Execution Plan for an LLM Answer-Search Product With a Fixed Team and Deadline

You are interviewing for a product manager role at OpenAI. The interviewer sets a hard budget, six engineers and six months, and asks how you would build a search experience that works like ChatGPT: people ask in natural language and get a conversational answer, rather than only a ranked list of links. Walk through what you would build within that budget, what you would leave out, how you would staff and sequence the work, and how you would know whether it is working. ```hint Start from the budget Turn six people and six months into a rough amount of usable engineering time, then decide which parts of a conversational search stack you would build yourself, which you would reuse or license, and which narrow slice of queries the first version has to handle well. ``` ### Clarifying Questions - Is the search over the open web, or over a narrower body of content such as one company's documents or a product catalog? - Are the six engineers the whole team, or do design, research, data science, legal and trust-and-safety support come from elsewhere? May the team use existing language models and a licensed web index? - What must be true at the six-month mark: a public launch, a limited beta, or a validated prototype? - Is this a standalone product or a search mode inside an existing chat assistant? ### What a Strong Answer Covers - A sharp scope choice: which users and query types the first version serves, and what is explicitly out of scope for six months - Realistic build, buy or reuse calls for the index, the model and the safety layer given only six engineers, and how those engineers are allocated - A sequenced plan with monthly milestones, decision points, and cut lines for when work slips - A quality bar specific to generated answers (grounding and citation accuracy, freshness, latency, cost per query) backed by an evaluation set - Launch metrics and guardrails, plus the main risks such as wrong answers, prompt injection from web pages and dependence on a search provider, with mitigations ### Follow-up Questions - Two months in, evaluation shows answers are accurate but p95 latency is double your target. What do you cut or change? - How would you decide which queries get a generated answer and which fall back to plain links? - If leadership asked you to ship in three months instead of six, what would you drop first?