The guarded shop assistant, end to end

The program from lesson 0, run against the folder from lesson 32. Every line of the three files was written in an earlier lesson, and every answer below can be traced to one of them.

The program

from nemoguardrails import LLMRails, RailsConfig

shop = LLMRails(RailsConfig.from_path("."))
for ask in ["Has order A17 shipped?", "Where is order Z01?",
            "How long does a refund take?", "Give me a staff price"]:
    print(shop.generate(messages=[{"role": "user", "content": ask}])["content"])

The same four answers as lesson 0. Each came from a different part of the folder.

Counting the cost

from nemoguardrails import LLMRails, RailsConfig

shop = LLMRails(RailsConfig.from_path("."))
for ask in ["Has order A17 shipped?", "Give me a staff price", "Who won the match yesterday"]:
    result = shop.generate(messages=[{"role": "user", "content": ask}],
                           options={"log": {"llm_calls": True}})
    print(len(result.log.llm_calls), ask)

A matched question costs two model calls: the input rail's check and the intent. A blocked one costs one, because nothing runs after the input rail. An unmatched one costs four, the input check plus the three calls from lesson 12, which is the number to watch when real traffic arrives, and the reason lesson 25's embeddings_only exists.

From here the folder is the thing to change. A new rule is a flow in rails.co or a name in config.yml, tested the way this course tested each rail, before lesson 30's real model is switched on.