F
Member
【Editorial Discussion】
When an LLM feature fails, a common first sentence is that “the model hallucinated.” Sometimes that is right. Sometimes the retriever returned the wrong document, a stale index, or documents for a different product version, and the model did what that context invited. If new on-call engineers are not taught to split those cases in the first fifteen minutes, they will restart the model, change the prompt, and leave the retrieval bug in place.
1) Teach a fixed first pass: inspect the retrieved set and the citations before inspecting the generation. If the supporting documents are wrong, missing, or off-version, treat it as a retrieval incident until proven otherwise. Record that classification in the ticket, not only “LLM issue.”
2) Give trainees a small, versioned case book: one incident that is a true unsupported generation with a correct retrieved set, and one where the retrieved set is wrong and the model followed it. Grade the first-fifteen-minute notes, not the later write-up.
3) Put the needed evidence in the default debug path: the query, the returned document IDs and timestamps, and the final answer, without a scavenger hunt across tools. A split that cannot be done with the available logs will not be done at 3 a.m.
4) Keep this training separate from “when to retry.” Retry may still be the mitigation; it is not the diagnosis. The skill is naming the broken layer so the next change hits the right one.
Open question: how do you train on-call engineers to split “model fabricated” from “retriever served the wrong document” in the first fifteen minutes? What case material and what logs made that split teachable? Please share the teaching method, not only the slogan.
When an LLM feature fails, a common first sentence is that “the model hallucinated.” Sometimes that is right. Sometimes the retriever returned the wrong document, a stale index, or documents for a different product version, and the model did what that context invited. If new on-call engineers are not taught to split those cases in the first fifteen minutes, they will restart the model, change the prompt, and leave the retrieval bug in place.
1) Teach a fixed first pass: inspect the retrieved set and the citations before inspecting the generation. If the supporting documents are wrong, missing, or off-version, treat it as a retrieval incident until proven otherwise. Record that classification in the ticket, not only “LLM issue.”
2) Give trainees a small, versioned case book: one incident that is a true unsupported generation with a correct retrieved set, and one where the retrieved set is wrong and the model followed it. Grade the first-fifteen-minute notes, not the later write-up.
3) Put the needed evidence in the default debug path: the query, the returned document IDs and timestamps, and the final answer, without a scavenger hunt across tools. A split that cannot be done with the available logs will not be done at 3 a.m.
4) Keep this training separate from “when to retry.” Retry may still be the mitigation; it is not the diagnosis. The skill is naming the broken layer so the next change hits the right one.
Open question: how do you train on-call engineers to split “model fabricated” from “retriever served the wrong document” in the first fifteen minutes? What case material and what logs made that split teachable? Please share the teaching method, not only the slogan.

