feat: add models for physical files, policies, and user management

- Implement PhysicalFile model to manage physical file references and reference counting. - Create Policy model with associated options and group links for storage policies. - Introduce Redeem and Report models for handling redeem codes and reports. - Add Settings model for site configuration and user settings management. - Develop Share model for sharing objects with unique codes and associated metadata. - Implement SourceLink model for managing download links associated with objects. - Create StoragePack model for managing user storage packages. - Add Tag model for user-defined tags with manual and automatic types. - Implement Task model for managing background tasks with status tracking. - Develop User model with comprehensive user management features including authentication. - Introduce UserAuthn model for managing WebAuthn credentials. - Create WebDAV model for managing WebDAV accounts associated with users.
2026-02-10 16:25:49 +08:00
parent 62c671e07b
commit 209cb24ab4
92 changed files with 3640 additions and 1444 deletions
--- a/sqlmodels/base/README.md
+++ b/sqlmodels/base/README.md
@@ -0,0 +1,657 @@
+# SQLModels Base Module
+
+This module provides `SQLModelBase`, the root base class for all SQLModel models in this project. It includes a custom metaclass with automatic type injection and Python 3.14 compatibility.
+
+**Note**: Table base classes (`TableBaseMixin`, `UUIDTableBaseMixin`) and polymorphic utilities have been migrated to the [`sqlmodels.mixin`](../mixin/README.md) module. See the mixin documentation for CRUD operations, polymorphic inheritance patterns, and pagination utilities.
+
+## Table of Contents
+
+- [Overview](#overview)
+- [Migration Notice](#migration-notice)
+- [Python 3.14 Compatibility](#python-314-compatibility)
+- [Core Component](#core-component)
+  - [SQLModelBase](#sqlmodelbase)
+- [Metaclass Features](#metaclass-features)
+  - [Automatic sa_type Injection](#automatic-sa_type-injection)
+  - [Table Configuration](#table-configuration)
+  - [Polymorphic Support](#polymorphic-support)
+- [Custom Types Integration](#custom-types-integration)
+- [Best Practices](#best-practices)
+- [Troubleshooting](#troubleshooting)
+
+## Overview
+
+The `sqlmodels.base` module provides `SQLModelBase`, the foundational base class for all SQLModel models. It features:
+
+- **Smart metaclass** that automatically extracts and injects SQLAlchemy types from type annotations
+- **Python 3.14 compatibility** through comprehensive PEP 649/749 support
+- **Flexible configuration** through class parameters and automatic docstring support
+- **Type-safe annotations** with automatic validation
+
+All models in this project should directly or indirectly inherit from `SQLModelBase`.
+
+---
+
+## Migration Notice
+
+As of the recent refactoring, the following components have been moved:
+
+| Component | Old Location | New Location |
+|-----------|-------------|--------------|
+| `TableBase` → `TableBaseMixin` | `sqlmodels.base` | `sqlmodels.mixin` |
+| `UUIDTableBase` → `UUIDTableBaseMixin` | `sqlmodels.base` | `sqlmodels.mixin` |
+| `PolymorphicBaseMixin` | `sqlmodels.base` | `sqlmodels.mixin` |
+| `create_subclass_id_mixin()` | `sqlmodels.base` | `sqlmodels.mixin` |
+| `AutoPolymorphicIdentityMixin` | `sqlmodels.base` | `sqlmodels.mixin` |
+| `TableViewRequest` | `sqlmodels.base` | `sqlmodels.mixin` |
+| `now()`, `now_date()` | `sqlmodels.base` | `sqlmodels.mixin` |
+
+**Update your imports**:
+
+```python
+# ❌ Old (deprecated)
+from sqlmodels.base import TableBase, UUIDTableBase
+
+# ✅ New (correct)
+from sqlmodels.mixin import TableBaseMixin, UUIDTableBaseMixin
+```
+
+For detailed documentation on table mixins, CRUD operations, and polymorphic patterns, see [`sqlmodels/mixin/README.md`](../mixin/README.md).
+
+---
+
+## Python 3.14 Compatibility
+
+### Overview
+
+This module provides full compatibility with **Python 3.14's PEP 649** (Deferred Evaluation of Annotations) and **PEP 749** (making it the default).
+
+**Key Changes in Python 3.14**:
+- Annotations are no longer evaluated at class definition time
+- Type hints are stored as deferred code objects
+- `__annotate__` function generates annotations on demand
+- Forward references become `ForwardRef` objects
+
+### Implementation Strategy
+
+We use **`typing.get_type_hints()`** as the universal annotations resolver:
+
+```python
+def _resolve_annotations(attrs: dict[str, Any]) -> tuple[...]:
+    # Create temporary proxy class
+    temp_cls = type('AnnotationProxy', (object,), dict(attrs))
+
+    # Use get_type_hints with include_extras=True
+    evaluated = get_type_hints(
+        temp_cls,
+        globalns=module_globals,
+        localns=localns,
+        include_extras=True  # Preserve Annotated metadata
+    )
+
+    return dict(evaluated), {}, module_globals, localns
+```
+
+**Why `get_type_hints()`?**
+- ✅ Works across Python 3.10-3.14+
+- ✅ Handles PEP 649 automatically
+- ✅ Preserves `Annotated` metadata (with `include_extras=True`)
+- ✅ Resolves forward references
+- ✅ Recommended by Python documentation
+
+### SQLModel Compatibility Patch
+
+**Problem**: SQLModel's `get_sqlalchemy_type()` doesn't recognize custom types with `__sqlmodel_sa_type__` attribute.
+
+**Solution**: Global monkey-patch that checks for SQLAlchemy type before falling back to original logic:
+
+```python
+if sys.version_info >= (3, 14):
+    def _patched_get_sqlalchemy_type(field):
+        annotation = getattr(field, 'annotation', None)
+        if annotation is not None:
+            # Priority 1: Check __sqlmodel_sa_type__ attribute
+            # Handles NumpyVector[dims, dtype] and similar custom types
+            if hasattr(annotation, '__sqlmodel_sa_type__'):
+                return annotation.__sqlmodel_sa_type__
+
+            # Priority 2: Check Annotated metadata
+            if get_origin(annotation) is Annotated:
+                for metadata in get_args(annotation)[1:]:
+                    if hasattr(metadata, '__sqlmodel_sa_type__'):
+                        return metadata.__sqlmodel_sa_type__
+
+            # ... handle ForwardRef, ClassVar, etc.
+
+        return _original_get_sqlalchemy_type(field)
+```
+
+### Supported Patterns
+
+#### Pattern 1: Direct Custom Type Usage
+```python
+from sqlmodels.sqlmodel_types.dialects.postgresql import NumpyVector
+from sqlmodels.mixin import UUIDTableBaseMixin
+
+class SpeakerInfo(UUIDTableBaseMixin, table=True):
+    embedding: NumpyVector[256, np.float32]
+    """Voice embedding - sa_type automatically extracted"""
+```
+
+#### Pattern 2: Annotated Wrapper
+```python
+from typing import Annotated
+from sqlmodels.mixin import UUIDTableBaseMixin
+
+EmbeddingVector = Annotated[np.ndarray, NumpyVector[256, np.float32]]
+
+class SpeakerInfo(UUIDTableBaseMixin, table=True):
+    embedding: EmbeddingVector
+```
+
+#### Pattern 3: Array Type
+```python
+from sqlmodels.sqlmodel_types.dialects.postgresql import Array
+from sqlmodels.mixin import TableBaseMixin
+
+class ServerConfig(TableBaseMixin, table=True):
+    protocols: Array[ProtocolEnum]
+    """Allowed protocols - sa_type from Array handler"""
+```
+
+### Migration from Python 3.13
+
+**No code changes required!** The implementation is transparent:
+
+- Uses `typing.get_type_hints()` which works in both Python 3.13 and 3.14
+- Custom types already use `__sqlmodel_sa_type__` attribute
+- Monkey-patch only activates for Python 3.14+
+
+---
+
+## Core Component
+
+### SQLModelBase
+
+`SQLModelBase` is the root base class for all SQLModel models. It uses a custom metaclass (`__DeclarativeMeta`) that provides advanced features beyond standard SQLModel capabilities.
+
+**Key Features**:
+- Automatic `use_attribute_docstrings` configuration (use docstrings instead of `Field(description=...)`)
+- Automatic `validate_by_name` configuration
+- Custom metaclass for sa_type injection and polymorphic setup
+- Integration with Pydantic v2
+- Python 3.14 PEP 649 compatibility
+
+**Usage**:
+
+```python
+from sqlmodels.base import SQLModelBase
+
+class UserBase(SQLModelBase):
+    name: str
+    """User's display name"""
+
+    email: str
+    """User's email address"""
+```
+
+**Important Notes**:
+- Use **docstrings** for field descriptions, not `Field(description=...)`
+- Do NOT override `model_config` in subclasses (it's already configured in SQLModelBase)
+- This class should be used for non-table models (DTOs, request/response models)
+
+**For table models**, use mixins from `sqlmodels.mixin`:
+- `TableBaseMixin` - Integer primary key with timestamps
+- `UUIDTableBaseMixin` - UUID primary key with timestamps
+
+See [`sqlmodels/mixin/README.md`](../mixin/README.md) for complete table mixin documentation.
+
+---
+
+## Metaclass Features
+
+### Automatic sa_type Injection
+
+The metaclass automatically extracts SQLAlchemy types from custom type annotations, enabling clean syntax for complex database types.
+
+**Before** (verbose):
+```python
+from sqlmodels.sqlmodel_types.dialects.postgresql.numpy_vector import _NumpyVectorSQLAlchemyType
+from sqlmodels.mixin import UUIDTableBaseMixin
+
+class SpeakerInfo(UUIDTableBaseMixin, table=True):
+    embedding: np.ndarray = Field(
+        sa_type=_NumpyVectorSQLAlchemyType(256, np.float32)
+    )
+```
+
+**After** (clean):
+```python
+from sqlmodels.sqlmodel_types.dialects.postgresql import NumpyVector
+from sqlmodels.mixin import UUIDTableBaseMixin
+
+class SpeakerInfo(UUIDTableBaseMixin, table=True):
+    embedding: NumpyVector[256, np.float32]
+    """Speaker voice embedding"""
+```
+
+**How It Works**:
+
+The metaclass uses a three-tier detection strategy:
+
+1. **Direct `__sqlmodel_sa_type__` attribute** (Priority 1)
+   ```python
+   if hasattr(annotation, '__sqlmodel_sa_type__'):
+       return annotation.__sqlmodel_sa_type__
+   ```
+
+2. **Annotated metadata** (Priority 2)
+   ```python
+   # For Annotated[np.ndarray, NumpyVector[256, np.float32]]
+   if get_origin(annotation) is typing.Annotated:
+       for item in metadata_items:
+           if hasattr(item, '__sqlmodel_sa_type__'):
+               return item.__sqlmodel_sa_type__
+   ```
+
+3. **Pydantic Core Schema metadata** (Priority 3)
+   ```python
+   schema = annotation.__get_pydantic_core_schema__(...)
+   if schema['metadata'].get('sa_type'):
+       return schema['metadata']['sa_type']
+   ```
+
+After extracting `sa_type`, the metaclass:
+- Creates `Field(sa_type=sa_type)` if no Field is defined
+- Injects `sa_type` into existing Field if not already set
+- Respects explicit `Field(sa_type=...)` (no override)
+
+**Supported Patterns**:
+
+```python
+from sqlmodels.mixin import UUIDTableBaseMixin
+
+# Pattern 1: Direct usage (recommended)
+class Model(UUIDTableBaseMixin, table=True):
+    embedding: NumpyVector[256, np.float32]
+
+# Pattern 2: With Field constraints
+class Model(UUIDTableBaseMixin, table=True):
+    embedding: NumpyVector[256, np.float32] = Field(nullable=False)
+
+# Pattern 3: Annotated wrapper
+EmbeddingVector = Annotated[np.ndarray, NumpyVector[256, np.float32]]
+
+class Model(UUIDTableBaseMixin, table=True):
+    embedding: EmbeddingVector
+
+# Pattern 4: Explicit sa_type (override)
+class Model(UUIDTableBaseMixin, table=True):
+    embedding: NumpyVector[256, np.float32] = Field(
+        sa_type=_NumpyVectorSQLAlchemyType(128, np.float16)
+    )
+```
+
+### Table Configuration
+
+The metaclass provides smart defaults and flexible configuration:
+
+**Automatic `table=True`**:
+```python
+# Classes inheriting from TableBaseMixin automatically get table=True
+from sqlmodels.mixin import UUIDTableBaseMixin
+
+class MyModel(UUIDTableBaseMixin):  # table=True is automatic
+    pass
+```
+
+**Convenient mapper arguments**:
+```python
+# Instead of verbose __mapper_args__
+from sqlmodels.mixin import UUIDTableBaseMixin
+
+class MyModel(
+    UUIDTableBaseMixin,
+    polymorphic_on='_polymorphic_name',
+    polymorphic_abstract=True
+):
+    pass
+
+# Equivalent to:
+class MyModel(UUIDTableBaseMixin):
+    __mapper_args__ = {
+        'polymorphic_on': '_polymorphic_name',
+        'polymorphic_abstract': True
+    }
+```
+
+**Smart merging**:
+```python
+# Dictionary and keyword arguments are merged
+from sqlmodels.mixin import UUIDTableBaseMixin
+
+class MyModel(
+    UUIDTableBaseMixin,
+    mapper_args={'version_id_col': 'version'},
+    polymorphic_on='type'  # Merged into __mapper_args__
+):
+    pass
+```
+
+### Polymorphic Support
+
+The metaclass supports SQLAlchemy's joined table inheritance through convenient parameters:
+
+**Supported parameters**:
+- `polymorphic_on`: Discriminator column name
+- `polymorphic_identity`: Identity value for this class
+- `polymorphic_abstract`: Whether this is an abstract base
+- `table_args`: SQLAlchemy table arguments
+- `table_name`: Override table name (becomes `__tablename__`)
+
+**For complete polymorphic inheritance patterns**, including `PolymorphicBaseMixin`, `create_subclass_id_mixin()`, and `AutoPolymorphicIdentityMixin`, see [`sqlmodels/mixin/README.md`](../mixin/README.md).
+
+---
+
+## Custom Types Integration
+
+### Using NumpyVector
+
+The `NumpyVector` type demonstrates automatic sa_type injection:
+
+```python
+from sqlmodels.sqlmodel_types.dialects.postgresql import NumpyVector
+from sqlmodels.mixin import UUIDTableBaseMixin
+import numpy as np
+
+class SpeakerInfo(UUIDTableBaseMixin, table=True):
+    embedding: NumpyVector[256, np.float32]
+    """Speaker voice embedding - sa_type automatically injected"""
+```
+
+**How NumpyVector works**:
+
+```python
+# NumpyVector[dims, dtype] returns a class with:
+class _NumpyVectorType:
+    __sqlmodel_sa_type__ = _NumpyVectorSQLAlchemyType(dimensions, dtype)
+
+    @classmethod
+    def __get_pydantic_core_schema__(cls, source_type, handler):
+        return handler.generate_schema(np.ndarray)
+```
+
+This dual approach ensures:
+1. Metaclass can extract `sa_type` via `__sqlmodel_sa_type__`
+2. Pydantic can validate as `np.ndarray`
+
+### Creating Custom SQLAlchemy Types
+
+To create types that work with automatic injection, provide one of:
+
+**Option 1: `__sqlmodel_sa_type__` attribute** (preferred):
+
+```python
+from sqlalchemy import TypeDecorator, String
+
+class UpperCaseString(TypeDecorator):
+    impl = String
+
+    def process_bind_param(self, value, dialect):
+        return value.upper() if value else value
+
+class UpperCaseType:
+    __sqlmodel_sa_type__ = UpperCaseString()
+
+    @classmethod
+    def __get_pydantic_core_schema__(cls, source_type, handler):
+        return core_schema.str_schema()
+
+# Usage
+from sqlmodels.mixin import UUIDTableBaseMixin
+
+class MyModel(UUIDTableBaseMixin, table=True):
+    code: UpperCaseType  # Automatically uses UpperCaseString()
+```
+
+**Option 2: Pydantic metadata with sa_type**:
+
+```python
+def __get_pydantic_core_schema__(self, source_type, handler):
+    return core_schema.json_or_python_schema(
+        json_schema=core_schema.str_schema(),
+        python_schema=core_schema.str_schema(),
+        metadata={'sa_type': UpperCaseString()}
+    )
+```
+
+**Option 3: Using Annotated**:
+
+```python
+from typing import Annotated
+from sqlmodels.mixin import UUIDTableBaseMixin
+
+UpperCase = Annotated[str, UpperCaseType()]
+
+class MyModel(UUIDTableBaseMixin, table=True):
+    code: UpperCase
+```
+
+---
+
+## Best Practices
+
+### 1. Inherit from correct base classes
+
+```python
+from sqlmodels.base import SQLModelBase
+from sqlmodels.mixin import TableBaseMixin, UUIDTableBaseMixin
+
+# ✅ For non-table models (DTOs, requests, responses)
+class UserBase(SQLModelBase):
+    name: str
+
+# ✅ For table models with UUID primary key
+class User(UserBase, UUIDTableBaseMixin, table=True):
+    email: str
+
+# ✅ For table models with custom primary key
+class LegacyUser(TableBaseMixin, table=True):
+    id: int = Field(primary_key=True)
+    username: str
+```
+
+### 2. Use docstrings for field descriptions
+
+```python
+from sqlmodels.mixin import UUIDTableBaseMixin
+
+# ✅ Recommended
+class User(UUIDTableBaseMixin, table=True):
+    name: str
+    """User's display name"""
+
+# ❌ Avoid
+class User(UUIDTableBaseMixin, table=True):
+    name: str = Field(description="User's display name")
+```
+
+**Why?** SQLModelBase has `use_attribute_docstrings=True`, so docstrings automatically become field descriptions in API docs.
+
+### 3. Leverage automatic sa_type injection
+
+```python
+from sqlmodels.mixin import UUIDTableBaseMixin
+
+# ✅ Clean and recommended
+class SpeakerInfo(UUIDTableBaseMixin, table=True):
+    embedding: NumpyVector[256, np.float32]
+    """Voice embedding"""
+
+# ❌ Verbose and unnecessary
+class SpeakerInfo(UUIDTableBaseMixin, table=True):
+    embedding: np.ndarray = Field(
+        sa_type=_NumpyVectorSQLAlchemyType(256, np.float32)
+    )
+```
+
+### 4. Follow polymorphic naming conventions
+
+See [`sqlmodels/mixin/README.md`](../mixin/README.md) for complete polymorphic inheritance patterns using `PolymorphicBaseMixin`, `create_subclass_id_mixin()`, and `AutoPolymorphicIdentityMixin`.
+
+### 5. Separate Base, Parent, and Implementation classes
+
+```python
+from abc import ABC, abstractmethod
+from sqlmodels.base import SQLModelBase
+from sqlmodels.mixin import UUIDTableBaseMixin, PolymorphicBaseMixin
+
+# ✅ Recommended structure
+class ASRBase(SQLModelBase):
+    """Pure data fields, no table"""
+    name: str
+    base_url: str
+
+class ASR(ASRBase, UUIDTableBaseMixin, PolymorphicBaseMixin, ABC):
+    """Abstract parent with table"""
+    @abstractmethod
+    async def transcribe(self, audio: bytes) -> str:
+        pass
+
+class WhisperASR(ASR, table=True):
+    """Concrete implementation"""
+    model_size: str
+
+    async def transcribe(self, audio: bytes) -> str:
+        # Implementation
+        pass
+```
+
+**Why?**
+- Base class can be reused for DTOs
+- Parent class defines the polymorphic hierarchy
+- Implementation classes are clean and focused
+
+---
+
+## Troubleshooting
+
+### Issue: ValueError: X has no matching SQLAlchemy type
+
+**Solution**: Ensure your custom type provides `__sqlmodel_sa_type__` attribute or proper Pydantic metadata with `sa_type`.
+
+```python
+# ✅ Provide __sqlmodel_sa_type__
+class MyType:
+    __sqlmodel_sa_type__ = MyCustomSQLAlchemyType()
+```
+
+### Issue: Can't generate DDL for NullType()
+
+**Symptoms**: Error during table creation saying a column has `NullType`.
+
+**Root Cause**: Custom type's `sa_type` not detected by SQLModel.
+
+**Solution**:
+1. Ensure your type has `__sqlmodel_sa_type__` class attribute
+2. Check that the monkey-patch is active (`sys.version_info >= (3, 14)`)
+3. Verify type annotation is correct (not a string forward reference)
+
+```python
+from sqlmodels.mixin import UUIDTableBaseMixin
+
+# ✅ Correct
+class Model(UUIDTableBaseMixin, table=True):
+    data: NumpyVector[256, np.float32]  # __sqlmodel_sa_type__ detected
+
+# ❌ Wrong (string annotation)
+class Model(UUIDTableBaseMixin, table=True):
+    data: 'NumpyVector[256, np.float32]'  # sa_type lost
+```
+
+### Issue: Polymorphic identity conflicts
+
+**Symptoms**: SQLAlchemy raises errors about duplicate polymorphic identities.
+
+**Solution**:
+1. Check that each concrete class has a unique identity
+2. Use `AutoPolymorphicIdentityMixin` for automatic naming
+3. Manually specify identity if needed:
+   ```python
+   class MyClass(Parent, polymorphic_identity='unique.name', table=True):
+       pass
+   ```
+
+### Issue: Python 3.14 annotation errors
+
+**Symptoms**: Errors related to `__annotations__` or type resolution.
+
+**Solution**: The implementation uses `get_type_hints()` which handles PEP 649 automatically. If issues persist:
+1. Check for manual `__annotations__` manipulation (avoid it)
+2. Ensure all types are properly imported
+3. Avoid `from __future__ import annotations` (can cause SQLModel issues)
+
+### Issue: Polymorphic and CRUD-related errors
+
+For issues related to polymorphic inheritance, CRUD operations, or table mixins, see the troubleshooting section in [`sqlmodels/mixin/README.md`](../mixin/README.md).
+
+---
+
+## Implementation Details
+
+For developers modifying this module:
+
+**Core files**:
+- `sqlmodel_base.py` - Contains `__DeclarativeMeta` and `SQLModelBase`
+- `../mixin/table.py` - Contains `TableBaseMixin` and `UUIDTableBaseMixin`
+- `../mixin/polymorphic.py` - Contains `PolymorphicBaseMixin`, `create_subclass_id_mixin()`, and `AutoPolymorphicIdentityMixin`
+
+**Key functions in this module**:
+
+1. **`_resolve_annotations(attrs: dict[str, Any])`**
+   - Uses `typing.get_type_hints()` for Python 3.14 compatibility
+   - Returns tuple: `(annotations, annotation_strings, globalns, localns)`
+   - Preserves `Annotated` metadata with `include_extras=True`
+
+2. **`_extract_sa_type_from_annotation(annotation: Any) -> Any | None`**
+   - Extracts SQLAlchemy type from type annotations
+   - Supports `__sqlmodel_sa_type__`, `Annotated`, and Pydantic core schema
+   - Called by metaclass during class creation
+
+3. **`_patched_get_sqlalchemy_type(field)`** (Python 3.14+)
+   - Global monkey-patch for SQLModel
+   - Checks `__sqlmodel_sa_type__` before falling back to original logic
+   - Handles custom types like `NumpyVector` and `Array`
+
+4. **`__DeclarativeMeta.__new__()`**
+   - Processes class definition parameters
+   - Injects `sa_type` into field definitions
+   - Sets up `__mapper_args__`, `__table_args__`, etc.
+   - Handles Python 3.14 annotations via `get_type_hints()`
+
+**Metaclass processing order**:
+1. Check if class should be a table (`_has_table_mixin`)
+2. Collect `__mapper_args__` from kwargs and explicit dict
+3. Process `table_args`, `table_name`, `abstract` parameters
+4. Resolve annotations using `get_type_hints()`
+5. For each field, try to extract `sa_type` and inject into Field
+6. Call parent metaclass with cleaned kwargs
+
+For table mixin implementation details, see [`sqlmodels/mixin/README.md`](../mixin/README.md).
+
+---
+
+## See Also
+
+**Project Documentation**:
+- [SQLModel Mixin Documentation](../mixin/README.md) - Table mixins, CRUD operations, polymorphic patterns
+- [Project Coding Standards (CLAUDE.md)](/mnt/c/Users/Administrator/PycharmProjects/emoecho-backend-server/CLAUDE.md)
+- [Custom SQLModel Types Guide](/mnt/c/Users/Administrator/PycharmProjects/emoecho-backend-server/sqlmodels/sqlmodel_types/README.md)
+
+**External References**:
+- [SQLAlchemy Joined Table Inheritance](https://docs.sqlalchemy.org/en/20/orm/inheritance.html#joined-table-inheritance)
+- [Pydantic V2 Documentation](https://docs.pydantic.dev/latest/)
+- [SQLModel Documentation](https://sqlmodel.tiangolo.com/)
+- [PEP 649: Deferred Evaluation of Annotations](https://peps.python.org/pep-0649/)
+- [PEP 749: Implementing PEP 649](https://peps.python.org/pep-0749/)
+- [Python Annotations Best Practices](https://docs.python.org/3/howto/annotations.html)
--- a/sqlmodels/base/init.py
+++ b/sqlmodels/base/init.py
@@ -0,0 +1,12 @@
+"""
+SQLModel 基础模块
+
+包含：
+- SQLModelBase: 所有 SQLModel 类的基类（真正的基类）
+
+注意：
+    TableBase, UUIDTableBase, PolymorphicBaseMixin 已迁移到 sqlmodels.mixin
+    为了避免循环导入，此处不再重新导出它们
+    请直接从 sqlmodels.mixin 导入这些类
+"""
+from .sqlmodel_base import SQLModelBase
--- a/sqlmodels/base/sqlmodel_base.py
+++ b/sqlmodels/base/sqlmodel_base.py
@@ -0,0 +1,846 @@
+import sys
+import typing
+from typing import Any, Mapping, get_args, get_origin, get_type_hints
+
+from pydantic import ConfigDict
+from pydantic.fields import FieldInfo
+from pydantic_core import PydanticUndefined as Undefined
+from sqlalchemy.orm import Mapped
+from sqlmodel import Field, SQLModel
+from sqlmodel.main import SQLModelMetaclass
+
+# Python 3.14+ PEP 649支持
+if sys.version_info >= (3, 14):
+    import annotationlib
+
+    # 全局Monkey-patch: 修复SQLModel在Python 3.14上的兼容性问题
+    import sqlmodel.main
+    _original_get_sqlalchemy_type = sqlmodel.main.get_sqlalchemy_type
+
+    def _patched_get_sqlalchemy_type(field):
+        """
+        修复SQLModel的get_sqlalchemy_type函数，处理Python 3.14的类型问题。
+
+        问题：
+        1. ForwardRef对象（来自Relationship字段）会导致issubclass错误
+        2. typing._GenericAlias对象（如ClassVar[T]）也会导致同样问题
+        3. list/dict等泛型类型在没有Field/Relationship时可能导致错误
+        4. Mapped类型在Python 3.14下可能出现在annotation中
+        5. Annotated类型可能包含sa_type metadata（如Array[T]）
+        6. 自定义类型（如NumpyVector）有__sqlmodel_sa_type__属性
+        7. Pydantic已处理的Annotated类型会将metadata存储在field.metadata中
+
+        解决：
+        - 优先检查field.metadata中的__get_pydantic_core_schema__（Pydantic已处理的情况）
+        - 检测__sqlmodel_sa_type__属性（NumpyVector等）
+        - 检测Relationship/ClassVar等返回None
+        - 对于Annotated类型，尝试提取sa_type metadata
+        - 其他情况调用原始函数
+        """
+        # 优先检查 field.metadata（Pydantic已处理Annotated类型的情况）
+        # 当使用 Array[T] 或 Annotated[T, metadata] 时，Pydantic会将metadata存储在这里
+        metadata = getattr(field, 'metadata', None)
+        if metadata:
+            # metadata是一个列表，包含所有Annotated的元数据项
+            for metadata_item in metadata:
+                # 检查metadata_item是否有__get_pydantic_core_schema__方法
+                if hasattr(metadata_item, '__get_pydantic_core_schema__'):
+                    try:
+                        # 调用获取schema
+                        schema = metadata_item.__get_pydantic_core_schema__(None, None)
+                        # 检查schema的metadata中是否有sa_type
+                        if isinstance(schema, dict) and 'metadata' in schema:
+                            sa_type = schema['metadata'].get('sa_type')
+                            if sa_type is not None:
+                                return sa_type
+                    except (TypeError, AttributeError, KeyError):
+                        # Pydantic schema获取可能失败（类型不匹配、缺少属性等）
+                        # 这是正常情况，继续检查下一个metadata项
+                        pass
+
+        annotation = getattr(field, 'annotation', None)
+        if annotation is not None:
+            # 优先检查 __sqlmodel_sa_type__ 属性
+            # 这处理 NumpyVector[dims, dtype] 等自定义类型
+            if hasattr(annotation, '__sqlmodel_sa_type__'):
+                return annotation.__sqlmodel_sa_type__
+
+            # 检查自定义类型（如JSON100K）的 __get_pydantic_core_schema__ 方法
+            # 这些类型在schema的metadata中定义sa_type
+            if hasattr(annotation, '__get_pydantic_core_schema__'):
+                try:
+                    # 调用获取schema（传None作为handler，因为我们只需要metadata）
+                    schema = annotation.__get_pydantic_core_schema__(annotation, lambda x: None)
+                    # 检查schema的metadata中是否有sa_type
+                    if isinstance(schema, dict) and 'metadata' in schema:
+                        sa_type = schema['metadata'].get('sa_type')
+                        if sa_type is not None:
+                            return sa_type
+                except (TypeError, AttributeError, KeyError):
+                    # Schema获取失败，继续其他检查
+                    pass
+
+            anno_type_name = type(annotation).__name__
+
+            # ForwardRef: Relationship字段的annotation
+            if anno_type_name == 'ForwardRef':
+                return None
+
+            # AnnotatedAlias: 检查是否有sa_type metadata（如Array[T]）
+            if anno_type_name == 'AnnotatedAlias' or anno_type_name == '_AnnotatedAlias':
+                from typing import get_origin, get_args
+                import typing
+
+                # 尝试提取Annotated的metadata
+                if hasattr(typing, 'get_args'):
+                    args = get_args(annotation)
+                    # args[0]是实际类型，args[1:]是metadata
+                    for metadata in args[1:]:
+                        # 检查metadata是否有__get_pydantic_core_schema__方法
+                        if hasattr(metadata, '__get_pydantic_core_schema__'):
+                            try:
+                                # 调用获取schema
+                                schema = metadata.__get_pydantic_core_schema__(None, None)
+                                # 检查schema中是否有sa_type
+                                if isinstance(schema, dict) and 'metadata' in schema:
+                                    sa_type = schema['metadata'].get('sa_type')
+                                    if sa_type is not None:
+                                        return sa_type
+                            except (TypeError, AttributeError, KeyError):
+                                # Annotated metadata的schema获取可能失败
+                                # 这是正常的类型检查过程，继续检查下一个metadata
+                                pass
+
+            # _GenericAlias或GenericAlias: typing泛型类型
+            if anno_type_name in ('_GenericAlias', 'GenericAlias'):
+                from typing import get_origin
+                import typing
+                origin = get_origin(annotation)
+
+                # ClassVar必须跳过
+                if origin is typing.ClassVar:
+                    return None
+
+                # list/dict/tuple/set等内置泛型，如果字段没有明确的Field或Relationship，也跳过
+                # 这通常意味着它是Relationship字段或类变量
+                if origin in (list, dict, tuple, set):
+                    # 检查field_info是否存在且有意义
+                    # Relationship字段会有特殊的field_info
+                    field_info = getattr(field, 'field_info', None)
+                    if field_info is None:
+                        return None
+
+            # Mapped: SQLAlchemy 2.0的Mapped类型，SQLModel不应该处理
+            # 这可能是从父类继承的字段或Python 3.14注解处理的副作用
+            # 检查类型名称和annotation的字符串表示
+            if 'Mapped' in anno_type_name or 'Mapped' in str(annotation):
+                return None
+
+            # 检查annotation是否是Mapped类或其实例
+            try:
+                from sqlalchemy.orm import Mapped as SAMapped
+                # 检查origin（对于Mapped[T]这种泛型）
+                from typing import get_origin
+                if get_origin(annotation) is SAMapped:
+                    return None
+                # 检查类型本身
+                if annotation is SAMapped or isinstance(annotation, type) and issubclass(annotation, SAMapped):
+                    return None
+            except (ImportError, TypeError):
+                # 如果SQLAlchemy没有Mapped或检查失败，继续
+                pass
+
+        # 其他情况正常处理
+        return _original_get_sqlalchemy_type(field)
+
+    sqlmodel.main.get_sqlalchemy_type = _patched_get_sqlalchemy_type
+
+    # 第二个Monkey-patch: 修复继承表类中InstrumentedAttribute作为默认值的问题
+    # 在Python 3.14 + SQLModel组合下，当子类（如SMSBaoProvider）继承父类（如VerificationCodeProvider）时，
+    # 父类的关系字段（如server_config）会在子类的model_fields中出现，
+    # 但其default值错误地设置为InstrumentedAttribute对象，而不是None
+    # 这导致实例化时尝试设置InstrumentedAttribute为字段值，触发SQLAlchemy内部错误
+    import sqlmodel._compat as _compat
+    from sqlalchemy.orm import attributes as _sa_attributes
+
+    _original_sqlmodel_table_construct = _compat.sqlmodel_table_construct
+
+    def _patched_sqlmodel_table_construct(self_instance, values):
+        """
+        修复sqlmodel_table_construct，跳过InstrumentedAttribute默认值
+
+        问题：
+        - 继承自polymorphic基类的表类（如FishAudioTTS, SMSBaoProvider）
+        - 其model_fields中的继承字段default值为InstrumentedAttribute
+        - 原函数尝试将InstrumentedAttribute设置为字段值
+        - SQLAlchemy无法处理，抛出 '_sa_instance_state' 错误
+
+        解决：
+        - 只设置用户提供的值和非InstrumentedAttribute默认值
+        - InstrumentedAttribute默认值跳过（让SQLAlchemy自己处理）
+        """
+        cls = type(self_instance)
+
+        # 收集要设置的字段值
+        fields_to_set = {}
+
+        for name, field in cls.model_fields.items():
+            # 如果用户提供了值，直接使用
+            if name in values:
+                fields_to_set[name] = values[name]
+                continue
+
+            # 否则检查默认值
+            # 跳过InstrumentedAttribute默认值 - 这些是继承字段的错误默认值
+            if isinstance(field.default, _sa_attributes.InstrumentedAttribute):
+                continue
+
+            # 使用正常的默认值
+            if field.default is not Undefined:
+                fields_to_set[name] = field.default
+            elif field.default_factory is not None:
+                fields_to_set[name] = field.get_default(call_default_factory=True)
+
+        # 设置属性 - 只设置非InstrumentedAttribute值
+        for key, value in fields_to_set.items():
+            if not isinstance(value, _sa_attributes.InstrumentedAttribute):
+                setattr(self_instance, key, value)
+
+        # 设置Pydantic内部属性
+        object.__setattr__(self_instance, '__pydantic_fields_set__', set(values.keys()))
+        if not cls.__pydantic_root_model__:
+            _extra = None
+            if cls.model_config.get('extra') == 'allow':
+                _extra = {}
+                for k, v in values.items():
+                    if k not in cls.model_fields:
+                        _extra[k] = v
+            object.__setattr__(self_instance, '__pydantic_extra__', _extra)
+
+        if cls.__pydantic_post_init__:
+            self_instance.model_post_init(None)
+        elif not cls.__pydantic_root_model__:
+            object.__setattr__(self_instance, '__pydantic_private__', None)
+
+        # 设置关系
+        for key in self_instance.__sqlmodel_relationships__:
+            value = values.get(key, Undefined)
+            if value is not Undefined:
+                setattr(self_instance, key, value)
+
+        return self_instance
+
+    _compat.sqlmodel_table_construct = _patched_sqlmodel_table_construct
+else:
+    annotationlib = None
+
+
+def _extract_sa_type_from_annotation(annotation: Any) -> Any | None:
+    """
+    从类型注解中提取SQLAlchemy类型。
+
+    支持以下形式：
+    1. NumpyVector[256, np.float32] - 直接使用类型（有__sqlmodel_sa_type__属性）
+    2. Annotated[np.ndarray, NumpyVector[256, np.float32]] - Annotated包装
+    3. 任何有__get_pydantic_core_schema__且返回metadata['sa_type']的类型
+
+    Args:
+        annotation: 字段的类型注解
+
+    Returns:
+        提取到的SQLAlchemy类型，如果没有则返回None
+    """
+    # 方法1：直接检查类型本身是否有__sqlmodel_sa_type__属性
+    # 这涵盖了 NumpyVector[256, np.float32] 这种直接使用的情况
+    if hasattr(annotation, '__sqlmodel_sa_type__'):
+        return annotation.__sqlmodel_sa_type__
+
+    # 方法2：检查是否为Annotated类型
+    if get_origin(annotation) is typing.Annotated:
+        # 获取元数据项（跳过第一个实际类型参数）
+        args = get_args(annotation)
+        if len(args) >= 2:
+            metadata_items = args[1:]  # 第一个是实际类型，后面都是元数据
+
+            # 遍历元数据，查找包含sa_type的项
+            for item in metadata_items:
+                # 检查元数据项是否有__sqlmodel_sa_type__属性
+                if hasattr(item, '__sqlmodel_sa_type__'):
+                    return item.__sqlmodel_sa_type__
+
+                # 检查是否有__get_pydantic_core_schema__方法
+                if hasattr(item, '__get_pydantic_core_schema__'):
+                    try:
+                        # 调用该方法获取core schema
+                        schema = item.__get_pydantic_core_schema__(
+                            annotation,
+                            lambda x: None  # 虚拟handler
+                        )
+                        # 检查schema的metadata中是否有sa_type
+                        if isinstance(schema, dict) and 'metadata' in schema:
+                            sa_type = schema['metadata'].get('sa_type')
+                            if sa_type is not None:
+                                return sa_type
+                    except (TypeError, AttributeError, KeyError, ValueError):
+                        # Pydantic core schema获取可能失败：
+                        # - TypeError: 参数不匹配
+                        # - AttributeError: metadata不存在
+                        # - KeyError: schema结构不符合预期
+                        # - ValueError: 无效的类型定义
+                        # 这是正常的类型探测过程，继续检查下一个metadata项
+                        pass
+
+    # 方法3：检查类型本身是否有__get_pydantic_core_schema__
+    # （虽然NumpyVector已经在方法1处理，但这是通用的fallback）
+    if hasattr(annotation, '__get_pydantic_core_schema__'):
+        try:
+            schema = annotation.__get_pydantic_core_schema__(
+                annotation,
+                lambda x: None  # 虚拟handler
+            )
+            if isinstance(schema, dict) and 'metadata' in schema:
+                sa_type = schema['metadata'].get('sa_type')
+                if sa_type is not None:
+                    return sa_type
+        except (TypeError, AttributeError, KeyError, ValueError):
+            # 类型本身的schema获取失败
+            # 这是正常的fallback机制，annotation可能不支持此协议
+            pass
+
+    return None
+
+
+def _resolve_annotations(attrs: dict[str, Any]) -> tuple[
+    dict[str, Any],
+    dict[str, str],
+    Mapping[str, Any],
+    Mapping[str, Any],
+]:
+    """
+    Resolve annotations from a class namespace with Python 3.14 (PEP 649) support.
+
+    This helper prefers evaluated annotations (Format.VALUE) so that `typing.Annotated`
+    metadata and custom types remain accessible. Forward references that cannot be
+    evaluated are replaced with typing.ForwardRef placeholders to avoid aborting the
+    whole resolution process.
+    """
+    raw_annotations = attrs.get('__annotations__') or {}
+    try:
+        base_annotations = dict(raw_annotations)
+    except TypeError:
+        base_annotations = {}
+
+    module_name = attrs.get('__module__')
+    module_globals: dict[str, Any]
+    if module_name and module_name in sys.modules:
+        module_globals = dict(sys.modules[module_name].__dict__)
+    else:
+        module_globals = {}
+
+    module_globals.setdefault('__builtins__', __builtins__)
+    localns: dict[str, Any] = dict(attrs)
+
+    try:
+        temp_cls = type('AnnotationProxy', (object,), dict(attrs))
+        temp_cls.__module__ = module_name
+        extras_kw = {'include_extras': True} if sys.version_info >= (3, 10) else {}
+        evaluated = get_type_hints(
+            temp_cls,
+            globalns=module_globals,
+            localns=localns,
+            **extras_kw,
+        )
+    except (NameError, AttributeError, TypeError, RecursionError):
+        # get_type_hints可能失败的原因：
+        # - NameError: 前向引用无法解析（类型尚未定义）
+        # - AttributeError: 模块或类型不存在
+        # - TypeError: 无效的类型注解
+        # - RecursionError: 循环依赖的类型定义
+        # 这是正常情况，回退到原始注解字符串
+        evaluated = base_annotations
+
+    return dict(evaluated), {}, module_globals, localns
+
+
+def _evaluate_annotation_from_string(
+    field_name: str,
+    annotation_strings: dict[str, str],
+    current_type: Any,
+    globalns: Mapping[str, Any],
+    localns: Mapping[str, Any],
+) -> Any:
+    """
+    Attempt to re-evaluate the original annotation string for a field.
+
+    This is used as a fallback when the resolved annotation lost its metadata
+    (e.g., Annotated wrappers) and we need to recover custom sa_type data.
+    """
+    if not annotation_strings:
+        return current_type
+
+    expr = annotation_strings.get(field_name)
+    if not expr or not isinstance(expr, str):
+        return current_type
+
+    try:
+        return eval(expr, globalns, localns)
+    except (NameError, SyntaxError, AttributeError, TypeError):
+        # eval可能失败的原因：
+        # - NameError: 类型名称在namespace中不存在
+        # - SyntaxError: 注解字符串有语法错误
+        # - AttributeError: 访问不存在的模块属性
+        # - TypeError: 无效的类型表达式
+        # 这是正常的fallback机制，返回当前已解析的类型
+        return current_type
+
+
+class __DeclarativeMeta(SQLModelMetaclass):
+    """
+    一个智能的混合模式元类，它提供了灵活性和清晰度：
+
+    1.  **自动设置 `table=True`**: 如果一个类继承了 `TableBaseMixin`，则自动应用 `table=True`。
+    2.  **明确的字典参数**: 支持 `mapper_args={...}`, `table_args={...}`, `table_name='...'`。
+    3.  **便捷的关键字参数**: 支持最常见的 mapper 参数作为顶级关键字（如 `polymorphic_on`）。
+    4.  **智能合并**: 当字典和关键字同时提供时，会自动合并，且关键字参数有更高优先级。
+    """
+
+    _KNOWN_MAPPER_KEYS = {
+        "polymorphic_on",
+        "polymorphic_identity",
+        "polymorphic_abstract",
+        "version_id_col",
+        "concrete",
+    }
+
+    def __new__(cls, name, bases, attrs, **kwargs):
+        # 1. 约定优于配置：自动设置 table=True
+        is_intended_as_table = any(getattr(b, '_has_table_mixin', False) for b in bases)
+        if is_intended_as_table and 'table' not in kwargs:
+            kwargs['table'] = True
+
+        # 2. 智能合并 __mapper_args__
+        collected_mapper_args = {}
+
+        # 首先，处理明确的 mapper_args 字典 (优先级较低)
+        if 'mapper_args' in kwargs:
+            collected_mapper_args.update(kwargs.pop('mapper_args'))
+
+        # 其次，处理便捷的关键字参数 (优先级更高)
+        for key in cls._KNOWN_MAPPER_KEYS:
+            if key in kwargs:
+                # .pop() 获取值并移除，避免传递给父类
+                collected_mapper_args[key] = kwargs.pop(key)
+
+        # 如果收集到了任何 mapper 参数，则更新到类的属性中
+        if collected_mapper_args:
+            existing = attrs.get('__mapper_args__', {}).copy()
+            existing.update(collected_mapper_args)
+            attrs['__mapper_args__'] = existing
+
+        # 3. 处理其他明确的参数
+        if 'table_args' in kwargs:
+            attrs['__table_args__'] = kwargs.pop('table_args')
+        if 'table_name' in kwargs:
+            attrs['__tablename__'] = kwargs.pop('table_name')
+        if 'abstract' in kwargs:
+            attrs['__abstract__'] = kwargs.pop('abstract')
+
+        # 4. 从Annotated元数据中提取sa_type并注入到Field
+        # 重要：必须在调用父类__new__之前处理，因为SQLModel会消费annotations
+        #
+        # Python 3.14兼容性问题：
+        # - SQLModel在Python 3.14上会因为ClassVar[T]类型而崩溃（issubclass错误）
+        # - 我们必须在SQLModel看到annotations之前过滤掉ClassVar字段
+        # - 虽然PEP 749建议不修改__annotations__，但这是修复SQLModel bug的必要措施
+        #
+        # 获取annotations的策略：
+        # - Python 3.14+: 优先从__annotate__获取（如果存在）
+        # - fallback: 从__annotations__读取（如果存在）
+        # - 最终fallback: 空字典
+        annotations, annotation_strings, eval_globals, eval_locals = _resolve_annotations(attrs)
+
+        if annotations:
+            attrs['__annotations__'] = annotations
+            if annotationlib is not None:
+                # 在Python 3.14中禁用descriptor，转为普通dict
+                attrs['__annotate__'] = None
+
+        for field_name, field_type in annotations.items():
+            field_type = _evaluate_annotation_from_string(
+                field_name,
+                annotation_strings,
+                field_type,
+                eval_globals,
+                eval_locals,
+            )
+
+            # 跳过字符串或ForwardRef类型注解，让SQLModel自己处理
+            if isinstance(field_type, str) or isinstance(field_type, typing.ForwardRef):
+                continue
+
+            # 跳过特殊类型的字段
+            origin = get_origin(field_type)
+
+            # 跳过 ClassVar 字段 - 它们不是数据库字段
+            if origin is typing.ClassVar:
+                continue
+
+            # 跳过 Mapped 字段 - SQLAlchemy 2.0+ 的声明式字段，已经有 mapped_column
+            if origin is Mapped:
+                continue
+
+            # 尝试从注解中提取sa_type
+            sa_type = _extract_sa_type_from_annotation(field_type)
+
+            if sa_type is not None:
+                # 检查字段是否已有Field定义
+                field_value = attrs.get(field_name, Undefined)
+
+                if field_value is Undefined:
+                    # 没有Field定义，创建一个新的Field并注入sa_type
+                    attrs[field_name] = Field(sa_type=sa_type)
+                elif isinstance(field_value, FieldInfo):
+                    # 已有Field定义，检查是否已设置sa_type
+                    # 注意：只有在未设置时才注入，尊重显式配置
+                    # SQLModel使用Undefined作为"未设置"的标记
+                    if not hasattr(field_value, 'sa_type') or field_value.sa_type is Undefined:
+                        field_value.sa_type = sa_type
+                # 如果field_value是其他类型（如默认值），不处理
+                # SQLModel会在后续处理中将其转换为Field
+
+        # 5. 调用父类的 __new__ 方法，传入被清理过的 kwargs
+        result = super().__new__(cls, name, bases, attrs, **kwargs)
+
+        # 6. 修复：在联表继承场景下，继承父类的 __sqlmodel_relationships__
+        # SQLModel 为每个 table=True 的类创建新的空 __sqlmodel_relationships__
+        # 这导致子类丢失父类的关系定义，触发错误的 Column 创建
+        # 必须在 super().__new__() 之后修复，因为 SQLModel 会覆盖我们预设的值
+        if kwargs.get('table', False):
+            for base in bases:
+                if hasattr(base, '__sqlmodel_relationships__'):
+                    for rel_name, rel_info in base.__sqlmodel_relationships__.items():
+                        # 只继承子类没有重新定义的关系
+                        if rel_name not in result.__sqlmodel_relationships__:
+                            result.__sqlmodel_relationships__[rel_name] = rel_info
+                            # 同时修复被错误创建的 Column - 恢复为父类的 relationship
+                            if hasattr(base, rel_name):
+                                base_attr = getattr(base, rel_name)
+                                setattr(result, rel_name, base_attr)
+
+        # 7. 检测：禁止子类重定义父类的 Relationship 字段
+        # 子类重定义同名的 Relationship 字段会导致 SQLAlchemy 关系映射混乱，
+        # 应该在类定义时立即报错，而不是在运行时出现难以调试的问题。
+        for base in bases:
+            parent_relationships = getattr(base, '__sqlmodel_relationships__', {})
+            for rel_name in parent_relationships:
+                # 检查当前类是否在 attrs 中重新定义了这个关系字段
+                if rel_name in attrs:
+                    raise TypeError(
+                        f"类 {name} 不允许重定义父类 {base.__name__} 的 Relationship 字段 '{rel_name}'。"
+                        f"如需修改关系配置，请在父类中修改。"
+                    )
+
+        # 8. 修复：从 model_fields/__pydantic_fields__ 中移除 Relationship 字段
+        # SQLModel 0.0.27 bug：子类会错误地继承父类的 Relationship 字段到 model_fields
+        # 这导致 Pydantic 尝试为 Relationship 字段生成 schema，因为类型是
+        # Mapped[list['Character']] 这种前向引用，Pydantic 无法解析，
+        # 导致 __pydantic_complete__ = False
+        #
+        # 修复策略：
+        # - 检查类的 __sqlmodel_relationships__ 属性
+        # - 从 model_fields 和 __pydantic_fields__ 中移除这些字段
+        # - Relationship 字段由 SQLAlchemy 管理，不需要 Pydantic 参与
+        relationships = getattr(result, '__sqlmodel_relationships__', {})
+        if relationships:
+            model_fields = getattr(result, 'model_fields', {})
+            pydantic_fields = getattr(result, '__pydantic_fields__', {})
+
+            fields_removed = False
+            for rel_name in relationships:
+                if rel_name in model_fields:
+                    del model_fields[rel_name]
+                    fields_removed = True
+                if rel_name in pydantic_fields:
+                    del pydantic_fields[rel_name]
+                    fields_removed = True
+
+            # 如果移除了字段，重新构建 Pydantic 模式
+            # 注意：只在有字段被移除时才 rebuild，避免不必要的开销
+            if fields_removed and hasattr(result, 'model_rebuild'):
+                result.model_rebuild(force=True)
+
+        return result
+
+    def __init__(
+        cls,
+        classname: str,
+        bases: tuple[type, ...],
+        dict_: dict[str, typing.Any],
+        **kw: typing.Any,
+    ) -> None:
+        """
+        重写 SQLModel 的 __init__ 以支持联表继承（Joined Table Inheritance）
+
+        SQLModel 原始行为：
+        - 如果任何基类是表模型，则不调用 DeclarativeMeta.__init__
+        - 这阻止了子类创建自己的表
+
+        修复逻辑：
+        - 检测联表继承场景（子类有自己的 __tablename__ 且有外键指向父表）
+        - 强制调用 DeclarativeMeta.__init__ 来创建子表
+        """
+        from sqlmodel.main import is_table_model_class, DeclarativeMeta, ModelMetaclass
+
+        # 检查是否是表模型
+        if not is_table_model_class(cls):
+            ModelMetaclass.__init__(cls, classname, bases, dict_, **kw)
+            return
+
+        # 检查是否有基类是表模型
+        base_is_table = any(is_table_model_class(base) for base in bases)
+
+        if not base_is_table:
+            # 没有基类是表模型，走正常的 SQLModel 流程
+            # 处理关系字段
+            cls._setup_relationships()
+            DeclarativeMeta.__init__(cls, classname, bases, dict_, **kw)
+            return
+
+        # 关键：检测联表继承场景
+        # 条件：
+        # 1. 当前类的 __tablename__ 与父类不同（表示需要新表）
+        # 2. 当前类有字段带有 foreign_key 指向父表
+        current_tablename = getattr(cls, '__tablename__', None)
+
+        # 查找父表信息
+        parent_table = None
+        parent_tablename = None
+        for base in bases:
+            if is_table_model_class(base) and hasattr(base, '__tablename__'):
+                parent_tablename = base.__tablename__
+                break
+
+        # 检查是否有不同的 tablename
+        has_different_tablename = (
+            current_tablename is not None
+            and parent_tablename is not None
+            and current_tablename != parent_tablename
+        )
+
+        # 检查是否有外键字段指向父表的主键
+        # 注意：由于字段合并，我们需要检查直接基类的 model_fields
+        # 而不是当前类的合并后的 model_fields
+        has_fk_to_parent = False
+
+        def _normalize_tablename(name: str) -> str:
+            """标准化表名以进行比较（移除下划线，转小写）"""
+            return name.replace('_', '').lower()
+
+        def _fk_matches_parent(fk_str: str, parent_table: str) -> bool:
+            """检查 FK 字符串是否指向父表"""
+            if not fk_str or not parent_table:
+                return False
+            # FK 格式: "tablename.column" 或 "schema.tablename.column"
+            parts = fk_str.split('.')
+            if len(parts) >= 2:
+                fk_table = parts[-2]  # 取倒数第二个作为表名
+                # 标准化比较（处理下划线差异）
+                return _normalize_tablename(fk_table) == _normalize_tablename(parent_table)
+            return False
+
+        if has_different_tablename and parent_tablename:
+            # 首先检查当前类的 model_fields
+            for field_name, field_info in cls.model_fields.items():
+                fk = getattr(field_info, 'foreign_key', None)
+                if fk is not None and isinstance(fk, str) and _fk_matches_parent(fk, parent_tablename):
+                    has_fk_to_parent = True
+                    break
+
+            # 如果没找到，检查直接基类的 model_fields（解决 mixin 字段被覆盖的问题）
+            if not has_fk_to_parent:
+                for base in bases:
+                    if hasattr(base, 'model_fields'):
+                        for field_name, field_info in base.model_fields.items():
+                            fk = getattr(field_info, 'foreign_key', None)
+                            if fk is not None and isinstance(fk, str) and _fk_matches_parent(fk, parent_tablename):
+                                has_fk_to_parent = True
+                                break
+                    if has_fk_to_parent:
+                        break
+
+        is_joined_inheritance = has_different_tablename and has_fk_to_parent
+
+        if is_joined_inheritance:
+            # 联表继承：需要创建子表
+
+            # 修复外键字段：由于字段合并，外键信息可能丢失
+            # 需要从基类的 mixin 中找回外键信息，并重建列
+            from sqlalchemy import Column, ForeignKey, inspect as sa_inspect
+            from sqlalchemy.dialects.postgresql import UUID as SA_UUID
+            from sqlalchemy.exc import NoInspectionAvailable
+            from sqlalchemy.orm.attributes import InstrumentedAttribute
+
+            # 联表继承：子表只应该有 id（FK 到父表）+ 子类特有的字段
+            # 所有继承自祖先表的列都不应该在子表中重复创建
+
+            # 收集整个继承链中所有祖先表的列名（这些列不应该在子表中重复）
+            # 需要遍历整个 MRO，因为可能是多级继承（如 Tool -> Function -> GetWeatherFunction）
+            ancestor_column_names: set[str] = set()
+            for ancestor in cls.__mro__:
+                if ancestor is cls:
+                    continue  # 跳过当前类
+                if is_table_model_class(ancestor):
+                    try:
+                        # 使用 inspect() 获取 mapper 的公开属性
+                        # 源码确认: mapper.local_table 是公开属性 (mapper.py:979-998)
+                        mapper = sa_inspect(ancestor)
+                        for col in mapper.local_table.columns:
+                            # 跳过 _polymorphic_name 列（鉴别器，由根父表管理）
+                            if col.name.startswith('_polymorphic'):
+                                continue
+                            ancestor_column_names.add(col.name)
+                    except NoInspectionAvailable:
+                        continue
+
+            # 找到子类自己定义的字段（不在父类中的）
+            child_own_fields: set[str] = set()
+            for field_name in cls.model_fields:
+                # 检查这个字段是否是在当前类直接定义的（不是继承的）
+                # 通过检查父类是否有这个字段来判断
+                is_inherited = False
+                for base in bases:
+                    if hasattr(base, 'model_fields') and field_name in base.model_fields:
+                        is_inherited = True
+                        break
+                if not is_inherited:
+                    child_own_fields.add(field_name)
+
+            # 从子类类属性中移除父表已有的列定义
+            # 这样 SQLAlchemy 就不会在子表中创建这些列
+            fk_field_name = None
+            for base in bases:
+                if hasattr(base, 'model_fields'):
+                    for field_name, field_info in base.model_fields.items():
+                        fk = getattr(field_info, 'foreign_key', None)
+                        pk = getattr(field_info, 'primary_key', False)
+                        if fk is not None and isinstance(fk, str) and _fk_matches_parent(fk, parent_tablename):
+                            fk_field_name = field_name
+                            # 找到了外键字段，重建它
+                            # 创建一个新的 Column 对象包含外键约束
+                            new_col = Column(
+                                field_name,
+                                SA_UUID(as_uuid=True),
+                                ForeignKey(fk),
+                                primary_key=pk if pk else False
+                            )
+                            setattr(cls, field_name, new_col)
+                            break
+                    else:
+                        continue
+                    break
+
+            # 移除继承自祖先表的列属性（除了 FK/PK 和子类自己的字段）
+            # 这防止 SQLAlchemy 在子表中创建重复列
+            # 注意：在 __init__ 阶段，列是 Column 对象，不是 InstrumentedAttribute
+            for col_name in ancestor_column_names:
+                if col_name == fk_field_name:
+                    continue  # 保留 FK/PK 列（子表的主键，同时是父表的外键）
+                if col_name == 'id':
+                    continue  # id 会被 FK 字段覆盖
+                if col_name in child_own_fields:
+                    continue  # 保留子类自己定义的字段
+
+                # 检查类属性是否是 Column 或 InstrumentedAttribute
+                if col_name in cls.__dict__:
+                    attr = cls.__dict__[col_name]
+                    # Column 对象或 InstrumentedAttribute 都需要删除
+                    if isinstance(attr, (Column, InstrumentedAttribute)):
+                        try:
+                            delattr(cls, col_name)
+                        except AttributeError:
+                            pass
+
+            # 找到子类自己定义的关系（不在父类中的）
+            # 继承的关系会从父类自动获取，只需要设置子类新增的关系
+            child_own_relationships: set[str] = set()
+            for rel_name in cls.__sqlmodel_relationships__:
+                is_inherited = False
+                for base in bases:
+                    if hasattr(base, '__sqlmodel_relationships__') and rel_name in base.__sqlmodel_relationships__:
+                        is_inherited = True
+                        break
+                if not is_inherited:
+                    child_own_relationships.add(rel_name)
+
+            # 只为子类自己定义的新关系调用关系设置
+            if child_own_relationships:
+                cls._setup_relationships(only_these=child_own_relationships)
+
+            # 强制调用 DeclarativeMeta.__init__
+            DeclarativeMeta.__init__(cls, classname, bases, dict_, **kw)
+        else:
+            # 非联表继承：单表继承或正常 Pydantic 模型
+            ModelMetaclass.__init__(cls, classname, bases, dict_, **kw)
+
+    def _setup_relationships(cls, only_these: set[str] | None = None) -> None:
+        """
+        设置 SQLAlchemy 关系字段（从 SQLModel 源码复制）
+
+        Args:
+            only_these: 如果提供，只设置这些关系（用于 joined table inheritance 子类）
+                       如果为 None，设置所有关系（默认行为）
+        """
+        from sqlalchemy.orm import relationship, Mapped
+        from sqlalchemy import inspect
+        from sqlmodel.main import get_relationship_to
+        from typing import get_origin
+
+        for rel_name, rel_info in cls.__sqlmodel_relationships__.items():
+            # 如果指定了 only_these，只设置这些关系
+            if only_these is not None and rel_name not in only_these:
+                continue
+            if rel_info.sa_relationship:
+                setattr(cls, rel_name, rel_info.sa_relationship)
+                continue
+
+            raw_ann = cls.__annotations__[rel_name]
+            origin: typing.Any = get_origin(raw_ann)
+            if origin is Mapped:
+                ann = raw_ann.__args__[0]
+            else:
+                ann = raw_ann
+                cls.__annotations__[rel_name] = Mapped[ann]
+
+            relationship_to = get_relationship_to(
+                name=rel_name, rel_info=rel_info, annotation=ann
+            )
+            rel_kwargs: dict[str, typing.Any] = {}
+            if rel_info.back_populates:
+                rel_kwargs["back_populates"] = rel_info.back_populates
+            if rel_info.cascade_delete:
+                rel_kwargs["cascade"] = "all, delete-orphan"
+            if rel_info.passive_deletes:
+                rel_kwargs["passive_deletes"] = rel_info.passive_deletes
+            if rel_info.link_model:
+                ins = inspect(rel_info.link_model)
+                local_table = getattr(ins, "local_table")
+                if local_table is None:
+                    raise RuntimeError(
+                        f"Couldn't find secondary table for {rel_info.link_model}"
+                    )
+                rel_kwargs["secondary"] = local_table
+
+            rel_args: list[typing.Any] = []
+            if rel_info.sa_relationship_args:
+                rel_args.extend(rel_info.sa_relationship_args)
+            if rel_info.sa_relationship_kwargs:
+                rel_kwargs.update(rel_info.sa_relationship_kwargs)
+
+            rel_value = relationship(relationship_to, *rel_args, **rel_kwargs)
+            setattr(cls, rel_name, rel_value)
+
+
+class SQLModelBase(SQLModel, metaclass=__DeclarativeMeta):
+    """此类必须和TableBase系列类搭配使用"""
+
+    model_config = ConfigDict(use_attribute_docstrings=True, validate_by_name=True)