Grounding / RAG

Why this bot doesn't make things up

The number one fear people have about an AI chatbot is that “the bot will give out wrong information and do real damage” — this is the mechanism that prevents it.

Retrieve first, answer second

Rather than letting the AI answer from its own memory.

The customer asks

The system reads both the current message and the earlier conversation history.

Search the shop's real data

Prices and stock come straight from the shop's database; the bot does not invent numbers.

Filter for relevance

Anything retrieved that isn't genuinely related to the question is discarded.

Compose the answer

From that information only.

Edge cases

If nothing is found, or the system goes down

Nothing found = no guessing

The bot says honestly that it doesn't have that information yet and offers to put the customer in touch with an admin — it never guesses or makes anything up.

There is a fallback search

If searching by meaning turns up nothing, the system falls back to keyword search so the customer isn't met with silence.

The AI service fails

The bot sends a polite status message — raw data and error messages are never thrown out at the customer.

An admin can take over

Turn the bot on or off in one click; while it is off, every customer message waits for an admin to answer in person.

What that means for the business: you never have to worry that the bot will give out wrong information and do real damage — quoting the wrong price, saying something is in stock when it isn't, or claiming benefits the shop has never claimed.

Scope of control

How much the shop controls

You can adjust these

Your type of business — telling the bot what your shop sells, which shapes its judgement about which questions are in scope to answer
Bot persona — tone of voice, how it addresses people and how formal it is, to match your brand
Canned messages — the greeting, the reply to out-of-scope questions, the message when the system is unavailable, and the message when a product has no description

If the shop sets nothing at all, the system uses the platform's standard values, which have already been tuned.

You can't adjust these (by design)

The bot may not claim product benefits beyond what the shop has provided — this matters enormously for beauty and supplement businesses.
The bot may not disclose internal system information.
Message length is capped so it stays readable on LINE.

These rules sit at the platform level and shops cannot change them — they protect the shop and the customer alike.

Security

How shop data and customer data are looked after

LINE connection credentials are encrypted

Stored encrypted, and the interface never displays the real values back.

Every message signature is checked

The signature on every message from LINE is verified each time, to confirm it really came from LINE and isn't a forgery.

Row-Level Security

Data isolation at the database level — one shop's data cannot be reached from another shop's context, not even at the level of a database query.

bcrypt + tokens that expire

Passwords are hashed with bcrypt and never stored readably · tokens expire, with a renewal mechanism and immediate revocation. A permission change takes effect on access straight away.

Rate limiting

Protects against flooding the system and against abuse.

No cloud keys stored on the machine

Authentication is handled by the cloud provider, which reduces the risk of a leaked key.

Cancelling the service and deleting data (in line with PDPA guidance) — when a shop stops using the service, there is a procedure that exports the shop's data first and then permanently deletes everything from every system: the database, the search index and the stored files.