Railguarding Response

[anisbhsl]

LLM responses are sometimes inconsistent and we need a way to railguard those responses to prevent unwanted text from being thrown in response to ensure the responses are appropriate & accurate.

This might involve safeguarding against inappropriate content, user privacy and so on.

[NISH1001]

LLM responses are sometimes inconsistent and we need a way to railguard those responses to prevent unwanted text from being thrown in response to ensure the responses are appropriate & accurate.

This might involve safeguarding against inappropriate content, user privacy and so on.

Possibility
https://github.com/guardrails-ai/guardrails

Have to check. I am thinking having larch.railguards as a subpackage/module that will be applied as post-processing to any larch.search.SearchEngine component. I am also thinking about non-llm based, with some pattern-matching thing that can at least does 50% of heavylifting. LLM-based we can later do. Something like

Something like:

from larch.railguards import BasicRailGuard

railguard = BasicRailGuard(patterns=[...])

response = SearchEngine(...)(query=..., top_k=...)

response = raiguard(response)

Either this pattern or we can pass the railguard object to each SearchEngine constructor and if available, use that.

---