LLM output is untrusted
Here is a critical rule: LLM output is untrusted input. Just like you validate user input from a web form, you must validate what the LLM returns before acting on it. It can return invalid room names, out-of-range temperatures, or malformed data.
# Real examples of bad LLM outputs:
# Wrong room name
{"room": "ktichen", "temp": 22} # Typo!
# Invalid temperature
{"room": "kitchen", "temp": "warm"} # String, not number
# Missing required field
{"room": "kitchen"} # No temperature!
# Hallucinated room
{"room": "garage", "temp": 22} # No garage in system!LLMs can return typos, wrong types, missing fields, or hallucinated values. Never trust raw output.
Trust boundary between LLM and your system
Quiz: Quiz
Loading practice…