jhzhang/rag
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

fix(exam): 修复出题批卷 API 契约一致性问题

Browse Source 
P0:
- 修复 V2 降级路径响应格式，fallback 输出经过 enrich + dedup + balance
- 补充主观题 format example，要求 LLM 生成 scoring_points 评分标准

P1:
- 补充多选题、判断题 format example，明确 answer 为列表格式
- 统一批卷响应格式，所有题型增加 grading_status 字段
- 出题数量不足时返回 requested_types/actual_types/warnings
- 修复 _call_llm 未配置时抛异常而非静默返回 0 分

P2:
- API 层增加 question_types 和 difficulty 入参校验
- 修复 generate_questions_from_content 方法名和参数类型 bug
This commit is contained in:
lacerate551
2026-06-05 12:06:49 +08:00
parent 42434781f7
commit b6631c626b
4 changed files with 957 additions and 793 deletions
Download Patch File Download Diff File Expand all files Collapse all files

70
exam_pkg/api.py
View File

@@ -32,6 +32,35 @@ from auth.gateway import (
from core.status_codes import EXAM_SUCCESS, GRADE_SUCCESS, EXAM_ERROR, GRADE_ERROR, BAD_REQUEST, NO_CONTENT, LLM_ERROR
from api.response_utils import success_response, error_response
# 合法题型
VALID_QUESTION_TYPES = {'single_choice', 'multiple_choice', 'true_false', 'fill_blank', 'subjective'}
def validate_question_types(question_types: dict) -> str:
"""
校验 question_types 参数
Returns:
None 表示校验通过，字符串表示错误信息
"""
if not isinstance(question_types, dict):
return "question_types 必须为对象"
for q_type, count in question_types.items():
if q_type not in VALID_QUESTION_TYPES:
return f"不支持的题型: {q_type}，合法值: {', '.join(sorted(VALID_QUESTION_TYPES))}"
if not isinstance(count, int) or count < 0:
return f"题型 {q_type} 的数量必须为非负整数，当前值: {count}"
return None
def validate_difficulty(difficulty) -> str:
"""校验 difficulty 参数"""
if not isinstance(difficulty, int) or difficulty < 1 or difficulty > 5:
return "difficulty 必须为 1-5 的整数"
return None
# 创建蓝图
exam_bp = Blueprint('exam', __name__)
@@ -69,12 +98,16 @@ def api_generate_questions():
"request_id": "uuid",
"questions": [
{
"metadata": {"question_id": "...", "question_type": "...", ...},
"source_trace": {"document_name": "...", "sources": [...], ...},
"content": {"stem": "...", "data": {...}, "answer": "...", ...}
"question_type": "single_choice",
"difficulty": 3,
"content": {"stem": "...", "data": {...}, "answer": "...", "explanation": "..."},
"source_trace": {"document_name": "...", "chunk_ids": [...], "sources": [...]}
}
],
"total": 10,
"requested_types": {"single_choice": 3, "fill_blank": 2},
"actual_types": {"single_choice": 3, "fill_blank": 2},
"warnings": ["fill_blank: 请求 3 道，实际生成 2 道"],
"source_chunks_used": 15
}
"""
@@ -91,6 +124,17 @@ def api_generate_questions():
if not question_types:
return error_response("MISSING_PARAMS", BAD_REQUEST, "缺少 question_types 参数", http_status=400)
# 校验题型和数量
type_error = validate_question_types(question_types)
if type_error:
return error_response("INVALID_PARAMS", BAD_REQUEST, type_error, http_status=400)
# 校验难度
difficulty = data.get('difficulty', 3)
diff_error = validate_difficulty(difficulty)
if diff_error:
return error_response("INVALID_PARAMS", BAD_REQUEST, diff_error, http_status=400)
# 获取当前用户
user = get_current_user()
if not user:
@@ -149,6 +193,12 @@ def api_generate_smart():
if not file_path or not collection:
return error_response("MISSING_PARAMS", BAD_REQUEST, "缺少 file_path 或 collection 参数", http_status=400)
# 校验难度
difficulty = data.get('difficulty', 3)
diff_error = validate_difficulty(difficulty)
if diff_error:
return error_response("INVALID_PARAMS", BAD_REQUEST, diff_error, http_status=400)
# 获取当前用户
user = get_current_user()
if not user:
@@ -237,11 +287,23 @@ def api_grade_answers():
{
"success": true,
"request_id": "uuid",
"results": [...],
"results": [
{
"question_id": "uuid",
"score": 2,
"max_score": 2,
"grading_status": "success",
"details": {"correct": true, "student_answer": "A", "correct_answer": "B", "feedback": "..."}
}
],
"total_score": 10.5,
"total_max_score": 16.0,
"score_rate": 65.6
}
grading_status 取值：
- "success": 评分成功
- "failed": 评分失败（LLM 超时/解析失败等），score 为 0
"""
try:
data = request.json

106
exam_pkg/generator.py
View File

@@ -322,13 +322,17 @@ class QuestionGenerator:
) -> List[Dict]:
"""
降级出题方法：当知识点提取失败时，直接使用 chunks 出题
输出格式与正常路径一致（嵌套结构），确保 API 响应格式统一
"""
source_content = build_source_context(chunks)
total_questions = sum(question_types.values())
if not source_content or total_questions == 0:
return []
# 构造出题 prompt
q_type_str = ", ".join(f"{t}:{n}道" for t, n in question_types.items())
# 构造出题 prompt —— 使用与正常路径相同的嵌套格式
prompt = f"""请根据以下参考资料生成题目。
## 参考资料
@@ -337,41 +341,41 @@ class QuestionGenerator:
## 要求
- 题型及数量：{json.dumps(question_types, ensure_ascii=False)}
- 题型及数量：{q_type_str}
- 难度等级：{difficulty}/5
- 每道题必须包含：题干、答案、解析
- 答案必须基于参考资料
## 输出格式
## 输出格式（JSON 数组）
{self._get_format_examples()}
返回 JSON 数组，每个元素格式：
{{"stem": "题干", "type": "single_choice/multiple_choice/true_false/fill_blank/subjective", "answer": "答案", "explanation": "解析", "options": [{{"key": "A", "content": "选项内容"}}]}}
只输出 JSON 数组，不要其他内容。"""
请直接输出 JSON 数组："""
try:
# 使用 llm_utils 统一调用
messages = [{"role": "user", "content": prompt}]
messages = [
{"role": "system", "content": "你是一个专业的出题专家，必须严格按照JSON格式输出。"},
{"role": "user", "content": prompt}
]
content = call_llm(
client=self.client,
prompt=prompt,
model=self.model,
temperature=0.7,
max_tokens=2000,
max_tokens=4000,
messages=messages
)
if not content:
return []
# 解析 JSON
questions = parse_json_list_from_response(content)
if questions:
# 添加元数据
for q in questions:
q["source_trace"] = {
"document_name": document_name,
"chunks_count": len(chunks)
}
return questions
raw_questions = parse_json_list_from_response(content)
if not raw_questions:
return []
# Schema 校验
validated = validate_questions_schema(raw_questions)
# 补充溯源信息（与正常路径格式统一）
enriched = self._enrich_with_source_trace(validated, chunks, document_name)
return enriched
except Exception as e:
logger.error(f"[降级出题失败] {e}")
@@ -510,7 +514,7 @@ class QuestionGenerator:
return result
def _get_format_examples(self) -> str:
"""返回各题型格式示例"""
"""返回各题型格式示例（覆盖全部 5 种题型）"""
return '''
### 单选题示例
{
@@ -524,6 +528,30 @@ class QuestionGenerator:
"referenced_chunk_ids": ["chunk_001"]
}
### 多选题示例
{
"type": "multiple_choice",
"content": {
"stem": "题干内容",
"data": {"options": [{"key": "A", "content": "选项A"}, {"key": "B", "content": "选项B"}, {"key": "C", "content": "选项C"}, {"key": "D", "content": "选项D"}]},
"answer": ["A", "C"],
"explanation": "解析..."
},
"referenced_chunk_ids": ["chunk_001"]
}
### 判断题示例
{
"type": "true_false",
"content": {
"stem": "判断：某个陈述内容",
"data": {},
"answer": "对",
"explanation": "解析..."
},
"referenced_chunk_ids": ["chunk_002"]
}
### 填空题示例
{
"type": "fill_blank",
@@ -535,6 +563,24 @@ class QuestionGenerator:
},
"referenced_chunk_ids": ["chunk_003"]
}
### 主观题示例
{
"type": "subjective",
"content": {
"stem": "请简述某个概念的含义及其应用场景。",
"data": {
"scoring_points": [
{"point": "要点1：概念定义", "weight": 0.4},
{"point": "要点2：应用场景", "weight": 0.3},
{"point": "要点3：举例说明", "weight": 0.3}
]
},
"answer": "参考答案范文...",
"explanation": "解析..."
},
"referenced_chunk_ids": ["chunk_004"]
}
'''
def _extract_knowledge_points(
@@ -862,9 +908,14 @@ def generate_questions_from_content(
question_types: Dict[str, int],
difficulty: int = 3
) -> List[Dict]:
"""便捷函数：从内容生成题目"""
"""便捷函数：从文本内容生成题目（将纯文本包装为单 chunk 后调用结构化出题）"""
if not source_content or not source_content.strip():
return []
# 将纯文本包装为一个 chunk
chunks = [{"content": source_content, "section": "", "page": None, "chunk_id": "content_0"}]
generator = QuestionGenerator()
return generator.generate_questions(source_content, document_name, question_types, difficulty)
return generator.generate_questions_structured(chunks, document_name, question_types, difficulty)
def analyze_document_for_exam(chunks: List[Dict]) -> Dict[str, Any]:
@@ -1114,9 +1165,14 @@ def generate_questions_structured_v2(
logger.info(f" 全局唯一知识点: {len(all_knowledge_points)}")
if not all_knowledge_points:
# 降级：直接使用 chunks 出题
# 降级：直接使用 chunks 出题（格式已由 fallback 内部统一）
logger.warning("[v2] 知识点提取失败，降级使用 chunks 出题")
return generator._generate_questions_fallback(chunks, document_name, question_types, difficulty)
fallback_questions = generator._generate_questions_fallback(chunks, document_name, question_types, difficulty)
# 降级路径也执行去重和数量平衡
deduped = _deduplicate_questions(fallback_questions)
final = _balance_question_types(deduped, question_types)
logger.info(f" [降级] 最终题目数: {len(final)}")
return final
# AI 分配：哪些知识点出什么题型
assignments = _ai_assign_question_types(all_knowledge_points, question_types)

69
exam_pkg/grader.py
View File

@@ -97,8 +97,13 @@ def grade_objective(answer: Dict) -> Dict:
"question_id": answer.get('question_id'),
"score": max_score if correct else 0,
"max_score": max_score,
"correct": correct,
"feedback": f"正确答案: {correct_answer}" if not correct else "正确！"
"grading_status": "success",
"details": {
"correct": correct,
"student_answer": student_answer,
"correct_answer": correct_answer,
"feedback": f"正确答案: {correct_answer}" if not correct else "正确！"
}
}
@@ -119,7 +124,8 @@ def grade_fill_blank(answer: Dict) -> Dict:
"question_id": answer.get('question_id'),
"score": 0,
"max_score": max_score,
"details": {"error": "答案格式错误"}
"grading_status": "failed",
"details": {"error": "答案格式错误，correct_answers 或 student_answer 为空"}
}
# 计算每空分数
@@ -149,9 +155,11 @@ def grade_fill_blank(answer: Dict) -> Dict:
"question_id": answer.get('question_id'),
"score": round(total_score, 1),
"max_score": max_score,
"grading_status": "success",
"details": {
"blank_scores": blank_scores,
"correct_answers": correct_answers
"total_blanks": len(correct_answers),
"correct_blanks": sum(1 for s in blank_scores if s > 0)
}
}
@@ -243,22 +251,38 @@ class AnswerGrader:
}
# 收集结果
for future in as_completed(future_to_qid, timeout=60):
qid = future_to_qid[future]
try:
result = future.result(timeout=15)
results_map[qid] = result
except Exception as e:
# 失败时返回默认结果
results_map[qid] = {
"question_id": qid,
"score": 0,
"max_score": next(
(a.get('max_score', 10) for a in questions if a.get('question_id') == qid),
10
),
"error": str(e)
}
try:
for future in as_completed(future_to_qid, timeout=60):
qid = future_to_qid[future]
try:
result = future.result(timeout=15)
results_map[qid] = result
except Exception as e:
# 失败时返回默认结果
results_map[qid] = {
"question_id": qid,
"score": 0,
"max_score": next(
(a.get('max_score', 10) for a in questions if a.get('question_id') == qid),
10
),
"grading_status": "failed",
"details": {"error": str(e)}
}
except TimeoutError:
# 超时：为未完成的任务设置失败状态
for future, qid in future_to_qid.items():
if qid not in results_map:
results_map[qid] = {
"question_id": qid,
"score": 0,
"max_score": next(
(a.get('max_score', 10) for a in questions if a.get('question_id') == qid),
10
),
"grading_status": "failed",
"details": {"error": "批阅超时"}
}
@retry(times=2, delay=1)
def _grade_subjective(self, answer: Dict) -> Dict:
@@ -315,8 +339,7 @@ class AnswerGrader:
def _call_llm(self, prompt: str) -> str:
"""调用本地 LLM"""
if not self.client:
# 无 LLM 客户端，返回默认评分
return json.dumps({"score": 0, "overall_feedback": "LLM 未配置"})
raise ValueError("LLM 未配置，无法进行主观题评分")
messages = [
{"role": "system", "content": "你是一个专业的阅卷老师，请严格按照JSON格式输出评分结果。"},
@@ -347,6 +370,7 @@ class AnswerGrader:
"question_id": answer.get('question_id'),
"score": score,
"max_score": max_score,
"grading_status": "success",
"details": {
"scoring_breakdown": result.get('scoring_breakdown', []),
"highlights": result.get('highlights', []),
@@ -360,6 +384,7 @@ class AnswerGrader:
"question_id": answer.get('question_id'),
"score": 0,
"max_score": max_score,
"grading_status": "failed",
"details": {"error": "评分结果解析失败"}
}

1505
exam_pkg/manager.py
View File

File diff suppressed because it is too large Load Diff