Skip to content
anythingai.space
Conversational AIintermediate13 min · 9 steps

LLM-Based Assistants

How one language model runs a whole conversation from its context window, and how system prompts, few-shot examples, RLHF, tools and retrieval shape what it does.

context0tok
grounded answers0%
3 requests in 1out-of-scopeONE MESSAGE, THREE THINGSmake it two roomsslot changedoes it have a pool?hotel factis March cold there?off-topicUserIntent classifiertop score 0.22“Didn't understand”
requesterrortokenresponsedatasignalheavy worktool call
metrics are illustrative
1 / 9
ProblemStep 1 of 9

Users don't stay on the script

A classical booking bot is mid-flow, waiting for a check-in date. The user writes: “Actually make it two rooms, and does it have a pool? Also, is March cold there?”

That's a slot change, a hotel-facts question and an off-topic question in one message. None matches a trained intent, so the bot falls back to “Sorry, I didn't understand.”

Watch outRigid intents break on exactly the messages people send most: corrections, combinations and small talk.

Lesson 4 of 8 in Conversational AI