I found guardrails AI really useful for my research on LLMs. Otherwise I would have wasted a lot of time trying to curate the outputs from my experiments.
Don’t cram too much in a single prompt though. Prompt structures like guard rails naturally carry high cognitive load for the LLM, which leads to biased outputs. I found the best practice is to alleviate it by using multiple prompts and using a guard rail as an end-step rather than one big prompt for the LLM. (https://arxiv.org/abs/2402.01740)
Pretty interesting to go from a world of deterministic code, to LLMs which can do incredible things, but unreliably. In a world of LLMs, I could imagine guardrails being a table-stakes part of engineering an ML system, just like unit tests, and CI/CD would be for traditional software.
Don’t cram too much in a single prompt though. Prompt structures like guard rails naturally carry high cognitive load for the LLM, which leads to biased outputs. I found the best practice is to alleviate it by using multiple prompts and using a guard rail as an end-step rather than one big prompt for the LLM. (https://arxiv.org/abs/2402.01740)