defending against these attacks becomes a mandatory part of going to production.

PII is personally identifiable data, and data minimization means passing into an AI scenario only what the answer and the measurement actually require.

For a shopping assistant that can mean separating behavioral signals from identifying data and controlling what reaches logs and external models.