fix：处理模型错误添加大量文本的bug

2026-10-14 01:08:50 -07:00
parent a80727b7a0
commit 41b769adde
35 changed files with 388 additions and 149 deletions
@@ -1,7 +1,10 @@
 """按物料展开字段定义：解析共享分组引用、取字段默认值、派生前端属性面板 setters"""

+import re
+
 from pydantic import JsonValue

+from app.schemas.base import JsonObject
 from app.schemas.component_resource import (
    ComponentCatalog,
    ComponentDefinition,
@@ -61,13 +64,35 @@ def panel_setters(groups: list[FieldGroup]) -> list[dict]:


 def frontend_catalog(catalog: ComponentCatalog) -> dict:
-    """前端物料接口的数据：在保存的定义上附加由字段定义派生的 setters"""
+    """前端物料接口的数据：在保存的定义上附加由字段定义派生的 setters 和 defaults"""
    data = catalog.model_dump(mode="json", by_alias=True)
    for item, component in zip(data["components"], catalog.components, strict=True):
        item["setters"] = panel_setters(resolve_field_groups(catalog, component))
+        item["defaults"] = structural_defaults(catalog, component)
    return data


+def structural_defaults(
+    catalog: ComponentCatalog, component: ComponentDefinition
+) -> JsonObject:
+    """默认配置去掉示例内容（字段定义里标了 example 的标题、数据……）后剩下的结构默认值
+
+    前端打开大屏时用它给旧节点补上物料后来新增的字段；不用完整的默认配置，否则用户删掉的标题
+    会被补回成示例标题，真实大屏上又出现演示内容
+    """
+    defaults = component.template()
+    for field in component_fields(catalog, component):
+        if field.example:
+            *parents, last = field.segments
+            try:
+                container = lookup(defaults, parents)
+            except ValueError:
+                continue
+            if isinstance(container, dict) and isinstance(last, str):
+                container.pop(last, None)
+    return defaults
+
+
 # ---- 字段查询：MCP 工具、动作提示词和动作校验共用同一套逻辑 ----


@@ -128,7 +153,15 @@ def describe_field(
    """给模型看的字段说明：省略空值和属性面板专用的信息"""
    data = field.model_dump(
        mode="json",
-        exclude={"default", "panel", "span", "editable", "themed"},
+        exclude={
+            "default",
+            "panel",
+            "span",
+            "editable",
+            "themed",
+            "example",
+            "identifies",
+        },
        exclude_none=True,
        exclude_defaults=True,
    )
@@ -136,6 +169,9 @@ def describe_field(
    if field.themed:
        # 默认为 null：没单独设置时画面上按画布主题取色
        data["follows_theme"] = True
+    if field.example:
+        # 默认值是演示内容，要按实际内容写，不能当成合理的默认
+        data["default_is_example"] = True
    if not field.editable:
        data["editable"] = False
    data["group"] = group.key
@@ -257,3 +293,139 @@ def check_update_field(
        raise ValueError(f"字段 {path}（{field.label}）不允许修改")
    if error := field.value_error(action.value):
        raise ValueError(f"字段 {path}（{field.label}）{error}")
+
+
+def component_fields(
+    catalog: ComponentCatalog, component: ComponentDefinition
+) -> list[MaterialField]:
+    return [
+        field
+        for group in resolve_field_groups(catalog, component)
+        for field in group.fields
+    ]
+
+
+def check_add_fields(
+    catalog: ComponentCatalog, node_type: str, written: JsonValue
+) -> None:
+    """新增组件时模型显式写的字段，和 update 用同一套字段定义检查；不合法时一次列出全部问题
+
+    没写的字段来自默认配置，载入目录时已经检查过；物料没有定义的路径不检查，新增时可以整段写 option
+    """
+    problems = []
+    for field in component_fields(catalog, find_material(catalog, node_type)):
+        try:
+            value = lookup(written, field.segments)
+        except ValueError:
+            problems.append(
+                f"字段 {field.path}（{field.label}）所在的上一层结构和物料不同，"
+                "按 get_component_materials 返回的默认配置写"
+            )
+            continue
+        if value is not MISSING and (error := field.value_error(value)):
+            problems.append(f"字段 {field.path}（{field.label}）{error}")
+    if problems:
+        raise ValueError("；".join(problems))
+
+
+def retained_examples(
+    catalog: ComponentCatalog | None, node: dict
+) -> list[tuple[MaterialField, JsonValue]]:
+    """组件上还和默认配置一样的示例内容（标题、系列名、数据……），按字段定义的 example 判断
+
+    只比较现在的值，不管是新增时没写、写了一样的值，还是之后又改回来；物料不存在时为空
+    """
+    if catalog is None:
+        return []
+    try:
+        component = find_material(catalog, node["type"])
+    except ValueError:
+        return []
+    template = component.template()
+    retained = []
+    for field in component_fields(catalog, component):
+        if not field.example:
+            continue
+        try:
+            current = lookup(node, field.segments)
+        except ValueError:
+            continue
+        default = lookup(template, field.segments)
+        if current is not MISSING and current == default:
+            retained.append((field, default))
+    return retained
+
+
+# 摘要里每段文字的长度上限：文本组件的内容可能很长，认出组件只需要开头
+IDENTITY_MAX_CHARS = 40
+
+
+def identifying_texts(catalog: ComponentCatalog | None, node: dict) -> list[str]:
+    """组件在画面上的名字或内容：字段定义里标了 identifies 的字段的文字，过长时截断
+
+    路径途经数组而字段按对象写时（例如 ECharts 的多个 title），每一项都取
+    """
+    if catalog is None:
+        return []
+    try:
+        component = find_material(catalog, node["type"])
+    except ValueError:
+        return []
+    texts = []
+    for field in component_fields(catalog, component):
+        if field.identifies:
+            texts.extend(
+                text
+                if len(text) <= IDENTITY_MAX_CHARS
+                else text[:IDENTITY_MAX_CHARS] + "…"
+                for value in _values_at(node, field.segments)
+                if isinstance(value, str) and (text := value.strip())
+            )
+    return texts
+
+
+# 识别字段（标题、名称、文本内容）在画面上是一行短文字。换行、超过这个长度或带句号分号时是一段话：
+# 线上问答之后"把结果放到大屏上"，模型把"关键发现 1. …"、"注：……；……。"写进了文本组件
+LABEL_MAX_CHARS = 60
+SENTENCE_MARKS = re.compile(r"[。；;]")
+
+
+def is_label(text: str) -> bool:
+    text = text.strip()
+    return (
+        "\n" not in text
+        and len(text) <= LABEL_MAX_CHARS
+        and not SENTENCE_MARKS.search(text)
+    )
+
+
+def paragraph_texts(
+    catalog: ComponentCatalog | None, node: dict, paths: set[str] | None = None
+) -> list[tuple[MaterialField, str]]:
+    """组件的识别字段里写成一段话的文字；paths 给出时只看这些字段"""
+    if catalog is None:
+        return []
+    try:
+        component = find_material(catalog, node["type"])
+    except ValueError:
+        return []
+    return [
+        (field, value)
+        for field in component_fields(catalog, component)
+        if field.identifies and (paths is None or field.path in paths)
+        for value in _values_at(node, field.segments)
+        if isinstance(value, str) and not is_label(value)
+    ]
+
+
+def _values_at(value: JsonValue, segments: list[str | int]) -> list[JsonValue]:
+    if not segments:
+        return [value]
+    head, *rest = segments
+    if isinstance(value, dict) and isinstance(head, str) and head in value:
+        return _values_at(value[head], rest)
+    if isinstance(value, list) and isinstance(head, int) and 0 <= head < len(value):
+        return _values_at(value[head], rest)
+    if isinstance(value, list) and isinstance(head, str):
+        return [found for item in value for found in _values_at(item, segments)]
+    return []