A Databricks agent answers a follow-up question using only the words "what about last month?" as its retrieval query. The previous turn identified a specific region and product. What should the developer improve?

1. Embed only the phrase last month and increase vector dimension.
2. Increase output temperature so the model guesses the missing region.
3. Construct a standalone retrieval query using the relevant prior-turn entities.
4. Discard all conversation state before every retrieval call.