Intermediate
Open
Pro
Choosing a Tool-Choice Mode for Structured Extraction
You need a pipeline step that reads a free-text customer email and
always outputs a structured object: {sentiment, requested_action, urgency}. There is no external system to call — you just want
reliable structured data out of every email, with no exceptions.
- Which tool-choice mode would you use, and why is it a better fit here than prompting the model to "respond in JSON" without a tool?
- What could still go wrong even with that mode set correctly, and how would you defend against it?
- Contrast this with a case elsewhere in the same pipeline where "auto" is clearly the right mode instead. Why does that case differ?
Share this question