Vietnamese LLM for Retail and Consumer Goods in Practice

01/10/2026

An online home-goods shop has two thousand SKUs, and about half of the descriptions are pasted straight from the supplier, typos included. Whoever manages content there knows that fixing them one by one is unrealistic. This is the kind of problem where a Vietnamese LLM fits retail: high volume, templated text, and small errors are tolerable if there is a review step. But do not rush, because each use has its own trap.

Shopping chatbot: do not let it make promises for you

Customers ask "does this pot work on an induction hob", "will size M fit someone 165 centimeters tall", "when will my order arrive". A chatbot on an LLM, connected to product data and the order system, can answer in natural Vietnamese even when the customer types without diacritics, with abbreviations or slang. That is an edge over scripted bots, which only work when the customer types the expected keyword.

The trap is that a chatbot tends to say what the customer wants to hear. Ask "can this shirt go in the washing machine", and if the data is silent the model may say "yes" to be polite. The defense is to force it to use only the product data, and when information is missing to say plainly that it does not have it and hand over to staff. The same applies to promotions, prices and warranty: anything that can become a legal or commercial commitment must come from the official data source, not from the model's own phrasing.

Situations to hand to a person

• The customer is upset or threatening a complaint.
• A return request outside the standard policy.
• Products with a safety dimension, like cosmetics for sensitive skin or food for people with allergies.
• A question the chatbot has answered twice without resolving it.

Product descriptions: fast, but with a guardrail

Writing descriptions for thousands of products from a spec sheet is something an LLM does reasonably well. Feed in material, dimensions and use, and get back text in a consistent brand voice, already tuned for search. A benefit few people think of is consistency: the whole catalog sounds as if one person wrote it.

Two failures keep showing up, though. The first is embellishment: the spec says cotton, and the generated text claims "excellent sweat absorption, antibacterial", which is not in the data. For consumer goods, especially food, cosmetics and supplements, lines like that can be false advertising. The second is descriptions that all sound alike, stuffed with phrases such as "delivering a great experience". A shopper spots machine writing at first read. The remedy is to give the model real samples of the brand's voice and forbid it from inferring beyond the input data.

Customer feedback analysis: the quiet value

Star ratings, marketplace comments, page messages, complaint emails. Very few teams read them all. An LLM can, and it groups them into themes: slow delivery in which region, which product draws complaints about smell, which packaging model arrives dented. Vietnamese social media text is a real test: abbreviations ("ko", "dc", "sp"), missing diacritics, slang, sarcasm, emoji standing in for feeling. Is "delivered so fast, only three weeks" praise or sarcasm? Machines often guess wrong, so sentiment analysis still needs human spot checks.

I advise against using it only to score sentiment and draw pretty charts. Use it to answer specific questions: what did customers complain about more this week than last? Is that complaint concentrated in one batch? When you attach a verbatim customer quote, the product team trusts the result far more.

Choosing a model

In retail, cost and speed matter as much as quality because query volume is large. Often a compact model finetuned on industry data and brand voice fits better than a huge general model used for everything. AIVISION trains and finetunes LLMs for Vietnamese and for specific domains, and has released L1.0, an LLM for Vietnamese. I am not quoting performance or cost figures, because they depend on how you deploy.

How to begin

Pick a narrow category, say five hundred products from one department. Generate descriptions, have editors review them, and count how much they change and in what way. That number tells you whether to tune the model or fix the input data. More about the models is on the AIVISION site.