Most chatbots are easy to notice for the wrong reason.
They answer one message at a time. They forget what you said a minute ago. They sound confident even when they are guessing. And when you ask something practical, like whether a product is available or what happened to an order, they often push you toward a generic support page.
That is not how people actually shop for fragrance.
Buying perfume is personal, but it is also practical. You may want something fresh for office, warm for winter, below a certain budget, available in 10ml, and not too loud around other people. You may ask one thing, change your mind, come back later, then ask for more options. A useful chat should follow that flow without making you start from zero every time.
That is the idea behind the Fragscentric chat.
Not to make it feel flashy. Not to make it pretend it knows everything. The goal is simpler: make it easier to choose, check, and continue.
It Remembers the Conversation
A regular chatbot often treats every message like a fresh ticket.
You say you want something for summer. Then you mention a budget. Then you ask for more options. Instead of connecting those messages, the bot responds as if each one arrived in isolation.
That gets tiring fast.
Our chat keeps useful context from the conversation. If you mention that you prefer fresh scents, dislike oud, want something for work, or need to stay within a budget, the next answer can use that context. If it already showed you a few products, it can avoid repeating the same cards when you ask for more.
This matters because fragrance advice is rarely a single-question problem. People refine their taste as they talk.
Sometimes the first request is vague: “Find me something good.”
Then it becomes clearer: “Not too sweet.”
Then clearer again: “Something I can wear daily in hot weather.”
A better chat should improve as the conversation improves.
It Uses Store Data, Not Just Nice Words
There is a big difference between a chatbot that can talk about fragrance and a shopping assistant that can help you buy the right thing.
For Fragscentric, product names, prices, stock, collections, order links, and policy details should come from the site data. Not from the model’s memory. Not from a confident guess. Not from a sentence that sounds good but may be wrong.
If the chat shows a product card, that card is built from validated store data. If it talks about price or stock, that answer should match what the site knows right now. If it links somewhere, the link should be a real internal route or an approved support link.
This is important because customers do not benefit from creative answers when the question is factual.
If a product is unavailable, the chat should not recommend it as if it can be ordered. If a price changed, the answer should follow the current price. If a policy question comes up, the answer should use the actual policy, not a generic ecommerce answer.
The chat can sound natural, but the facts need to stay grounded.
It Does Not Try to Force Every Question Into One Flow
One problem with many chatbots is that they get stuck.
If you start tracking an order, some bots keep treating every next message as an order-tracking message. You might ask about a perfume right after that, and the bot still asks for an order number.
That feels robotic because it ignores what you are actually saying.
The Fragscentric chat is designed to move between contexts. If you are asking about an order, it handles that as an order flow. If you switch back to fragrance advice, it should leave the order flow and continue normally.
That is a small detail, but it changes the experience. Real conversations are not perfectly organized. People correct themselves. They send a short phone number, then fix it. They type an incomplete order number, then send the full one. They ask one thing and then another.
The chat should handle that without making the customer feel at fault.
It Is Account-Aware, But Still Careful
If you are signed in and ask about your order, the chat should use your account order data instead of asking you to prove the order again like a guest.
That is the convenient path.
But convenience should not mean careless access. Guest order tracking still needs verification, because an order number alone should not expose private order details. Signed-in account context and guest tracking are different situations, so the chat treats them differently.
This is also why empty states matter.
If you are signed in and have no orders, the chat should simply say that and point you to the shop. If every order is already delivered, it should not pretend there is an active order to track. Clear answers are better than filler.
Smart Cards Make the Next Step Easier
Sometimes a good answer should not end with just text.
If the chat recommends a fragrance, a product card lets you open the detail page quickly. If it finds an order, an order card gives you the important details without making you search through your account. If it suggests a collection or shop page, the link is right there.
The point is simple: smart cards keep the conversation connected to quick navigation and useful details, so the next step is easier.
It Knows When Not to Guess
A good assistant should be helpful, but it should also know its limits.
If the chat cannot safely answer something from store data, it should not invent an answer. If a customer needs human help, it should point them to Messenger or WhatsApp. If a product detail needs to be checked before ordering, the answer should make that clear.
For fragrance advice, a little interpretation is useful. For stock, pricing, account, delivery, and policy questions, guessing is not useful.
That balance is the point.
For Technical Readers: The Short Version
The chat is not just a prompt wrapped in a bubble.
The language model handles natural phrasing, but deterministic code handles facts and flows. Recent messages, lightweight preferences, shown products, and active workflows are stored locally so the conversation can continue after a reload. Product, collection, stock, price, policy, account, and order answers are routed through site data where possible before the model is used.
After the model responds, structured output is validated. Product slugs must exist. Cards are generated from known records. Unsupported links and unsafe factual claims are rejected or replaced with a grounded fallback.
That architecture matters because a store assistant has two jobs: understand the customer and respect the source of truth.
The model helps with the first part. The site data protects the second.
The Real Goal
The goal is not to replace the human side of Fragscentric.
The goal is to make the first step easier.
If you already know what you want, the chat should help you get there faster. If you are unsure, it should help narrow the field. If you come back later, it should remember enough to continue. If the question is factual, it should check the site instead of guessing.
That is what makes a fragrance chat useful.
Not because it talks more.
Because it makes the next decision easier.
