Microsoft Foundry Generative AI Chat App Practice Question 9: Multi Turn Conversation State
Microsoft Applied Skills: Develop a Generative AI Chat App Using the Microsoft Foundry SDK · Test 1 · question 9 of 10
Testers report a bug in a Foundry chat app built on the Responses API. Asking "What is the size of France in square miles?" returns a correct answer, but the follow-up "And what is the capital city?" comes back with a request for clarification about which country is meant. Each turn calls responses.create() with only the new user message. What is the most direct fix?
Show answer and explanation
Answer:
Each responses.create() call is independent unless you tell the service how the turns relate. The documented pattern is to "chain turns by passing the previous response ID to previous_response_id," which carries the conversation forward without your code reassembling the history. The alternative, if you want to manage state yourself, is stateless input-array chaining: append the assistant's response.output to your input list and send the whole list on the next turn. Either way, the second call has to carry the first turn; sending only the new message is why the model can't tell which country is meant.
- A is wrong because max_output_tokens caps how long the reply can be. A longer reply to a question the model can't resolve is still a request for clarification. - B is wrong because streaming only changes how the response is delivered, token by token instead of all at once. It has no effect on what context the request carries. - C is wrong because temperature controls sampling randomness. The model is not discarding context; the context was never sent, so no sampling setting can recover it.
Reference: Use the Azure OpenAI Responses API