Quick definition
An LLM (Large Language Model) is an artificial intelligence model trained on enormous volumes of text to predict and generate language coherently. It can interpret natural-language questions, draft text, summarize information, reason over data it is given and, when connected to external tools, execute actions through agents.
What does it mean?
An LLM is built by training a neural network—typically with a transformer-style architecture—on massive amounts of text, adjusting billions of parameters until the model learns statistical patterns of how language fits together: grammar, facts, style, and the reasoning implicit in the training text. The result is not a database of memorized answers, but a system capable of generating new, coherent, contextually appropriate text.
A point that commerce vendors often omit from their definitions is that an LLM, on its own, has no access to a company's private or up-to-date data. Its knowledge is frozen at the time of its training. For an LLM to answer with a business's specific catalog, inventory, or policy data, it needs to connect to external sources—through RAG or tools exposed via protocols like MCP—rather than simply "knowing" them natively.
Modern LLMs can also use tools: they don't just generate text, they can decide to invoke an external function (searching a database, running a calculation, calling an API) as part of their response process, which makes them the core component of AI agents.
Why it matters
Before LLMs, interacting with a commerce system in free-form natural language—without being restricted to predefined commands or menus—was unreliable. LLMs solve that problem: they interpret intent in ambiguous, colloquial, or incomplete language, and can reason over provided context to produce a relevant response or action.
This enables real conversational interfaces in commerce—not chatbot scripts with fixed decision trees, but systems capable of understanding the infinite variations in how a shopper can express the same need.
How it works
An LLM processes input text by splitting it into tokens (word fragments) and computes, for each position, the probability of the next token given all preceding context. It generates text iteratively, predicting one token at a time, until it completes a response. The transformer architecture underlying most current LLMs uses a mechanism called attention, which lets the model weigh which parts of the input text are most relevant to each word it generates.
In commerce applications, the LLM rarely operates alone: it receives additional context retrieved through RAG, can invoke external tools or APIs via protocols like MCP, and operates within an agent flow that decides when to generate text and when to execute an action.
Applied example in AI Commerce
A shopper writes to a store assistant: "I want something similar to what I bought last month but one size up and not as expensive." The LLM interprets the full intent—reference to a previous purchase, size adjustment, price constraint—and, supported by data retrieved from the CDP (purchase history) and the PIM (catalog and pricing), generates a response with concrete options, instead of asking the user to rephrase with exact keywords.
Related concepts
An LLM is the generation component within RAG, which supplies it with context retrieved from an external source. It frequently works together with Embeddings and a Vector Database to access relevant information. It is the reasoning core of an AI Agent, and it connects to external tools through protocols like MCP. It is also the underlying technology behind conversational Semantic Search and AI Personalization.
Common mistakes
Assuming an LLM natively "knows" current data about a specific company: without connections to external sources, its knowledge is limited to what existed at training time. Confusing "large model" with "always-correct model": an LLM can generate fluent, confident-sounding responses that are factually wrong (hallucinations), especially without mechanisms like RAG anchoring it to verifiable data. Finally, believing all LLMs are interchangeable: they vary enormously in cost, speed, reasoning capability, and quality depending on the provider and model size.
The Edgebound Labs perspective
In the lab we don't evaluate an LLM by how impressive its answers sound, but by how well it anchors to real data when connected to the client's business. A brilliant model without access to up-to-date inventory generates convincing but potentially false answers—and in commerce, a false answer about availability or price has a direct cost. Measuring before trusting is the method.
Frequently asked questions about LLM
What does LLM stand for?
Large Language Model: an AI model trained on large volumes of text to generate and understand natural language.
Does an LLM have access to my product catalog automatically?
No. It needs to connect through RAG or external tools (for example, via MCP) to access a business's specific, up-to-date data.
What is a hallucination in an LLM?
It's when the model generates a response that appears certain but is actually incorrect or invented, usually due to a lack of verifiable context.
Can LLMs execute actions, not just generate text?
Yes. When combined with external tools and an agent flow, they can invoke functions or APIs to execute concrete tasks.
Are all LLMs equally good for commerce?
No. Performance varies by model, size, cost, and how well it connects to the specific business data.
Is retraining an LLM the way to keep it up to date?
It's not the most practical approach for data that changes constantly; RAG solves the freshness problem without retraining the model.
Keep exploring the glossary
Applying LLM in your operation?
We audit your commerce stack and tell you exactly what you need to scale with AI — no generic slide decks, with clearly defined success metrics.