🤖 AI / ML

Large Language Model (LLM)

LLM · Foundation Model · GPT · Llama · Claude

Definition

An LLM (large language model) is a language model with billions of parameters, trained on large text corpora by autoregressive token prediction, capable of understanding language, following instructions and generating text.

How an LLM is built

An LLM is a transformer-architecture neural network with billions of parameters. Training runs
over enormous text corpora — web pages, books, code — by autoregression: the model learns to
predict the next token from the preceding ones.

After that base training the model goes through RLHF (reinforcement learning from human feedback),
a tuning pass on human ratings that makes it useful and safe in conversation.

Input context:  "Looking for a jacket for winter walks down to -4°F,
                 budget up to $200"

Tokenization → [Looking] [for] [a] [jacket] ... → 18 tokens
The transformer processes the context; attention weighs the relationships
The answer is generated token by token: [For] [temperatures] [down] [to] [-4°F] ...

Where LLMs are used in e-commerce

AI Shopping Assistant

The main commercial application of LLMs in retail is the conversational assistant. The shopper
describes the task in free form — “a gift for my mum’s 60th, she likes gardening, budget $40–70” —
the LLM reads the intent and, using RAG to reach the current catalogue, returns specific
recommendations with an explanation.

Semantic search

Exact-match search cannot handle synonyms or typos. LLM embeddings let you find products by
semantic proximity between the query and the description — “a case that doesn’t slip” surfaces
products with a “non-slip coating” attribute, even though the words never match.

Content generation

Writing product descriptions from attributes, SEO-oriented headlines and answers to common shopper
questions — these are the operational uses of LLMs that need no complex integration.

Limits in commercial products

Risk	What it looks like	Mitigation
Hallucination	The model invents specifications	RAG — facts from the catalogue only
Stale knowledge	The base model was trained on past data	Current context supplied through RAG
Latency	Token generation takes time	Streaming responses, context caching
Cost	LLM API calls cost more than ordinary queries	Caching, prompt optimisation

Frequently asked questions

How is an LLM different from an ordinary chatbot?

A classic chatbot follows rigid scripts — a decision tree, or template answers keyed to keywords. An LLM reads intent from arbitrary phrasing, keeps the context of the conversation and generates answers shaped to the request. Instead of “pick a command from the list”, an LLM handles “I want something warm for autumn walks, under $70”.

What are LLM hallucinations and how do you deal with them?

A hallucination is when an LLM generates a plausible but false statement. In e-commerce this is critical: the model must not invent product specifications. The fix is RAG (retrieval-augmented generation) — the LLM answers only from current catalogue data, not from its own parameters.

How are LLMs used in on-site search?

Semantic search built on an LLM copes with typos, synonyms and intents that never match a product name exactly. A query such as “what to wear to an office party” is decomposed into a category (clothing) and a context (formal event), and returns relevant products even though the keyword never appears verbatim.

Do you need to fine-tune an LLM on your store's data?

For the basic cases — search, description generation — no. Prompt engineering plus RAG over a current catalogue is enough. Fine-tuning earns its cost only for specific tasks with an unusual tone of voice or a subject area the base model never covered.

Gravity Field