An input rail you wrote yourself

Lessons 16 and 17 built the parts. This lesson puts them together into the rail lesson 2 tried to write in plain Python, and the comparison is the point of the whole part.

The same check, as a rail

define bot too loud
  "Please do not shout at me."

define flow no shouting
  $verdict = execute check_shouting
  if $verdict.is_blocked
    bot too loud
    stop

A different check this time, so the two are easy to tell apart: a message in capitals gets a polite refusal. The shape is the same as self check input, which is the point.

Returning a decision, not a boolean

import pretend_nemo
from nemoguardrails import LLMRails, RailsConfig
from nemoguardrails.actions import action
from nemoguardrails.actions.rail_outcome import RailOutcome


@action(name="check_shouting")
async def check_shouting(context: dict):
    text = context.get("user_message", "")
    if text.isupper():
        return RailOutcome.block(reason="all capitals")
    return RailOutcome.allow()

RailOutcome is the object $response.is_blocked was reading back in lesson 14. block and allow are the two decisions used here, and reason is a sentence for a human reading a log.

rails = LLMRails(RailsConfig.from_path("."))
rails.register_action(check_shouting, "check_shouting")
for ask in ["WHERE IS MY ORDER", "When do I get my money back?"]:
    print(ask, "->", rails.generate(messages=[{"role": "user", "content": ask}])["content"])

Against lesson 2

The plain Python guard in lesson 2 had the check, the refusal sentence and the decision to stop all tangled in one function, in the middle of the application. Here the check is a function, the sentence is in rails.co, and the order the checks run in is a list in config.yml. Somebody who does not write Python can change the refusal.

It is also composable in a way the if was not. Add a second name to rails.input.flows and it runs after this one, sees the same context, and stops the turn the same way.