Integration guides ยท 2026-09-22

Structured outputs and minimal reasoning_effort on Solar Pro 2 for data extraction

How to combine Upstage Solar Pro 2's structured outputs support with its lowest reasoning_effort level for a predictable, low-latency enterprise data extraction flow.

Diagram showing text sent to Solar Pro 2 converted into low-latency, fixed-shape JSON output through minimal reasoning_effort and a structured outputs schema.

Minimal is a distinct point among four reasoning_effort levels

Solar Pro 2's catalog page lists four reasoning_effort levels: minimal, low, medium, and high; that's one step lower than the three-level (low/medium/high) structure many models offer. The model also carries `structured_outputs` as a separate flag; unlike simply requesting free-form JSON with `json_mode`, this is a mechanism that guarantees the output conforms to a specific schema.

Using these two features together lets you build a data extraction flow that conforms to a fixed schema and returns with low latency.

When to use minimal, when to go higher

On a relatively simple, single-step extraction task like pulling an amount and a date from an invoice, using `reasoning_effort: "minimal"` together with `structured_outputs` lets the model produce a schema-conforming response directly without running a long thinking process; this keeps cost and latency low on a high-volume, latency-sensitive production pipeline.

On a more complex task, like resolving conflicting information across multiple documents or cross-checking the extracted data's consistency, a higher reasoning_effort level lets the model evaluate more thoroughly before drawing a conclusion; the minimal level can fall short in that case.

  • Simple, single-step extraction: the minimal level plus structured_outputs gives low latency.
  • Conflict resolution, cross-checking tasks: a higher level gives a more reliable result.
  • Test task complexity against a small sample set first, then pick the level accordingly.

The thinking trace is not returned in the response

Solar Pro 2's catalog page states that the model's reasoning trace is not returned in the response. That's not a downside for a data extraction flow only interested in the final structured output, but if your use case wants to inspect the model's intermediate steps for debugging, you need to know this limit upfront.

Frequently asked questions

Are structured_outputs and json_mode the same thing?

No. json_mode guarantees the output is valid JSON but doesn't enforce a schema; structured_outputs guarantees the output conforms to a specific schema. Solar Pro 2 lists both as separate capabilities.

Does the minimal level always give the lowest cost on every task?

It usually gives the lowest latency, but on a complex task it can fall short and require a retry, making net cost higher than a higher level would have been. Test against your task's complexity.

Related posts