大模型聊天机器人
A chat assistant that understands what was asked, rather than matching keywords.
The chatbots most people have met were keyword machines. They matched phrases to canned replies, and the moment a customer typed something slightly different they collapsed into "Sorry, I did not understand that." Everyone learned to type "agent" immediately.
A chatbot built on a language model reads the question the way a person does. "Do you deliver to Klang?", "can send to klang ah?" and "delivery available Selangor?" are the same question, and it answers all three from your actual delivery policy.
The engineering that matters is not the model — anyone can call an API. It is grounding: making sure it answers from your price list and your policies rather than from its own imagination, and that when it does not know, it says so and fetches a person instead of inventing a price.
Enquiries come in at 9pm. Nobody sees them until morning. By then two of the five have bought elsewhere, and your team spends the first hour re-typing the same answers about delivery and sizing.
All five get an accurate answer within seconds of asking. Three self-serve completely. Two are qualified, captured with name and number, and sitting in your inbox as leads when you open your laptop.
How we actually build it
-
01
We collect what it should know
Price list, delivery terms, returns policy, opening hours, the ten questions you answer daily. If these only exist in someone's head, writing them down is the first deliverable and it is valuable on its own.
-
02
We ground it in those documents
It retrieves the relevant passage and answers from it, rather than generating from memory. This is what stops a chatbot inventing a discount you never offered.
-
03
We define the handover
Which questions it must never attempt — complaints, disputes, anything about money it is not certain of — and how it passes a person the full conversation, not just a name.
-
04
We put it where customers already are
Your site, WhatsApp, Instagram. One assistant, several front doors.
- You answer the same ten questions every day across WhatsApp, web and social
- Enquiries arrive outside working hours and go cold by morning
- You want leads qualified before a person spends time on them
- Your questions are genuinely complex and each answer is bespoke
- You have no written policies or price list for it to learn from — that comes first
- The conversation is emotionally sensitive. Complaints belong with a person
- Your price list and policies in writing
- A WhatsApp Business number if you want that channel
- Someone to review the first week of conversations
1–2 weeks for web; add a week for WhatsApp Business approval.
What people ask about this
Will it make things up?
Grounded properly, it answers only from your documents and says "let me get someone" outside them. That behaviour is designed in and tested, not hoped for.
Can it speak Malay and Chinese?
Yes, and it will usually reply in whatever language the customer used. Worth testing with real phrasing including the mixed English-Malay people actually type.
Often paired with
Would this work for you?
The first call is free, and we will say plainly if this is the wrong tool for your problem.
预约免费通话