fix(exam): 修复出题批卷 API 契约一致性问题
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
@@ -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)
|
||||
|
||||
Reference in New Issue
Block a user