The output is unstable.

[Linz1248]

When I use JsonObjectConcept for concept extraction, I find that adding or removing fields in this data structure can affect the extraction results of the original fields. For example, the result of my first extraction is
"value": {
"name": "Jack",
"price": "50.0",
}.
However, when I add a new field "date", the output may become
"value": {
"name": "Smith",
"price": "50.0",
"date": "2025-8-10"
}.
The values of the original fields may change. What could be the reason for this?

[SergiiShcherbak]

@Linz1248 LLM outputs are not fully deterministic, so results may vary even with the same prompt and seed. When you add a new field to the JsonObjectConcept structure, the prompt also changes and gets slightly more complex, which can shift the model's interpretation.

Suggestions:

• Reduce sampling randomness: set temperature=0.0, top_p=0.0, and a seed (if your LLM supports it).
• Strengthen field definitions: be precise in field names (e.g. "name" -> "full_name") and description; add examples to anchor expected outputs.
• Use references/justifications: turn these on to see which text drives each value; helpful to diagnose shifts.
• Optionally test with another capable LLM.

Hope this helps.