Updates

.env.example

@@ -34,7 +34,7 @@ DEPARTMENT_GENERATION_MODEL=gpt-5.4-nano
 LECTURE_GENERATION_MODEL=claude-sonnet-4-6
 LECTURE_SUMMARY_MODEL=gpt-5.4-nano
 PROFESSOR_GENERATION_MODEL=gpt-5.4-nano
-TOPICS_GENERATION_MODEL=gemini-3.5-flash
+TOPICS_GENERATION_MODEL=gemini-3.6-flash
 IMAGE_GENERATION_MODEL=gemini-3.1-flash-lite-image
 TTS_VOICE_MODEL=eleven_flash_v2_5
 
@@ -45,6 +45,7 @@ TTS_VOICE_MODEL=eleven_flash_v2_5
 # claude-sonnet-4-6
 # claude-haiku-4-5
 # gemini-3.1-pro-preview
+# gemini-3.6-flash
 # gemini-3.5-flash
 # gpt-5.5
 # gpt-5.4-mini

artificial_u/api/routers/jobs.py

@@ -117,7 +117,7 @@ def _coerce_int(value: Any) -> Optional[int]:
     return None
 
 
-def _job_link_path(
+def _job_link_path(  # noqa: C901
     payload: Any, result: Any, factory=None, cache: Optional[dict] = None
 ) -> Optional[str]:
     """

artificial_u/config/settings.py

@@ -140,7 +140,7 @@ class Settings(BaseSettings):
     # Professor generation model
     PROFESSOR_GENERATION_MODEL: str = "gpt-5.4-nano"
     # Topics generation model
-    TOPICS_GENERATION_MODEL: str = "gemini-3.5-flash"
+    TOPICS_GENERATION_MODEL: str = "gemini-3.6-flash"
     # Course tags generation model
     TAGS_GENERATION_MODEL: str = "gpt-5.4-nano"
     # Image generation model
@@ -220,7 +220,7 @@ def set_default_model(cls, v, info):
             if backend == "openai":
                 return "gpt-5.4-nano"
             elif backend == "gemini":
-                return "gemini-3.5-flash"
+                return "gemini-3.6-flash"
             elif backend == "anthropic":
                 return "claude-sonnet-4-6"
             else:

artificial_u/services/content_service.py

@@ -32,6 +32,27 @@ def _is_gemini_3_plus(model: str) -> bool:
     return bool(match) and int(match.group(1)) >= 3
 
 
+def _deprecates_gemini_sampling_params(model: str) -> bool:
+    """Check if a Gemini model deprecates sampling parameters (temperature, top_p, top_k).
+
+    Starting with Gemini 3.6 Flash and Gemini 3.5 Flash-Lite, sampling parameters
+    are deprecated and ignored (and will return a 400 error in future model generations).
+    """
+    if not model or not model.startswith("gemini-"):
+        return False
+
+    if "gemini-3.5-flash-lite" in model:
+        return True
+
+    match = re.match(r"gemini-(\d+)(?:\.(\d+))?", model)
+    if match:
+        major = int(match.group(1))
+        minor = int(match.group(2)) if match.group(2) else 0
+        return (major, minor) >= (3, 6)
+
+    return False
+
+
 class ContentService:
     """
     Service for generating text content using various AI models/backends.
@@ -628,19 +649,23 @@ async def _generate_gemini(
             # which in turn burns the output budget and truncates the response).
             is_gemini_3_plus = _is_gemini_3_plus(model)
 
-            if temperature is not None:
-                effective_temperature = temperature
-            elif is_gemini_3_plus:
-                effective_temperature = 1.0
-            else:
-                effective_temperature = DEFAULT_TEMPERATURE
-
-            # Create generation config with all parameters including system_instruction
+            # Create generation config with max_output_tokens
             config_params = {
-                "temperature": effective_temperature,
                 "max_output_tokens": max_tokens if max_tokens is not None else DEFAULT_MAX_TOKENS,
             }
 
+            # Sampling parameters (temperature, top_p, top_k) are deprecated starting
+            # with Gemini 3.6 Flash and Gemini 3.5 Flash-Lite. Omit temperature for models
+            # that deprecate it to avoid errors or ignored parameter warnings.
+            if not _deprecates_gemini_sampling_params(model):
+                if temperature is not None:
+                    effective_temperature = temperature
+                elif is_gemini_3_plus:
+                    effective_temperature = 1.0
+                else:
+                    effective_temperature = DEFAULT_TEMPERATURE
+                config_params["temperature"] = effective_temperature
+
             # Control reasoning effort on models that support it (Gemini 3+). Using a
             # lower thinking level reserves more of the output budget for the visible
             # response and reduces the risk of MAX_TOKENS truncation.

cdk/cdk/cdk_stack.py

@@ -264,7 +264,7 @@ def __init__(self, scope: Construct, construct_id: str, **kwargs) -> None:
             "STORAGE_LECTURES_BUCKET": lectures_bucket.bucket_name,
             "STORAGE_REGION": self.region,
             "STORAGE_TYPE": "s3",
-            "TOPICS_GENERATION_MODEL": "gemini-3.5-flash",
+            "TOPICS_GENERATION_MODEL": "gemini-3.6-flash",
             "TTS_VOICE_MODEL": "eleven_flash_v2_5",
             # Database connection pool settings (conservative for db.t4g.small ~110 max_connections)
             # These ensure the app uses a shared connection pool and doesn't exhaust RDS connections

docs/configuration.md

@@ -148,7 +148,7 @@ LECTURE_GENERATION_MODEL=claude-sonnet-4-6
 PROFESSOR_GENERATION_MODEL=gpt-5.4-nano
 
 # Topics generation model
-TOPICS_GENERATION_MODEL=gemini-3.5-flash
+TOPICS_GENERATION_MODEL=gemini-3.6-flash
 
 # Image generation model (gemini-3.1-flash-lite-image, gemini-3.1-flash-image, or gemini-3-pro-image)
 IMAGE_GENERATION_MODEL=gemini-3.1-flash-lite-image
@@ -312,7 +312,7 @@ TESTING=true
 | `DEPARTMENT_GENERATION_MODEL` | Model for department generation | `gpt-5.4-nano` | No |
 | `LECTURE_GENERATION_MODEL` | Model for lecture generation | `claude-sonnet-4-6` | No |
 | `LECTURE_SUMMARY_MODEL` | Model for lecture summary generation | `gpt-5.4-nano` | No |
-| `TOPICS_GENERATION_MODEL` | Model for topics generation | `gemini-3.5-flash` | No |
+| `TOPICS_GENERATION_MODEL` | Model for topics generation | `gemini-3.6-flash` | No |
 | `PROFESSOR_GENERATION_MODEL` | Model for professor generation | `gpt-5.4-nano` | No |
 | `IMAGE_GENERATION_MODEL` | Model for image generation (`gemini-3.1-flash-lite-image`, `gemini-3.1-flash-image`, `gemini-3-pro-image`) | `gemini-3.1-flash-lite-image` | No |
 | `TTS_VOICE_MODEL` | Model for text-to-speech voice | `eleven_flash_v2_5` | No |

mise.toml

@@ -1,5 +1,5 @@
 [tools]
+pnpm = "11.15.1"
 node = "24"
-pnpm = "latest"
 python = "latest"
 uv = "latest"

tests/unit/test_content_service.py

@@ -11,7 +11,10 @@
 
 import pytest
 
-from artificial_u.services.content_service import ContentService
+from artificial_u.services.content_service import (
+    ContentService,
+    _deprecates_gemini_sampling_params,
+)
 
 
 def _build_service():
@@ -175,3 +178,95 @@ async def test_generate_anthropic_uses_sampling_params_for_sonnet_4_6(monkeypatc
     call_kwargs = mock_client.messages.create.await_args.kwargs
     assert "thinking" not in call_kwargs
     assert call_kwargs["temperature"] == pytest.approx(0.3)
+
+
+@pytest.mark.parametrize(
+    "model,expected",
+    [
+        ("gemini-3.6-flash", True),
+        ("gemini-3.5-flash-lite", True),
+        ("gemini-3.5-flash", False),
+        ("gemini-3.1-pro-preview", False),
+        ("gemini-3.1-flash-lite", False),
+        ("gemini-4.0-flash", True),
+        ("claude-sonnet-4-6", False),
+        ("", False),
+    ],
+)
+def test_deprecates_gemini_sampling_params(model, expected):
+    assert _deprecates_gemini_sampling_params(model) is expected
+
+
+@pytest.mark.asyncio
+async def test_generate_gemini_omits_temperature_for_gemini_3_6_flash(monkeypatch):
+    service = _build_service()
+
+    part_mock = MagicMock()
+    part_mock.text = "hello from gemini"
+    part_mock.thought = False
+
+    candidate_mock = MagicMock()
+    candidate_mock.content.parts = [part_mock]
+
+    response_mock = MagicMock()
+    response_mock.candidates = [candidate_mock]
+
+    mock_client = MagicMock()
+    mock_client.aio.models.generate_content = AsyncMock(return_value=response_mock)
+    monkeypatch.setattr("artificial_u.services.content_service.gemini_client", mock_client)
+
+    result = await service._generate_gemini(
+        prompt="Test prompt",
+        model="gemini-3.6-flash",
+        system_prompt="Test system",
+        temperature=0.3,
+        max_tokens=2048,
+        prefill=None,
+        thinking_level="low",
+    )
+
+    assert result == "hello from gemini"
+    call_kwargs = mock_client.aio.models.generate_content.await_args.kwargs
+    config = call_kwargs["config"]
+    # Config should not have temperature set because gemini-3.6-flash deprecates sampling params
+    assert not hasattr(config, "temperature") or getattr(config, "temperature", None) is None
+    assert config.max_output_tokens == 2048
+    assert (
+        getattr(
+            config.thinking_config.thinking_level, "value", config.thinking_config.thinking_level
+        ).lower()
+        == "low"
+    )
+
+
+@pytest.mark.asyncio
+async def test_generate_gemini_includes_temperature_for_gemini_3_5_flash(monkeypatch):
+    service = _build_service()
+
+    part_mock = MagicMock()
+    part_mock.text = "hello from gemini 3.5"
+    part_mock.thought = False
+
+    candidate_mock = MagicMock()
+    candidate_mock.content.parts = [part_mock]
+
+    response_mock = MagicMock()
+    response_mock.candidates = [candidate_mock]
+
+    mock_client = MagicMock()
+    mock_client.aio.models.generate_content = AsyncMock(return_value=response_mock)
+    monkeypatch.setattr("artificial_u.services.content_service.gemini_client", mock_client)
+
+    result = await service._generate_gemini(
+        prompt="Test prompt",
+        model="gemini-3.5-flash",
+        system_prompt="Test system",
+        temperature=None,
+        max_tokens=2048,
+        prefill=None,
+    )
+
+    assert result == "hello from gemini 3.5"
+    call_kwargs = mock_client.aio.models.generate_content.await_args.kwargs
+    config = call_kwargs["config"]
+    assert config.temperature == 1.0