GroundingMetadata.Chunks is empty even when GroundingSearchesUsed is > 0

[eezdev]

Hi all,
There is a critical issue where GroundingMetadata.Chunks returns an empty list despite the model successfully performing web searches. While SearchEntryPoint contains valid HTML/chips and the model text contains citations, the programmatically accessible metadata required for verification is missing.

Technical Details:

Package: Google.GenAI

Models Tested: gemini-3-flash-preview, gemini-3-1

Feature: GoogleSearch

Observed Behavior:

Search Execution: GroundingMetadata.GroundingSearchesUsed reports a positive integer (e.g., 10 or 15), proving the tool was triggered.

Visual Grounding: The SearchEntryPoint.RenderedContent is correctly populated with HTML source chips and links.

Missing Data: GroundingMetadata.Chunks is empty (Count = 0).

Broken References: GroundingMetadata.Supports may contain GroundingChunkIndices, but since the Chunks list is empty, these indices point to non-existent data, making it impossible to map segments to URIs.

Impact:

Without the Chunks list, it is impossible to programmatically verify the provenance of the AI's claims or distinguish between a grounded citation and a text-hallucinated URL. This prevents the implementation of automated QA filters.

Example Response Snippet (Logs):

FinishReason: STOP Grounding searches used: 15 GroundingChunks: [] <-- Should contain URIs and Titles GroundingSupports: [indices pointing to non-existent chunks]

[shivvaam0001]

Hi @eezdev

Thank you for bringing this to our attention.
Could you please share the complete payload details along with a sample of the code you are using and the steps to reproduce the issue? This will help us replicate the problem and present it to the appropriate teams. We would like to reproduce the issue on our end.

[eezdev]

Hi @eezdev

Thank you for bringing this to our attention. Could you please share the complete payload details along with a sample of the code you are using and the steps to reproduce the issue? This will help us replicate the problem and present it to the appropriate teams. We would like to reproduce the issue on our end.

Hi @shivvaam0001, thank you for the follow-up.

To reproduce this issue, I have created a neutral example using an environmental research prompt. This demonstrates the instability where the model performs many searches but fails to return the structured metadata required to verify claims.

1. Reproduction Environment & Steps

• Package: Google.GenAI (latest)
• Model: gemini-3-flash-preview
• Step 1: Use a prompt that requires multiple specific data points (e.g., historical weather records and local environmental reports).
• Step 2: Observe that SearchEntryPoint contains HTML chips for all searches, but GroundingMetadata.Chunks is either empty or contains only a fraction of the sources.

2. Sample Code (Minimal Reproducible Example)

var config = new GenerateContentConfig
{
    Tools = new List<Tool> { new() { GoogleSearch = new GoogleSearch() } },
    Temperature = 0.1,
   MaxOutputTokens = 2000,
    SystemInstruction = new Content
    {
        Parts = new List<Part> { new Part { Text = "You are a climate researcher. Provide a detailed report with specific data points." } }
    },
    ResponseMimeType = "text/plain"
};

var userPrompt = "Provide the official 2024 rainfall totals for the Amazon Basin, the current water levels of Lake Mead, and list the official governmental agencies responsible for these reports.";

var response = await _client.Models.GenerateContentAsync(model: "gemini-3-flash-preview", contents, config: config);

var metadata = response.Candidates[0].GroundingMetadata;

// LOGGING THE DISCONNECT
Console.WriteLine($"Searches Executed: {CountGroundingSearches(response)}"); 
Console.WriteLine($"Chunks Count: {metadata.Chunks.Count}"); // CAN RETURN 0 or 1 despite many searches
Console.WriteLine($"HTML Chips Provided: {metadata.SearchEntryPoint != null}");


private static int CountGroundingSearches(GenerateContentResponse response)
{
    if (response?.Candidates == null)
        return 0;

    var count = 0;

    foreach (var candidate in response.Candidates)
    {
        var metadata = candidate?.GroundingMetadata;
        if (metadata == null)
            continue;

        if (metadata.WebSearchQueries != null)
            count += metadata.WebSearchQueries.Count;
        else if (metadata.GroundingChunks != null && metadata.GroundingChunks.Count > 0)
            count += 1;
    }

    return count;
}

3. Observed Payload Evidence (Logs)
   In this trace, the model executed 20 searches and successfully found enough data to provide deep links in the text, yet the groundingChunks array is empty. This prevents any programmatic verification.

Console Output:

FinishReason: STOP
Grounding searches used: 20
GroundingChunks Count: 0  <-- THE BUG
GroundingSupports Count: 0
SearchEntryPoint: [Contains valid HTML with 5+ chips for .gov and .org sites]

Partial Metadata JSON Trace:

{
    "searchEntryPoint": {
        "renderedContent": "<div>...Valid HTML containing multiple <a> tags to environmental reports...</div>"
    },
    "webSearchQueries": [
        "2024 rainfall totals Amazon Basin official report",
        "Lake Mead current water levels March 2026",
        "US Bureau of Reclamation Lake Mead levels",
        "...[17 more queries]..."
    ],
    "groundingChunks": [],   <-- instead of empty array this prop is not even in the returned json despite +10 Grounding searches used 
    "groundingSupports": []  <-- instead of empty array this prop is not even in the returned json despite +10 Grounding searches used 
}

4. Key Issues Observed

• Inconsistent Execution: The primary issue is instability. Running the exact same code and prompt multiple times results in different metadata outcomes. One execution may return a populated GroundingChunks list, while the next—under identical conditions—returns an empty list []/no GroundingChunks prop at all or a single source, even when GroundingSearchesUsed remains high (10–20 searches).
• Broken Programmatic References: When the Chunks list is empty or partially populated, the GroundingSupports indices are often broken.

Thanks for your help and available to provide more inputs if needed.