update Gemini API : use thinking levels

copienator/commands/correction.py

@@ -20,6 +20,8 @@ from copienator import configuration as config
 from copienator.commands import grouping
 from copienator.answer_origins import save_origins
 from copienator import prompting
+from copienator.gemini_config import thinking_config
+from copienator import configuration as settings
 from copienator.feedback_boxes import valid_feedback_box
 from copienator.latex_json import repair_latex_json
 from copienator import (
@@ -279,6 +281,9 @@ def call_gemini_with_retries(model_id, contents, config,
         # Switch to fallback immediately if quota was exhausted by another thread
         if model_id == MODEL_ID_pro and pro_quota_exhausted and fallback_model_id:
             model_id = fallback_model_id
+            config = config.model_copy(update={"thinking_config": thinking_config(
+                settings.THINKING_LEVEL_FLASH
+            )})
 
         try:
             full_response_text = ""
@@ -312,6 +317,9 @@ def call_gemini_with_retries(model_id, contents, config,
             if is_quota_error and model_id == MODEL_ID_pro and fallback_model_id:
                 tprint(f"\tGemini Pro quota hit ({e}). \n\n\tFalling back to Flash permanently...")
                 model_id = fallback_model_id
+                config = config.model_copy(update={"thinking_config": thinking_config(
+                    settings.THINKING_LEVEL_FLASH
+                )})
                 pro_quota_exhausted = True
                 continue # Retry immediately with Flash
 
@@ -506,7 +514,11 @@ def process_single_task(task_tuple, precomputed_response=None):
                     flash_count += 1
 
         try:
-            contents, config = prompting.generate_request(INPUT_DIR, file_path, label)
+            contents, config = prompting.generate_request(
+                INPUT_DIR, file_path, label,
+                thinking_level=(settings.THINKING_LEVEL_FLASH if use_flash
+                                else settings.THINKING_LEVEL_PRO),
+            )
             model_to_use = MODEL_ID_flash if use_flash else MODEL_ID_pro
 
             if precomputed_response:
@@ -891,13 +903,10 @@ def run_configured(args: argparse.Namespace) -> ExitCode:
                                 ],
                             }
                         ],
-                        "generation_config": {
-                            "temperature": 1.0,
-                            "topP": 0.95,
-                            "maxOutputTokens": 65535,
-                            "responseMimeType": "application/json",
-                            "responseSchema": prompting.UNROLLED_SCHEMA,
-                        },
+                        "generation_config": prompting.batch_generation_config(
+                            settings.THINKING_LEVEL_FLASH if use_flash
+                            else settings.THINKING_LEVEL_PRO
+                        ),
                     },
                 }
                 line = json.dumps(request)

copienator/commands/gemini_for_enonce.py

@@ -26,6 +26,7 @@ from copienator import (
     workspace_from_args,
 )
 from copienator import configuration as config
+from copienator.gemini_config import thinking_config
 from copienator.filesystem import staged_files
 from copienator.latex_json import normalize_literal_newlines, repair_latex_json
 from copienator.platform import validate_windows_labels
@@ -249,7 +250,7 @@ def _request_rubrics(client, group_context_text: str):
         config=types.GenerateContentConfig(
             automatic_function_calling=types.AutomaticFunctionCallingConfig(disable=True),
             system_instruction="Rédige tous les barèmes et consignes de notation en français. Conserve les clés JSON, les labels et les formules mathématiques.",
-            temperature=0.2,
+            thinking_config=thinking_config(config.THINKING_LEVEL_LITE),
             response_mime_type="application/json",
             response_json_schema=GroupRubrics.model_json_schema(),
         ),
@@ -370,7 +371,7 @@ def refine_existing(workspace: EvaluationWorkspace, mode: str, *, api_client=Non
             ))])],
             config=types.GenerateContentConfig(
                 automatic_function_calling=types.AutomaticFunctionCallingConfig(disable=True),
-                temperature=0.1, response_mime_type="application/json",
+                thinking_config=thinking_config(config.THINKING_LEVEL_LITE), response_mime_type="application/json",
                 response_json_schema=LabelGroups.model_json_schema(),
             ),
         )
@@ -488,7 +489,7 @@ def process_exam(
 
     config_1 = types.GenerateContentConfig(
         automatic_function_calling=types.AutomaticFunctionCallingConfig(disable=True),
-        temperature=0.1,
+        thinking_config=thinking_config(config.THINKING_LEVEL_LITE),
         response_mime_type="application/json",
         response_json_schema=ExamQuestions.model_json_schema(),
     )
@@ -505,7 +506,7 @@ def process_exam(
 
     config_2 = types.GenerateContentConfig(
         automatic_function_calling=types.AutomaticFunctionCallingConfig(disable=True),
-        temperature=0.1,
+        thinking_config=thinking_config(config.THINKING_LEVEL_LITE),
         response_mime_type="application/json",
         response_json_schema=ExamSolutions.model_json_schema(),
     )
@@ -540,7 +541,7 @@ def process_exam(
 
     config_3 = types.GenerateContentConfig(
         automatic_function_calling=types.AutomaticFunctionCallingConfig(disable=True),
-        temperature=0.1,
+        thinking_config=thinking_config(config.THINKING_LEVEL_LITE),
         response_mime_type="application/json",
         response_json_schema=ExamContext.model_json_schema(),
     )

copienator/commands/gemini_for_labels.py

@@ -15,6 +15,7 @@ from google.genai import types
 from pydantic import BaseModel, Field
 
 from copienator import configuration as config
+from copienator.gemini_config import thinking_config
 from copienator import (
     CliError,
     EvaluationWorkspace,
@@ -277,8 +278,7 @@ def generate_request(
             retry_options=types.HttpRetryOptions(attempts=1),
         ),
         automatic_function_calling=types.AutomaticFunctionCallingConfig(disable=True),
-        temperature=1.0,
-        top_p=0.95,
+        thinking_config=thinking_config(config.THINKING_LEVEL_FOR_LABEL),
         seed=seed,
         max_output_tokens=65535,
         response_mime_type= "application/json",

copienator/configuration.py

@@ -32,6 +32,10 @@ else:
     _configuration = _load_user_config(CONFIG_PATH)
 
 # Keep new optional settings compatible with older personal configuration files.
+THINKING_LEVEL_LITE = "low"
+THINKING_LEVEL_FOR_LABEL = "medium"
+THINKING_LEVEL_FLASH = "medium"
+THINKING_LEVEL_PRO = "high"
 ALWAYS_CROP = False
 SHOW_PERSONAL_STEPS = False
 RETURN_JPEG_ENABLED = True

copienator/gemini_config.py

@@ -0,0 +1,13 @@
+from google.genai import types
+
+
+def thinking_config(level: str | None) -> types.ThinkingConfig | None:
+    """Use a supported reasoning level, or omit it for the model default."""
+    if level is None:
+        return None
+    if level not in {"minimal", "low", "medium", "high"}:
+        raise ValueError(
+            f"Invalid Gemini thinking level {level!r}; "
+            'expected "minimal", "low", "medium", "high", or None'
+        )
+    return types.ThinkingConfig(thinking_level=level)

copienator/prompting.py

@@ -1,6 +1,8 @@
 from pathlib import Path
 import io
 from . import utils
+from . import configuration as settings
+from .gemini_config import thinking_config
 
 PERSPECTIVE_GUIDANCE = (
     "Ce barème est indicatif, si une réponse est entièrement correcte mais utilise "
@@ -173,8 +175,20 @@ UNROLLED_SCHEMA = {
 
 from google.genai import types
 
+def batch_generation_config(thinking_level):
+    """Build the REST generation config for a grading batch request."""
+    config = {
+        "maxOutputTokens": 65535,
+        "responseMimeType": "application/json",
+        "responseSchema": UNROLLED_SCHEMA,
+    }
+    thinking = thinking_config(thinking_level)
+    if thinking is not None:
+        config["thinkingConfig"] = thinking.model_dump(mode="json", by_alias=True, exclude_none=True)
+    return config
+
 # The root model for parsing is be: List[EvaluationEntry]
-def generate_request(input_dir, file, full_label):
+def generate_request(input_dir, file, full_label, *, thinking_level=settings.THINKING_LEVEL_PRO):
     """Generates request for Gemini."""
     prompt = make_prompt(input_dir, full_label)
     image_path = Path(file)
@@ -194,8 +208,7 @@ def generate_request(input_dir, file, full_label):
 
     generate_content_config = types.GenerateContentConfig(
         automatic_function_calling=types.AutomaticFunctionCallingConfig(disable=True),
-        temperature=1.0,
-        top_p=0.95,
+        thinking_config=thinking_config(thinking_level),
         seed=0,
         max_output_tokens=65535,
         response_mime_type= "application/json",
@@ -272,7 +285,7 @@ Commentaires d'origine :
 
     config = types.GenerateContentConfig(
         automatic_function_calling=types.AutomaticFunctionCallingConfig(disable=True),
-        temperature=1.0,
+        thinking_config=thinking_config(settings.THINKING_LEVEL_FLASH),
         response_mime_type="application/json",
         response_json_schema=TypeAdapter(List[FeedbackItem]).json_schema()
     )
@@ -303,7 +316,7 @@ Voici les labels possibles. Ta réponse doit être l'un d'entre eux :
             types.Part.from_text(text=prompt)])]
     config = types.GenerateContentConfig(
         automatic_function_calling=types.AutomaticFunctionCallingConfig(disable=True),
-        temperature=1.0,
+        thinking_config=thinking_config(settings.THINKING_LEVEL_FLASH),
     )
     return contents, config
 
@@ -334,7 +347,7 @@ d'entre eux :
     ])]
     config = types.GenerateContentConfig(
         automatic_function_calling=types.AutomaticFunctionCallingConfig(disable=True),
-        temperature=1.0,
+        thinking_config=thinking_config(settings.THINKING_LEVEL_FLASH),
         response_mime_type="application/json",
     )
     return contents, config