conductor-oss/conductor · error · RuntimeException

No embeddings returned from Gemini API

Error message

No embeddings returned from Gemini API

What it means

GeminiVertex.generateEmbeddings() throws this when the Gemini embedContent API returns a response whose embedding object or its values list is null. This means the API call succeeded (no IOException) but produced no usable embedding vector — the response shape indicates either the model cannot embed the input, or the API returned an empty/null result for an edge-case input.

Source

Thrown at ai/src/main/java/org/conductoross/conductor/ai/providers/gemini/GeminiVertex.java:88

    public String getModelProvider() {
        return NAME;
    }

    @Override
    public List<String> getProviderAliases() {
        return List.of(ALIAS);
    }

    @Override
    public List<Float> generateEmbeddings(EmbeddingGenRequest embeddingGenRequest) {
        try {
            GeminiApi.EmbedContentResponse resp =
                    geminiApi.embedContent(
                            embeddingGenRequest.getModel(),
                            embeddingGenRequest.getText(),
                            embeddingGenRequest.getDimensions());
            if (resp.embedding() == null || resp.embedding().values() == null) {
                throw new RuntimeException("No embeddings returned from Gemini API");
            }
            return resp.embedding().values();
        } catch (RuntimeException e) {
            throw e;
        } catch (Exception e) {
            throw new RuntimeException("Gemini embedContent failed", e);
        }
    }

    @Override
    public ChatModel getChatModel() {
        return new GeminiChatModel(geminiApi);
    }

    @Override
    public ChatOptions getChatOptions(ChatCompletion input) {
        return GeminiChatOptions.builder()
                .model(input.getModel())

View on GitHub (pinned to cf7c3e4a8a)

Solutions

1. Verify the model name is a Gemini embedding model: 'text-embedding-004', 'gemini-embedding-001', or 'text-embedding-005'.
2. Check the input text is non-empty and non-blank before calling generateEmbeddings.
3. Log the raw API response to diagnose whether the embedding field is truly null vs. malformed.
4. Verify the API key has access to the embeddings endpoint (separate from chat access).

Example fix

// before
String model = "gemini-2.5-flash"; // chat model — won't produce embeddings
List<Float> emb = vertex.generateEmbeddings(
    new EmbeddingGenRequest(model, text, null));

// after
String model = "text-embedding-004";
List<Float> emb = vertex.generateEmbeddings(
    new EmbeddingGenRequest(model, text, null));

// Validate embedding model and input before calling
private static final Set<String> GEMINI_EMBEDDING_MODELS =
    Set.of("text-embedding-004", "text-embedding-005", "gemini-embedding-001");

void validateEmbeddingRequest(String model, String text) {
    if (text == null || text.isBlank()) {
        throw new IllegalArgumentException("Embedding input text must be non-empty");
    }
    if (!GEMINI_EMBEDDING_MODELS.contains(model)) {
        throw new IllegalArgumentException(
            "Use a Gemini embedding model (text-embedding-004, gemini-embedding-001). Got: " + model);
    }
}

try {
    return vertex.generateEmbeddings(request);
} catch (RuntimeException e) {
    if (e.getMessage().equals("No embeddings returned from Gemini API")) {
        throw new IllegalArgumentException(
            "Gemini returned no embedding — verify model is an embedding model "
            + "(text-embedding-004) and input is non-empty", e);
    }
    throw e;
}

When it happens

Trigger: Calling embedContent with a model that doesn't support embeddings (e.g. a chat-only or image-only model), passing empty or whitespace-only text, or hitting an API edge case where Google returns a valid HTTP 200 with a null embedding field.

Common situations: Using a chat model name (e.g. 'gemini-2.5-flash') for embeddings instead of an embedding model (e.g. 'text-embedding-004', 'gemini-embedding-001'). Empty input text. API key enabled for chat but not for the embeddings endpoint. Model deprecated or not available in the region.

Related errors

• Gemini embedContent failed
• Empty response downloading from {url}
• Empty response downloading image from {url}
• {"error":{errors},"response":{output}}
• Not supported

AI-assisted analysis of conductor-oss/conductor@cf7c3e4a8a (2026-08-14). Data as JSON: /api/errors/1108ea139b96bc3b. Report an issue: GitHub.