120 lines
6.4 KiB
Python
120 lines
6.4 KiB
Python
import tempfile
|
|
import unittest
|
|
from pathlib import Path
|
|
from types import SimpleNamespace
|
|
from unittest.mock import Mock, patch
|
|
|
|
from google.genai import types
|
|
|
|
from copienator import configuration, prompting
|
|
from copienator.commands import correction, gemini_for_enonce, gemini_for_labels
|
|
from copienator.gemini_config import fresh_seed, thinking_config
|
|
|
|
|
|
class GeminiConfigTests(unittest.TestCase):
|
|
def test_fresh_seed_excludes_previous_even_at_int32_boundary(self):
|
|
for previous in (0, 1234, 2**31 - 1):
|
|
for random_value in (0, 2**31 - 2):
|
|
with patch("copienator.gemini_config.secrets.randbelow", return_value=random_value):
|
|
seed = fresh_seed(previous)
|
|
self.assertNotEqual(seed, previous)
|
|
self.assertTrue(0 <= seed < 2**31)
|
|
|
|
def test_correction_retries_change_seed_without_mutating_config(self):
|
|
client = Mock()
|
|
client.models.generate_content_stream.side_effect = [
|
|
TimeoutError(), RuntimeError("429 minute limit retry in 1s"),
|
|
[SimpleNamespace(text="result")],
|
|
]
|
|
config = types.GenerateContentConfig(seed=0, thinking_config=thinking_config("medium"))
|
|
with patch.object(correction, "client", client), \
|
|
patch.object(correction, "fresh_seed", side_effect=[10, 20, 30]), \
|
|
patch.object(correction, "tprint"), \
|
|
patch.object(correction, "stop_requested") as stop:
|
|
stop.is_set.return_value = False
|
|
stop.wait.return_value = False
|
|
self.assertEqual(correction.call_gemini_with_retries("model", [], config), "result")
|
|
configs = [call.kwargs["config"] for call in client.models.generate_content_stream.call_args_list]
|
|
self.assertEqual([cfg.seed for cfg in configs], [10, 20, 30])
|
|
self.assertEqual(config.seed, 0)
|
|
self.assertIsNone(config.http_options)
|
|
for cfg in configs:
|
|
self.assert_request_config(cfg, "medium")
|
|
self.assertEqual(cfg.http_options.retry_options.attempts, 1)
|
|
|
|
def assert_request_config(self, config, level):
|
|
data = config.model_dump(mode="json", by_alias=True, exclude_none=True)
|
|
for key in ("temperature", "topP", "topK", "thinkingBudget"):
|
|
self.assertNotIn(key, data)
|
|
self.assertNotIn(key, data.get("thinkingConfig", {}))
|
|
if level is None:
|
|
self.assertNotIn("thinkingConfig", data)
|
|
else:
|
|
self.assertEqual(data["thinkingConfig"], {"thinkingLevel": level.upper()})
|
|
|
|
def test_grading_live_and_batch_settings(self):
|
|
with tempfile.TemporaryDirectory() as directory:
|
|
image = Path(directory) / "answer.jpg"
|
|
image.write_bytes(b"image")
|
|
image.with_suffix(".json").write_text('[["01", 0, 400, 1.0, "Ex 1"]]')
|
|
for level in ("minimal", "low", "medium", "high", None):
|
|
with self.subTest(level=level), patch.object(prompting, "make_prompt", return_value="Grade"):
|
|
_, config = prompting.generate_request(directory, image, "Ex 1", thinking_level=level)
|
|
self.assert_request_config(config, level)
|
|
batch = prompting.batch_generation_config(level)
|
|
self.assert_request_config(types.GenerateContentConfig.model_validate(batch), level)
|
|
self.assertEqual(batch["responseSchema"], prompting.UNROLLED_SCHEMA)
|
|
# Let Gemini randomize each execution, including resubmission.
|
|
self.assertNotIn("seed", batch)
|
|
|
|
def test_labels_and_statement_use_their_own_settings(self):
|
|
with tempfile.TemporaryDirectory() as directory:
|
|
image = Path(directory) / "answer.jpg"
|
|
image.write_bytes(b"image")
|
|
with patch.object(configuration, "THINKING_LEVEL_FOR_LABEL", "low"):
|
|
_, config = gemini_for_labels.generate_request(image, "Ex 1", "Student", [], [])
|
|
self.assert_request_config(config, "low")
|
|
client = Mock()
|
|
with patch.object(configuration, "THINKING_LEVEL_LITE", "high"):
|
|
gemini_for_enonce._request_rubrics(client, "Question")
|
|
self.assert_request_config(client.models.generate_content.call_args.kwargs["config"], "high")
|
|
self.assertIsNone(client.models.generate_content.call_args.kwargs["config"].seed)
|
|
|
|
def test_correction_helpers_use_flash_level(self):
|
|
with patch.object(configuration, "THINKING_LEVEL_FLASH", "low"), patch.object(
|
|
prompting, "get_single_image_bytes", return_value=b"image"
|
|
):
|
|
requests = [
|
|
prompting.request_for_box_correction("answer.pdf", []),
|
|
prompting.request_for_wrong_label("answer.pdf", "Ex 1", "Question", "Ex 1"),
|
|
prompting.request_for_additional_answer("answer.pdf", "Ex 1", "Question", "Ex 1"),
|
|
]
|
|
for _, config in requests:
|
|
self.assert_request_config(config, "low")
|
|
|
|
def test_quota_fallback_switches_thinking_level_without_mutating_request(self):
|
|
client = Mock()
|
|
client.models.generate_content_stream.side_effect = [
|
|
RuntimeError("429 daily quota exhausted"), [SimpleNamespace(text="result")],
|
|
]
|
|
config = types.GenerateContentConfig(thinking_config=thinking_config("high"))
|
|
with patch.object(correction, "client", client), patch.object(
|
|
correction, "MODEL_ID_pro", "pro-model"
|
|
), patch.object(correction, "pro_quota_exhausted", False), patch.object(
|
|
configuration, "THINKING_LEVEL_FLASH", "low"
|
|
), patch.object(correction, "tprint"), patch.object(correction, "stop_requested") as stop:
|
|
stop.is_set.return_value = False
|
|
result = correction.call_gemini_with_retries("pro-model", [], config, "flash-model")
|
|
self.assertEqual(result, "result")
|
|
calls = client.models.generate_content_stream.call_args_list
|
|
self.assert_request_config(calls[0].kwargs["config"], "high")
|
|
self.assert_request_config(calls[1].kwargs["config"], "low")
|
|
self.assertEqual(calls[1].kwargs["model"], "flash-model")
|
|
self.assertNotEqual(calls[0].kwargs["config"].seed, calls[1].kwargs["config"].seed)
|
|
self.assertIsNone(config.seed)
|
|
self.assert_request_config(config, "high")
|
|
|
|
def test_invalid_level_fails_before_submission(self):
|
|
with self.assertRaisesRegex(ValueError, "Invalid Gemini thinking level"):
|
|
thinking_config("ultra")
|