RAG & Agents Engineering

M37 lab, multi-agent applied builds

Test the router without calling a model.

The hotel coordinator’s router from notebook 02b, ported line for line: a confidence gate, a loop guard and routing by the agent a reply names. Move the threshold and the seven router unit tests run again. The table says which rule decided each route.

Tests passing
Routes the gate decided
Known issue
Guest asked to repeat at scores

Route your own reply

Compliance verdict

Route

What the router reads, oldest first

    The seven router tests

    From notebook 02b, Part 4

    TestConfidenceExpectedGot, and the rule that decided itResult

    Every test is a hand-built state: a guest message, an optional verdict tagged compliance, and the coordinator’s reply. END means the turn ends with the coordinator’s reply.

    Every threshold from 0 to 100

    ThresholdPassingKnown issue

    Computed by running the seven tests at all 101 thresholds and grouping runs with the same results. The notebook’s assessor prompt scores 80 to 100 for a clear intent, 50 to 79 for a clear intent with details missing, and 0 to 49 for an intent it cannot tell.