gemini‑2.0 suddenly drops confidenceScore when grounding with a data store ?

[Quentin_Lior]

Hi everyone,

Since Sunday 4 May 2025, every request I send to gemini‑2.0‑flash‑001 that is grounded with a Vertex AI Search data store comes back without the `confidenceScore` array inside `groundingMetadata.groundingSupport`.The grounding docs promise a numeric score (0–1) for each support chunk so we can audit the model’s answer and filter out weak references. (https://cloud.google.com/vertex-ai/generative-ai/docs/grounding/overview)

The schema example on the Grounding with your data page still shows the field: https://cloud.google.com/vertex-ai/generative-ai/docs/grounding/grounding-with-your-data

"groundingSupport": [
  {
    "segment": { "startIndex": 25, "endIndex": 147 },
    "segment_text": "ipsum lorem ...",
    "supportChunkIndices": [1, 2],
    "confidenceScore": [0.9541752, 0.97726375]
  }
]

and here as well.

What I get today (2.0 Flash 001 + data store grounding)

The same request now returns groundingSupport without confidenceScore — only segment, segment_text, and supportChunkIndices are present. However, it looks like with gemini‑2.5‑flash‑preview‑04‑17 with data store grounding → confidenceScores are present. This looks like a regression limited to the 2.0 Flash model.

Questions

1. Is the omission of confidenceScore in 2.0 Flash + data‑store grounding a known issue?

2. Has the response schema changed without documentation, or is this an unexpected bug?

Any insight or workaround would be greatly appreciated! I know that this feature is still in preview, but I'm quite surprised by how it suddenly stopped working.

Thanks in advance !

[a_aleinikov]

Hi @Quentin_Lior ,

Yes, this sounds like an unexpected regression specific to gemini‑2.0‑flash‑001 when grounding with Vertex AI Search.

- There has been no documented change to the response schema — the groundingMetadata.groundingSupport should still include confidenceScore as per the official docs.
- Since you confirmed gemini‑2.5‑flash‑preview‑04‑17 returns the scores, it suggests this issue is isolated to the 2.0 model.

I recommend:
- File a support ticket with Google Cloud → attach request + response samples.

[Quentin_Lior]

Thank you for your answer.

I just opened a ticket, will update once I have more info.