SimpleChat does not work with data model properly.

[polina-c]

It happens often for TextField in simple-chat-like renderers.

May be it makes sense to add validation to our evals to check for leading '/' in text field value.

---

2 comment(s) from the original issue

polina-c commented on 2026-05-11:

@gspencergoog , @jiahaog

What do you think about adding validation to evals to catch this issue?

Or it will be better to add e2e test to simple chat to try to catch this?

---

jiahaog commented on 2026-05-11:

It sounds like the a2ui json is not generated by the model correctly? Agreed on evals, we can probably add a set of programatic validations for evals in the A2UI repo.

[josemontespg]

A2UI Triage: @jiahaog any updates on this?

[polina-c]

This is how it looks like in json:

        {
          "id": "story_input_field",
          "component": "TextField",
          "variant": "longText",
          "label": "Your story goes here...",
          "value": "/user_story"            <------- error !!!!!!!
        },
        {
          "id": "submit_story_button",
          "component": "Button",
          "child": "submit_label",
          "action": {
            "event": {
              "name": "submit_user_story",
              "context": {
                "story_text": {
                  "path": "/user_story"
                }
              }
            }
          },
          "variant": "primary"
        },

[polina-c]

@gspencergoog

This thing appears too often for me in simple-chat-like apps.

It seems we want to test cover it and fix system prompt.

Is it complicated to update our evals to catch such cases?

@jacobsimionato , FYI

[gspencergoog]

The evals don't work like that: They are more like benchmarks than tests: they make sure we reach a certain "correct" percentage. We can test that we succeed a certain amount of the time, but can't test that we don't fail in a certain way. We could add more cases that might cause this to happen and make sure they are correctly generated, and if they aren't, we can do prompt engineering on the descriptions and prompt to help improve the success rate. (and it may also have to do with which model is doing the generation).

[polina-c]

Thank you and yes, it is what I mean: this issue is happening too often and we want to figure out how to reduce percentage with prompt engineering. At least we want to repro and measure it.

Another thing we can do is to compensate in the TextField component:

• in applications like simple chat initial value never starts with '/'
• in 99% of apps initial value for text field is always empty and it may be we want to remove it from schema at all

@gspencergoog, thoughts?