Siri Grounded Response Generation

Design or case-study a Siri response-generation system where answers must stay grounded in tool results, avoid hallucination, fit a voice-chat length budget, and remain evaluable as context grows. Requirements Design a Siri response-generation flow for an assistant that calls tools and then speaks a concise answer. You should cover: How the system grounds the response in tool output instead of unsupported model memory. How to evaluate factuality, usefulness, and voice-chat length. Why hallucination happens and what product / model guardrails reduce it. How to handle long context from conversation history, user state, and tool payloads. How to keep the answer short enough for spoken interaction without dropping required facts. Notes A strong answer separates retrieval / tool execution, evidence selection, response generation, verification, and final voice formatting. Treat the tool result as the source of truth; the model should summarize, transform, or ask for clarification rather than invent missing data. Evaluation should mix offline labeled sets, adversarial queries, tool-result consistency checks, human preference review, and online metrics such as correction rate, repeat-query rate, and user abandonment. For long context, discuss summarization with citations to retained evidence, recency windows, structured memory, and hard limits on what the generator may rely on. Preparation Practice a 10-minute whiteboard flow: user query -> intent -> tool call -> evidence extraction -> grounded generation -> verifier -> voice response. Prepare a compact hallucination taxonomy: missing evidence, stale memory, ambiguous tool result, prompt injection, and over-compression. Drill eval metrics for conversational assistants: factuality, helpfulness, latency, answer length, and user-correction signals. Be ready to explain why voice UX changes the design: shorter responses, explicit uncertainty, and follow-up questions beat long text-style answers.