sebastien/Copies
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Gemini_for_enonce : améliorations

Browse Source
This commit is contained in:
sebastien
2026-08-08 12:30:14 +02:00
parent 9d7a4f37b6
commit 137a9a7868
5 changed files with 262 additions and 135 deletions
Download Patch File Download Diff File Expand all files Collapse all files
gemini_for_enonce.py
+172 -75
View File
@@ -16,7 +16,12 @@ def get_lcp(s1: str, s2: str) -> str:
i = 0
while i < len(s1) and i < len(s2) and s1[i] == s2[i]:
i += 1
return s1[:i]
lcp = s1[:i]
if ')' in s1 or ')' in s2:
last_paren = lcp.rfind(')')
if last_paren != -1:
return lcp[:last_paren + 1]
return lcp
import config
@@ -114,7 +119,18 @@ PROMPT_3 = """I am providing:
2. The source code of the exam questions (`enonce` file).
Your task:
Extract important information necessary to understand the questions (e.g., definitions of objects, global notations, hypotheses, context) that are NOT part of the question texts themselves.
Extract important information necessary to understand the questions (e.g., definitions of objects, global notations, hypotheses, context) that are NOT part of the question texts themselves. Often, this information can be in a previous \\item that is not itself a question, but contains the question items.
For example, given LaTeX code like
\\item Let N, M be two commutating matrices
\\begin{itemize}
\\item Prove that N, M have a common eigenvector
\\item Prove that N, M are co-trigonalizable.
\\end{itemize}
the `Let N, M be two commutating matrices` part is not a question itself, and is important information to understand the next two questions.
For each extracted piece of information, identify:
1. The label of the FIRST question that comes immediately AFTER this information in the exam.
2. The label of the LAST question that uses or relies on this information.
@@ -128,9 +144,16 @@ def find_file(folder: Path, base_name: str) -> Path:
return path
return None
def process_exam(folder_path: str):
def process_exam(folder_path: str, restart: bool = False):
folder = Path(folder_path)
cache_dir = folder / "Cache"
cache_dir.mkdir(exist_ok=True)
cache_q_file = cache_dir / "gemini_questions.json"
cache_s_file = cache_dir / "gemini_solutions.json"
cache_c_file = cache_dir / "gemini_context.json"
# 1. Resolve files
pdf_path = folder / "enonce.pdf"
enonce_path = find_file(folder, "enonce")
@@ -172,10 +195,8 @@ def process_exam(folder_path: str):
response_json_schema=ExamQuestions.model_json_schema(),
)
cache_q_file = folder / "gemini_questions.json"
if cache_q_file.is_file():
print("Loading cached questions from gemini_questions.json...")
if cache_q_file.is_file() and not restart:
print("Loading cached questions from Cache/gemini_questions.json...")
response_q_text = cache_q_file.read_text(encoding="utf-8")
else:
print("Sending request 1 (Questions) to Gemini...")
@@ -212,10 +233,8 @@ def process_exam(folder_path: str):
response_json_schema=ExamSolutions.model_json_schema(),
)
cache_s_file = folder / "gemini_solutions.json"
if cache_s_file.is_file():
print("Loading cached solutions from gemini_solutions.json...")
if cache_s_file.is_file() and not restart:
print("Loading cached solutions from Cache/gemini_solutions.json...")
response_s_text = cache_s_file.read_text(encoding="utf-8")
else:
print("Sending request 2 (Solutions) to Gemini...")
@@ -250,10 +269,8 @@ def process_exam(folder_path: str):
response_json_schema=ExamContext.model_json_schema(),
)
cache_c_file = folder / "gemini_context.json"
if cache_c_file.is_file():
print("Loading cached context from gemini_context.json...")
if cache_c_file.is_file() and not restart:
print("Loading cached context from Cache/gemini_context.json...")
response_c_text = cache_c_file.read_text(encoding="utf-8")
else:
print("Sending request 3 (Context) to Gemini...")
@@ -329,51 +346,70 @@ def process_exam(folder_path: str):
trunc_map = {}
# --- INITIAL GROUPING COMPUTATION ---
# 1. Normalize labels first
for item in extracted_data.items:
if isinstance(item, QuestionItem):
item.label = item.label.replace("Exercice", "Ex").replace(".", ")")
# 2. Extract questions and compute grouping indices
questions_only = [item for item in extracted_data.items if isinstance(item, QuestionItem)]
# questions_only = [item for item in extracted_data.items if isinstance(item, QuestionItem)]
q_group_indices = []
if questions_only:
current_g = [0]
for i in range(1, len(questions_only)):
p = get_lcp(questions_only[current_g[0]].label, questions_only[i].label)
proposed = current_g + [i]
valid = True
for k in range(len(proposed) - 1):
if get_lcp(questions_only[proposed[k]].label, questions_only[proposed[k+1]].label) != p:
valid = False
break
if valid:
current_g.append(i)
else:
q_group_indices.append(current_g)
current_g = [i]
q_group_indices.append(current_g)
n = len(questions_only)
if n == 1:
q_group_indices = [[0]]
else:
adj_lcp = [get_lcp(questions_only[i].label, questions_only[i+1].label) for i in range(n - 1)]
group_starter_labels = {questions_only[g[0]].label for g in q_group_indices[1:]} if q_group_indices else set()
current_g = [0]
for i in range(n - 1):
p = adj_lcp[i]
prev_p = adj_lcp[i - 1] if i > 0 else ""
next_p = adj_lcp[i + 1] if i < n - 2 else ""
# Group i and i+1 together if p is non-empty and at least as specific as adjacent LCPs
if p and len(p) >= len(prev_p) and len(p) >= len(next_p):
current_g.append(i + 1)
else:
q_group_indices.append(current_g)
current_g = [i + 1]
q_group_indices.append(current_g)
# Build list of unique ContextItems from extracted data
all_contexts = [item for item in extracted_data.items if isinstance(item, ContextItem)]
initial_groups = []
current_group = []
for g_indices in q_group_indices:
group_items = []
first_q_idx = g_indices[0]
last_q_idx = g_indices[-1]
for i, item in enumerate(extracted_data.items):
is_new_group = False
if isinstance(item, QuestionItem):
if item.label in group_starter_labels:
if not (len(current_group) > 0 and isinstance(current_group[-1], ContextItem)):
is_new_group = True
elif isinstance(item, ContextItem):
if i + 1 < len(extracted_data.items):
next_item = extracted_data.items[i+1]
if isinstance(next_item, QuestionItem) and next_item.label in group_starter_labels:
is_new_group = True
for q_idx in g_indices:
q_item = questions_only[q_idx]
if is_new_group and current_group:
initial_groups.append(current_group)
current_group = []
# 1. Collect contexts targeting this specific question
# 2. Or contexts carried over from an earlier group (only added at the start of the group)
for ctx in all_contexts:
target_idx = label_to_idx.get(ctx.target_question_label, -1)
last_idx = label_to_idx.get(ctx.last_question_label, -1)
current_group.append(item)
if target_idx != -1 and last_idx != -1:
is_exact_target = (target_idx == q_idx)
is_carried_over = (q_idx == first_q_idx and target_idx < first_q_idx and last_idx >= first_q_idx)
if current_group:
initial_groups.append(current_group)
if is_exact_target or is_carried_over:
group_items.append(ContextItem(
target_question_label=ctx.target_question_label,
last_question_label=ctx.last_question_label,
content=ctx.content
))
group_items.append(q_item)
initial_groups.append(group_items)
# ---- Transform labels, and check uniqueness
@@ -383,9 +419,6 @@ def process_exam(folder_path: str):
for item in group:
if isinstance(item, QuestionItem):
orig_label = item.label
# 1. Transform label
item.label = item.label.replace("Exercice", "Ex")
item.label = item.label.replace(".", ")")
# 2. Ensure uniqueness (prefix with XX)
while item.label in seen_labels:
@@ -469,11 +502,28 @@ def process_exam(folder_path: str):
input("Press ENTER to return to the editor...")
continue
# Map original contexts by normalized content
orig_contexts = {c.context_content.strip(): c for c in context_data.contexts}
# 2. Actual Parsing
grouped_items = []
current_raw_group = [] # Stores (is_context, label_or_flag, content)
all_new_q_labels = []
# Pass 1: Read all edited lines and collect question labels in sequence
for line in edited_lines:
if line == "---" or " ### " not in line:
continue
lbl, content_raw = line.split(" ### ", 1)
lbl = lbl.strip()
if lbl != "CONTEXT":
all_new_q_labels.append(lbl)
# Mapping from original question index to new label
idx_to_new_label = {i: all_new_q_labels[i] for i in range(min(len(questions_only), len(all_new_q_labels)))}
orig_q_idx = 0
current_group = []
labels_list = []
orig_idx = 0
for line in edited_lines:
if line == "---":
@@ -488,37 +538,82 @@ def process_exam(folder_path: str):
new_label, edited_content_raw = line.split(" ### ", 1)
new_label = new_label.strip()
if new_label != "CONTEXT":
labels_list.append(new_label)
if "…" in edited_content_raw and edited_content_raw in trunc_map:
edited_content_raw = trunc_map[edited_content_raw]
edited_content = edited_content_raw.replace(' \\n ', '\n')
if orig_idx < len(extracted_data.items):
orig_item = extracted_data.items[orig_idx] # <-- Add this line
if isinstance(orig_item, QuestionItem):
current_group.append(QuestionItem(
label=new_label,
question_content=edited_content,
solution_content=orig_item.solution_content
))
elif isinstance(orig_item, ContextItem):
current_group.append(ContextItem(
target_question_label=orig_item.target_question_label,
last_question_label=orig_item.last_question_label,
content=edited_content
))
orig_idx += 1
if new_label == "CONTEXT":
current_group.append(('CONTEXT', edited_content))
else:
sol_content = questions_only[orig_q_idx].solution_content if orig_q_idx < len(questions_only) else ""
current_group.append(QuestionItem(
label=new_label,
question_content=edited_content,
solution_content=sol_content
))
orig_q_idx += 1
if current_group:
grouped_items.append(current_group)
# If we reached here without 'continue', the data is valid
# Pass 2: Resolve ContextItem target/last labels per group
final_grouped_items = []
for group in grouped_items:
final_group = []
q_in_group = [item for item in group if isinstance(item, QuestionItem)]
g_first_label = q_in_group[0].label if q_in_group else ""
g_last_label = q_in_group[-1].label if q_in_group else ""
for i, item in enumerate(group):
if isinstance(item, tuple) and item[0] == 'CONTEXT':
c_text = item[1]
norm_text = c_text.strip()
# Find next question label in group following this context
next_q_label = g_first_label
for successor in group[i+1:]:
if isinstance(successor, QuestionItem):
next_q_label = successor.label
break
if norm_text in orig_contexts:
orig_c = orig_contexts[norm_text]
orig_target_idx = label_to_idx.get(orig_c.target_question_label, -1)
orig_last_idx = label_to_idx.get(orig_c.last_question_label, -1)
mapped_target = idx_to_new_label.get(orig_target_idx, next_q_label)
mapped_last = idx_to_new_label.get(orig_last_idx, g_last_label)
# Check if context's last question is BEFORE the first question of this group
first_q_idx_in_exam = all_new_q_labels.index(g_first_label) if g_first_label in all_new_q_labels else -1
last_q_idx_in_exam = all_new_q_labels.index(mapped_last) if mapped_last in all_new_q_labels else -1
if last_q_idx_in_exam != -1 and first_q_idx_in_exam != -1 and last_q_idx_in_exam < first_q_idx_in_exam:
mapped_last = g_last_label
final_group.append(ContextItem(
target_question_label=mapped_target,
last_question_label=mapped_last,
content=c_text
))
else:
# New context created by user
final_group.append(ContextItem(
target_question_label=next_q_label,
last_question_label=g_last_label,
content=c_text
))
else:
final_group.append(item)
final_grouped_items.append(final_group)
grouped_items = final_grouped_items
break
labels_list = [item.label for group in grouped_items for item in group if isinstance(item, QuestionItem)]
# Save labels and proceed
with open(folder / "labels", 'w', encoding='utf-8') as f_labels:
for label in labels_list:
@@ -697,6 +792,8 @@ if __name__ == "__main__":
parser = argparse.ArgumentParser(description="Extract exam and solution code via Gemini.")
parser.add_argument("folder", help="Directory containing the exam files")
parser.add_argument("--restart", action="store_true", help="Ignore cache files and re-run extraction requests.")
args = parser.parse_args()
process_exam(args.folder)
process_exam(args.folder, restart=args.restart)