Gemini 2.5 Flash Lite hallucinates grounding chunks

Gemini API
,
1

I’m making calls to Gemini 2.5 Flash Lite model from an Android app, using Firebase AI SDK.

I add Google Search in the tool list to enable grounding. I make a request. Response includes groundingMetadata with groundingSupports and inside groundingSupports there are chunk indexes for references. But, groundingChunks list is empty. So the references are to an empty list.

It looks like the model is hallucinating references. Eg. response:

{
  "groundingMetadata": {
    "groundingAttribution": [],
    "groundingChunks": [],
    "groundingSupports": [
      {
        "groundingChunkIndices": [
          0
        ],
        "segment": {
          "endIndex": 2062,
          "partIndex": 0,
          "startIndex": 1892,
          "text": "However, the museum always offers a rich program of talks, family activities, and tours, so it's always worth checking their schedule upon arrival for any ongoing events."
        }
      },
      {
        "groundingChunkIndices": [
          0
        ],
        "segment": {
          "endIndex": 2199,
          "partIndex": 0,
          "startIndex": 2091,
          "text": "The British Museum is generally open from 10:00 AM to 5:00 PM, with extended hours on Fridays until 8:30 PM."
        }
      },
      {
        "groundingChunkIndices": [
          0
        ],
        "segment": {
          "endIndex": 2341,
          "partIndex": 0,
          "startIndex": 2200,
          "text": "While general admission is free, booking a timed entry slot in advance is recommended, especially during peak times, to guarantee your entry."
        }
      },
      {
        "groundingChunkIndices": [
          0
        ],
        "segment": {
          "endIndex": 2690,
          "partIndex": 0,
          "startIndex": 2605,
          "text": "The entire museum is accessible, with lifts and ramps available for wheelchair users."
        }
      }
    ],
    "retrievalQueries": [],
    "webSearchQueries": []
  }
}
3

Hello,

Welcome to the Forum,

We have a dedicated team for support on Firebase relates issues, we recommend raising your concern on Firebase support for more accurate and prompt response.

5

It’s not really about the SDK though. It should be using the same API and same LLM.

Is it known that Gemini 2.5 Flash Lite hallucinates grounding?

7

Hello,

Could you please share the full payload details that you are using for the API call? We would like to reproduce your issue on our end to analyze it better and come up with a more accurate response.

9

Here’s the code I’m using (Kotlin for Android):

val modelName = "gemini-2.5-flash-lite"

val model = Firebase.ai.generativeModel(
    modelName = modelName,
    tools = listOf(Tool.googleSearch()),
)

model.generateContent(promptContent)

and here’s the promptContent referenced above:

Generate a summary visitor guide for the following point of interest:
Name: The British Museum
Address: Great Russell St, London, WC1E 7JW, United Kingdom
Categories: museum, tourist_attraction
It should be a simple, fun bullet list with emojis. Give the most important 3-4 list items. Start immediately with the first bullet, no intros.

It’s pretty simple really.

In the response I get, there’s grounding reference indices (inside groundingSupports) but it’s all fake, nothing in groundingChunks section, that part is an empty list.

Pretty easy to reproduce.

11

Hello,

We would recommend reaching out to the Firebase Support team, as they are best equipped to assist you with this concern.