Local LLM Incident Battlecards
Recall the first response line for common local-LLM operating incidents.
battlecard Data leak The bot returned another customer's name. Possible local RAG disclosure. Pause the endpoint, snapshot the index and logs, capture the prompt and retrieved chunks, then scope before rebuilding. Deleting the index first erases the evidence that defines scope. Containment and preservation keep the incident analyzable. Bad answer The local model contradicted policy. Potential hallucination or stale retrieval. Save the answer, prompt, citations, retrieved chunks, and model version before changing prompts or models. Swapping models first hides whether retrieval or review caused the failure. Analysis needs the chain from input to evidence to output. Overload Endpoint overload: retry…
Sign up free — one personalized lesson every day, matched to your role and goals.
Already have an account? Sign in