Chatbot

The scripted era

Early chatbots were lists of rules. Find a word that was set in advance somewhere in the incoming sentence, and send back whatever answer got attached to that word. A slightly more advanced version first sorted the sentence into one of a handful of categories, then pulled a prepared answer for that category.

The upside was clear: since every answer was a pre-reviewed sentence, nothing the company never approved could go out, and you could always explain why a given answer came out. This approach still handles the counters that quote policy or amounts today.

The problem was that people always phrase things differently. Catching "can I get a refund," "I want my money back," and "can I return this" as the same request meant filling in thousands of sample phrasings by hand, and doing that whole job over again every time a new product launched.

A counter that composes sentences

Today's chatbot doesn't pick a script. It builds the answer piece by piece, choosing the word most likely to come next after everything said so far, then looking again. That's why questions nobody prepared for still get an answer, and the same question can come back worded slightly differently each time.

Tone can shift the same way. Slip an instruction like "be friendly," "keep it short," or "use formal language" in ahead of the question, and the whole feel of the answer changes. In the old approach, changing tone meant rewriting the entire script.

Reading company documents alongside it

A language model learned from the world's writing, not your company's refund policy. Asked without that knowledge, it mixes in some rule from somewhere else on the internet and produces a sentence that sounds plausible anyway.

So a support chatbot today usually looks up the passage related to a question in internal documents first, hands that passage to the model along with the question, and tells it to answer based on that material. This method is called RAG (Retrieval-Augmented Generation).

Do it this way, and updating a policy just means swapping the document, no retraining required. Being able to cite which document and page an answer came from is a major upside too.

Why it seems to remember

A chatbot is, by default, meeting you for the first time with every message. Conversation appears to continue only because the entire exchange so far gets attached and resent along with each new question.

That causes two things once a conversation runs long: each message sent gets bigger, so answers slow down, and once the total crosses a limit, the earliest parts get cut off. Seeming to forget a name you gave at the very start, after a long chat, comes from exactly this, which is why long support conversations often get summarized and only the summary gets resent.

It doesn't just answer, it acts

Today's support counters don't stop at giving information. They take an order number and call up shipping status, change a reservation, or hand things off to a human agent when it gets too complicated. The model itself doesn't run these — it decides which function to call with which values, and a program actually carries it out.

That's why a well-built chatbot has clear boundaries for where it stops answering: hand off to a person when it's unsure, and always confirm once more before anything involving money.