fix: improve import, evidence indexing and material-scoped grading

Preserve document revisions and evidence, harden indexing and export behavior, add explicit material selection and direct-folder import, validate model JSON schemas with safe error messages, and retain student selection while loading scoring data.

backend/wisegrade/llm.py

@@ -77,6 +77,26 @@ class ModelError(RuntimeError):
     pass
 
 
+def _safe_error_detail(exc):
+    """Describe failures without provider bodies, headers, URLs or exception text."""
+    if isinstance(exc, httpx.HTTPStatusError):
+        code = exc.response.status_code
+        hint = {401: "认证失败，请检查 API Key", 403: "访问被拒绝，请检查模型权限",
+                402: "额度不足，请检查账户余额", 429: "请求限流，请稍后重试或降低并发",
+                400: "请求参数不被模型支持", 404: "端点或模型不存在"}.get(
+                    code, "模型服务异常，请稍后重试" if code >= 500 else "请求被拒绝，请检查模型设置")
+        return f"HTTP {code}：{hint}"
+    if isinstance(exc, httpx.ConnectError):
+        return "ConnectError：无法连接模型端点，请检查 Docker 网络、DNS 或代理"
+    if isinstance(exc, httpx.TimeoutException):
+        return "模型请求超时，请稍后重试或调整超时设置"
+    if isinstance(exc, ValidationError):
+        fields = sorted({str(error["loc"][0]) for error in exc.errors(include_input=False)
+                         if error["loc"] and str(error["loc"][0]).isidentifier()})
+        return "ValidationError：模型返回字段不符合格式" + (f"（{', '.join(fields)}）" if fields else "")
+    return type(exc).__name__
+
+
 class DeepSeekClient:
     def __init__(self, *, base_url=None, model=None, api_key=None, timeout=None, retries=None, transport=None):
         settings = get_settings()
@@ -96,8 +116,9 @@ class DeepSeekClient:
             if not path.is_file():
                 raise ModelError(f"页面图像不存在: {path.name}")
             content.append({"type": "image_url", "image_url": {"url": "data:image/png;base64," + base64.b64encode(path.read_bytes()).decode()}})
+        output_schema = json.dumps(schema.model_json_schema(), ensure_ascii=False)
         body = {"model": self.model, "response_format": {"type": "json_object"},
-                "messages": [{"role": "system", "content": prompt + " 严格输出一个JSON对象。"}, {"role": "user", "content": content}],
+                "messages": [{"role": "system", "content": prompt + " 严格输出一个JSON对象，字段与类型遵循以下JSON Schema：" + output_schema}, {"role": "user", "content": content}],
                 "temperature": 0}
         last_error = None
         for attempt in range(self.retries + 1):
@@ -110,7 +131,7 @@ class DeepSeekClient:
                 return schema.model_validate(json.loads(raw))
             except (httpx.HTTPError, ValueError, KeyError, IndexError, ValidationError) as exc:
                 # Never include response bodies or headers: providers sometimes echo credentials.
-                last_error = type(exc).__name__
+                last_error = _safe_error_detail(exc)
                 if attempt < self.retries:
                     time.sleep(min(2 ** attempt, 5))
         raise ModelError(f"模型请求或JSON校验失败（{last_error}），已重试{self.retries}次")
@@ -197,6 +218,7 @@ def validate_score(result: ScoreResult, criterion: dict, chunks: list[dict], *,
     by_target = {target["target_id"]: target for target in targets}
     default_content = next((target for target in targets if target["use"] == USE_CONTENT), None)
     by_id = {chunk["chunk_id"]: chunk for chunk in chunks}
+    has_content_evidence = False
 
     for evidence in result.evidence:
         chunk = by_id.get(evidence.chunk_id)
@@ -209,16 +231,30 @@ def validate_score(result: ScoreResult, criterion: dict, chunks: list[dict], *,
         if quote is None:
             raise ModelError("模型证据未能匹配检索原文")
         page_type = chunk.get("page_type") or "unknown"
-        target = by_target.get(evidence.target_id) or default_content or (targets[0] if targets else None)
+        if evidence.target_id and evidence.target_id not in by_target:
+            raise ModelError("模型证据包含未知 target_id")
+        target = by_target.get(evidence.target_id)
+        if target is None and not evidence.target_id:
+            # Empty identifiers are accepted only for pre-object legacy rubrics.
+            if any(isinstance(value, dict) for value in criterion.get("required_evidence", [])):
+                raise ModelError("模型证据缺少 target_id")
+            target = default_content or (targets[0] if targets else None)
         if target is None:
             raise ModelError("模型证据未关联任何必备证据目标")
+        if chunk.get("target_ids") and target["target_id"] not in chunk["target_ids"]:
+            raise ModelError("模型证据未关联该片段的检索目标")
         if not page_type_allowed(page_type, target["allowed_page_types"]):
             raise ModelError(f"模型证据来自不允许的页面类型（{page_type}），不能支持「{target['label']}」")
         if target["use"] == USE_CONTENT and not content_evidence_allowed(page_type):
             raise ModelError(f"目录、摘要或参考文献页不能作为「{target['label']}」的内容证据")
         if target["use"] == USE_CONTENT and heading_only_text(quote):
             raise ModelError(f"「{target['label']}」的证据只有章节标题，没有实质内容")
+        if any(char in quote for char in ("\ufffd", "\ufffe", "\uffff")):
+            raise ModelError("模型证据含未解析字符，请核对原页")
+        has_content_evidence |= target["use"] == USE_CONTENT
         evidence.chunk_id, evidence.page, evidence.quote = chunk["chunk_id"], chunk["page"], quote
         evidence.target_id = target["target_id"]
+    if result.score is not None and result.score > 0 and require_content_evidence and not has_content_evidence:
+        raise ModelError("正分内容判断必须包含合法内容证据，不能仅引用结构信息")
     return result