Guardrails
Also known as: Safety filters, Content moderation
Pre- and post-processing layers that block bad inputs/outputs and enforce policy on top of an LLM.
Input/output filters and policies layered around an LLM to block disallowed content, prevent prompt injection, enforce output schemas, or gate sensitive actions. Implemented via classifiers, regex, smaller LLMs, or rule engines.
In practice
Every production LLM app needs them. Interviews probe layered defense, latency budget, and false-positive trade-offs.
Related topics
Related terms
Hallucination
When a model confidently makes up something that isn't true.
LLM Evaluation
Measuring whether an LLM does what you want, beyond "looks fine to me".
Perplexity
Exp(average cross-entropy) on held-out text; lower means the model is less surprised by real data.
Reasoning Model
An LLM trained to reason at length internally before answering. Slower and more expensive, but much better on hard problems.
MMLU (Massive Multitask Language Understanding)
Multiple-choice benchmark across 57 academic subjects; the standard "raw knowledge" headline number.
HumanEval
Code-generation benchmark: 164 problems with hidden unit tests, scored by whether the generated code passes.