400 Invalid argument while using candidate_count>2 and long json in the prompt

Gemini API
1

Very specific issue, but I do not get this error when:

  • N=1
  • There is no JSON content in the prompt.

Is it just me? Anyone able to reproduce?

import json from vertexai.preview.generative_models import GenerativeModel, GenerationConfig # Create a JSON structure large enough to trigger the error (~100K chars) items = [] for i in range(1600): items.append({ "id": i, "name": f"Item {i}", }) test_json = json.dumps({"items": items}, indent=4) print(f"JSON size: {len(test_json):,} characters") # Initialize model model = GenerativeModel("gemini-2.0-flash") # Test: JSON with candidate_count=2 (will fail) prompt = f"Analyze this JSON data:\n\n{test_json}" # The key setting that triggers the error is candidate_count > 1 response = model.generate_content( prompt, generation_config=GenerationConfig(candidate_count=2) )
3

Hi @Mika_Myrseth

As you’re using Vertex AI I was curious whether it’s related to the platform. Here’s the same example for Google AI using the newer google-genai package.

import os import json from google import genai from google.genai import types from dotenv import load_dotenv from Imagen import response load_dotenv() # take environment variables from .env. client = genai.Client( api_key=os.environ['GOOGLE_API_KEY'] ) # Create a JSON structure large enough to trigger the error (~100K chars) items = [] for i in range(1600): items.append({ "id": i, "name": f"Item {i}", }) test_json = json.dumps({"items": items}, indent=4) print(f"JSON size: {len(test_json):,} characters") # Test: JSON with candidate_count=2 (will fail) prompt = f"Analyze this JSON data:\n\n{test_json}" # The key setting that triggers the error is candidate_count > 1 response = client.models.generate_content( model="gemini-2.0-flash", contents=prompt, config=types.GenerateContentConfig( candidate_count=1, ) ) print(response.text)

It’s giving me the same HTTP 400 error when candidate_count is > 1.

google.genai.errors.ClientError: 400 INVALID_ARGUMENT. {'error': {'code': 400, 'message': 'Request contains an invalid argument.', 'status': 'INVALID_ARGUMENT'}}

The token count (prompt ~ 46k) looks OK though.

However, using a different model works.

  • gemini-2.0-flash Error
  • gemini-2.0-flash-exp Error
  • gemini-2.0-flash-thinking-exp-01-21 OK
  • gemini-2.0-flash-thinking-exp-1219 OK
  • gemini-exp-1206 Timeout and OK
  • gemini-2.0-pro-exp-02-05 OK
  • gemini-2.0-flash-lite Error

OK results come with the following note
there are 2 candidates, returning text from the first candidate.Access response.candidates directly to get text from other candidates.

Seems that it is model-related.

Cheers

1 Like
4

Very interesting, thanks for testing!

1 Like