springbootAI 2.3.8__tar.gz → 2.3.9__tar.gz

{springbootai-2.3.8 → springbootai-2.3.9}/CHANGELOG.md

@@ -1,5 +1,14 @@
 # Changelog
 
+## [2.3.9] - 2026-08-28
+
+- 加固 OpenAI 兼容接口与 Ollama 的同步/异步流式协议：校验终止帧、错误信封和畸形响应，限制事件与响应大小，并完善超时、有限重试、重定向拒绝及 SDK 重试边界。
+- 修复 AI 异步流阻塞事件循环、请求参数覆盖受保护字段、输出 Token 上限绕过、工具调用闭环和多租户 Memory/RAG 隔离等问题。
+- 加固 Feign、Gateway、Config Center、Admin 与 Discovery：正确释放响应、执行路由谓词和过滤器、过滤逐跳头、传播请求 ID，并增加路径、响应大小、故障回退与生命周期保护。
+- 新增统一日志脱敏和日志注入防护，覆盖异常链、Bearer/Basic/Cookie、云事件、Nacos、Seata 与 SkyWalking，避免敏感负载进入日志或健康响应。
+- 修复原子脚手架写入、TextReader 文件边界、DevTools 停止/重启语义、SkyWalking 显式初始化和 Banner 版本读取错误。
+- 新增企业加固回归测试；完整测试结果为 3170 passed、11 skipped、0 failed，Bandit 中高危扫描为 0。
+
 ## [2.3.8] - 2026-08-24
 
 - 修复配置加载的原子性与数据库 URL 优先级：请求监控不再误落到根目录 `test.db`，SQLite 运行目录会自动创建，显式 URL/驱动/数据库名称按统一优先级生效。

{springbootai-2.3.8/springbootAI.egg-info → springbootai-2.3.9}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: springbootAI
-Version: 2.3.8
+Version: 2.3.9
 Summary: Spring-style Python framework with Web, ORM, Cloud and AI integrations
 Author-email: YuConggen <1516933915@qq.com>
 License-Expression: MIT
@@ -216,7 +216,7 @@ SpringBootAI 是一个采用 Spring 风格注解和分层结构的 Python 应用
 | AI 与编排 | 模型调用、Tools、RAG、Chain、状态图、MCP client/server | LangChain / LangGraph / 官方 MCP SDK |
 | 生产治理 | 健康检查、Prometheus、限流熔断、追踪、Swagger | prometheus-client / OpenTelemetry / OpenAPI |
 
-当前版本是 `2.3.8`；支持 Python 3.10、3.11 和 3.12，许可证为 MIT。项目仍标记为 Beta。用于公网高并发、合规敏感或支付/订单/库存等核心系统前，必须完成目标数据库、流量模型、故障恢复和安全基线验证。内嵌 Gateway 适合内部路由，不替代公网 Nginx/Kong/WAF；Seata `distributed` 对接官方 TC + TCC 回调；`at` 模式通过 ORM 拦截器自动生成 undo_log 实现自动回滚。
+当前版本是 `2.3.9`；支持 Python 3.10、3.11 和 3.12，许可证为 MIT。项目仍标记为 Beta。用于公网高并发、合规敏感或支付/订单/库存等核心系统前，必须完成目标数据库、流量模型、故障恢复和安全基线验证。内嵌 Gateway 适合内部路由，不替代公网 Nginx/Kong/WAF；Seata `distributed` 对接官方 TC + TCC 回调；`at` 模式通过 ORM 拦截器自动生成 undo_log 实现自动回滚。
 
 [新手指南](https://github.com/YUCONGGEN/springbootAI/blob/master/doc/BEGINNER_GUIDE.md) | [全部文档](https://github.com/YUCONGGEN/springbootAI/tree/master/doc) | [变更日志](https://github.com/YUCONGGEN/springbootAI/blob/master/CHANGELOG.md) | [安全报告](https://github.com/YUCONGGEN/springbootAI/blob/master/SECURITY.md) | [发布检查](https://github.com/YUCONGGEN/springbootAI/blob/master/doc/RELEASE_CHECKLIST.md)
 
@@ -347,10 +347,10 @@ SpringBootAI 是一个 **Python Web 框架**。它把 Java Spring Boot 的"注
 
 | 组件 | 当前版本 |
 |------|----------|
-| `springbootai` 框架 API | 2.3.8 |
-| `springbootai.orm.pymybatis` | 2.3.8 |
-| `springbootai.ai` AI 模块 | 2.3.8 |
-| `springbootai.langchain` LangChain 模块 | 2.3.8 |
+| `springbootai` 框架 API | 2.3.9 |
+| `springbootai.orm.pymybatis` | 2.3.9 |
+| `springbootai.ai` AI 模块 | 2.3.9 |
+| `springbootai.langchain` LangChain 模块 | 2.3.9 |
 | Python | 3.10+ |
 
 ### 1.4 适合什么场景
@@ -937,6 +937,16 @@ class InitService:
             self.connection.close()
 ```
 
+每个 HTTP 请求都会获得 `X-Request-ID`（可通过
+`server.request-id.header` 修改），该值会写回响应并自动进入框架日志，Feign
+和 Gateway 也会向下游透传。日志会遮蔽 Bearer Token、API Key、密码、Cookie
+等常见敏感字段；`logging.diagnose` 默认关闭，避免异常帧局部变量泄密。
+
+声明式 Feign 客户端默认复用连接池，对幂等请求执行有限退避重试，并支持连接/
+读取超时与响应大小限制。直接读取 RAG 文本文件时建议使用
+`TextReader.from_file(path, max_bytes=...)`，文件不存在或超限会明确失败，不会把
+错误路径误当成内联文本。
+
 ### 5.7 应用事件
 
 ```python
@@ -1895,9 +1905,13 @@ export LOG_DIR=logs
 
 # AI 模块
 export AI_PROVIDER=openai
-export AI_ALLOW_FAKE=true
+export AI_ALLOW_FAKE=false
 export OPENAI_API_KEY=sk-xxx
 export OPENAI_CHAT_MODEL=gpt-4o-mini
+export AI_REQUEST_TIMEOUT_SECONDS=60
+export AI_MAX_OUTPUT_TOKENS=4096
+export AI_MAX_TOTAL_TOKENS=100000
+export AI_MAX_TOOL_ITERATIONS=5
 export OLLAMA_BASE_URL=http://localhost:11434
 export OLLAMA_CHAT_MODEL=llama3
 

{springbootai-2.3.8 → springbootai-2.3.9}/README.md

@@ -54,7 +54,7 @@ SpringBootAI 是一个采用 Spring 风格注解和分层结构的 Python 应用
 | AI 与编排 | 模型调用、Tools、RAG、Chain、状态图、MCP client/server | LangChain / LangGraph / 官方 MCP SDK |
 | 生产治理 | 健康检查、Prometheus、限流熔断、追踪、Swagger | prometheus-client / OpenTelemetry / OpenAPI |
 
-当前版本是 `2.3.8`；支持 Python 3.10、3.11 和 3.12，许可证为 MIT。项目仍标记为 Beta。用于公网高并发、合规敏感或支付/订单/库存等核心系统前，必须完成目标数据库、流量模型、故障恢复和安全基线验证。内嵌 Gateway 适合内部路由，不替代公网 Nginx/Kong/WAF；Seata `distributed` 对接官方 TC + TCC 回调；`at` 模式通过 ORM 拦截器自动生成 undo_log 实现自动回滚。
+当前版本是 `2.3.9`；支持 Python 3.10、3.11 和 3.12，许可证为 MIT。项目仍标记为 Beta。用于公网高并发、合规敏感或支付/订单/库存等核心系统前，必须完成目标数据库、流量模型、故障恢复和安全基线验证。内嵌 Gateway 适合内部路由，不替代公网 Nginx/Kong/WAF；Seata `distributed` 对接官方 TC + TCC 回调；`at` 模式通过 ORM 拦截器自动生成 undo_log 实现自动回滚。
 
 [新手指南](https://github.com/YUCONGGEN/springbootAI/blob/master/doc/BEGINNER_GUIDE.md) | [全部文档](https://github.com/YUCONGGEN/springbootAI/tree/master/doc) | [变更日志](https://github.com/YUCONGGEN/springbootAI/blob/master/CHANGELOG.md) | [安全报告](https://github.com/YUCONGGEN/springbootAI/blob/master/SECURITY.md) | [发布检查](https://github.com/YUCONGGEN/springbootAI/blob/master/doc/RELEASE_CHECKLIST.md)
 
@@ -185,10 +185,10 @@ SpringBootAI 是一个 **Python Web 框架**。它把 Java Spring Boot 的"注
 
 | 组件 | 当前版本 |
 |------|----------|
-| `springbootai` 框架 API | 2.3.8 |
-| `springbootai.orm.pymybatis` | 2.3.8 |
-| `springbootai.ai` AI 模块 | 2.3.8 |
-| `springbootai.langchain` LangChain 模块 | 2.3.8 |
+| `springbootai` 框架 API | 2.3.9 |
+| `springbootai.orm.pymybatis` | 2.3.9 |
+| `springbootai.ai` AI 模块 | 2.3.9 |
+| `springbootai.langchain` LangChain 模块 | 2.3.9 |
 | Python | 3.10+ |
 
 ### 1.4 适合什么场景
@@ -775,6 +775,16 @@ class InitService:
             self.connection.close()
 ```
 
+每个 HTTP 请求都会获得 `X-Request-ID`（可通过
+`server.request-id.header` 修改），该值会写回响应并自动进入框架日志，Feign
+和 Gateway 也会向下游透传。日志会遮蔽 Bearer Token、API Key、密码、Cookie
+等常见敏感字段；`logging.diagnose` 默认关闭，避免异常帧局部变量泄密。
+
+声明式 Feign 客户端默认复用连接池，对幂等请求执行有限退避重试，并支持连接/
+读取超时与响应大小限制。直接读取 RAG 文本文件时建议使用
+`TextReader.from_file(path, max_bytes=...)`，文件不存在或超限会明确失败，不会把
+错误路径误当成内联文本。
+
 ### 5.7 应用事件
 
 ```python
@@ -1733,9 +1743,13 @@ export LOG_DIR=logs
 
 # AI 模块
 export AI_PROVIDER=openai
-export AI_ALLOW_FAKE=true
+export AI_ALLOW_FAKE=false
 export OPENAI_API_KEY=sk-xxx
 export OPENAI_CHAT_MODEL=gpt-4o-mini
+export AI_REQUEST_TIMEOUT_SECONDS=60
+export AI_MAX_OUTPUT_TOKENS=4096
+export AI_MAX_TOTAL_TOKENS=100000
+export AI_MAX_TOOL_ITERATIONS=5
 export OLLAMA_BASE_URL=http://localhost:11434
 export OLLAMA_CHAT_MODEL=llama3
 

{springbootai-2.3.8 → springbootai-2.3.9}/pyproject.toml

@@ -12,7 +12,7 @@ build-backend = "setuptools.build_meta"
 [project]
 # Distribution name: springbootAI (pip install springbootAI); import package: springbootai.
 name = "springbootAI"
-version = "2.3.8"
+version = "2.3.9"
 description = "Spring-style Python framework with Web, ORM, Cloud and AI integrations"
 readme = "README.md"
 license = "MIT"
@@ -256,6 +256,7 @@ full = [
 
 [tool.pytest.ini_options]
 pythonpath = [".", "examples"]
+testpaths = ["tests"]
 filterwarnings = [
     # 闈欓粯 langchain_classic 寮冪敤鍛婅鍜?community sunset 鍛婅
     # 鍘熷洜锛氭湰椤圭洰鏈夋剰閿佸畾 langchain classic (1.x) API锛岃繖浜涙槸棰勬湡鍐呯殑妗嗘灦绾у憡璀?

{springbootai-2.3.8 → springbootai-2.3.9/springbootAI.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: springbootAI
-Version: 2.3.8
+Version: 2.3.9
 Summary: Spring-style Python framework with Web, ORM, Cloud and AI integrations
 Author-email: YuConggen <1516933915@qq.com>
 License-Expression: MIT
@@ -216,7 +216,7 @@ SpringBootAI 是一个采用 Spring 风格注解和分层结构的 Python 应用
 | AI 与编排 | 模型调用、Tools、RAG、Chain、状态图、MCP client/server | LangChain / LangGraph / 官方 MCP SDK |
 | 生产治理 | 健康检查、Prometheus、限流熔断、追踪、Swagger | prometheus-client / OpenTelemetry / OpenAPI |
 
-当前版本是 `2.3.8`；支持 Python 3.10、3.11 和 3.12，许可证为 MIT。项目仍标记为 Beta。用于公网高并发、合规敏感或支付/订单/库存等核心系统前，必须完成目标数据库、流量模型、故障恢复和安全基线验证。内嵌 Gateway 适合内部路由，不替代公网 Nginx/Kong/WAF；Seata `distributed` 对接官方 TC + TCC 回调；`at` 模式通过 ORM 拦截器自动生成 undo_log 实现自动回滚。
+当前版本是 `2.3.9`；支持 Python 3.10、3.11 和 3.12，许可证为 MIT。项目仍标记为 Beta。用于公网高并发、合规敏感或支付/订单/库存等核心系统前，必须完成目标数据库、流量模型、故障恢复和安全基线验证。内嵌 Gateway 适合内部路由，不替代公网 Nginx/Kong/WAF；Seata `distributed` 对接官方 TC + TCC 回调；`at` 模式通过 ORM 拦截器自动生成 undo_log 实现自动回滚。
 
 [新手指南](https://github.com/YUCONGGEN/springbootAI/blob/master/doc/BEGINNER_GUIDE.md) | [全部文档](https://github.com/YUCONGGEN/springbootAI/tree/master/doc) | [变更日志](https://github.com/YUCONGGEN/springbootAI/blob/master/CHANGELOG.md) | [安全报告](https://github.com/YUCONGGEN/springbootAI/blob/master/SECURITY.md) | [发布检查](https://github.com/YUCONGGEN/springbootAI/blob/master/doc/RELEASE_CHECKLIST.md)
 
@@ -347,10 +347,10 @@ SpringBootAI 是一个 **Python Web 框架**。它把 Java Spring Boot 的"注
 
 | 组件 | 当前版本 |
 |------|----------|
-| `springbootai` 框架 API | 2.3.8 |
-| `springbootai.orm.pymybatis` | 2.3.8 |
-| `springbootai.ai` AI 模块 | 2.3.8 |
-| `springbootai.langchain` LangChain 模块 | 2.3.8 |
+| `springbootai` 框架 API | 2.3.9 |
+| `springbootai.orm.pymybatis` | 2.3.9 |
+| `springbootai.ai` AI 模块 | 2.3.9 |
+| `springbootai.langchain` LangChain 模块 | 2.3.9 |
 | Python | 3.10+ |
 
 ### 1.4 适合什么场景
@@ -937,6 +937,16 @@ class InitService:
             self.connection.close()
 ```
 
+每个 HTTP 请求都会获得 `X-Request-ID`（可通过
+`server.request-id.header` 修改），该值会写回响应并自动进入框架日志，Feign
+和 Gateway 也会向下游透传。日志会遮蔽 Bearer Token、API Key、密码、Cookie
+等常见敏感字段；`logging.diagnose` 默认关闭，避免异常帧局部变量泄密。
+
+声明式 Feign 客户端默认复用连接池，对幂等请求执行有限退避重试，并支持连接/
+读取超时与响应大小限制。直接读取 RAG 文本文件时建议使用
+`TextReader.from_file(path, max_bytes=...)`，文件不存在或超限会明确失败，不会把
+错误路径误当成内联文本。
+
 ### 5.7 应用事件
 
 ```python
@@ -1895,9 +1905,13 @@ export LOG_DIR=logs
 
 # AI 模块
 export AI_PROVIDER=openai
-export AI_ALLOW_FAKE=true
+export AI_ALLOW_FAKE=false
 export OPENAI_API_KEY=sk-xxx
 export OPENAI_CHAT_MODEL=gpt-4o-mini
+export AI_REQUEST_TIMEOUT_SECONDS=60
+export AI_MAX_OUTPUT_TOKENS=4096
+export AI_MAX_TOTAL_TOKENS=100000
+export AI_MAX_TOOL_ITERATIONS=5
 export OLLAMA_BASE_URL=http://localhost:11434
 export OLLAMA_CHAT_MODEL=llama3
 

{springbootai-2.3.8 → springbootai-2.3.9}/springbootAI.egg-info/SOURCES.txt

@@ -157,6 +157,7 @@ springbootai/langgraph/checkpoint.py
 springbootai/langgraph/config.py
 springbootai/langgraph/runtime.py
 springbootai/logging/__init__.py
+springbootai/logging/context.py
 springbootai/logging/loguru_logger.py
 springbootai/mcp/__init__.py
 springbootai/mcp/annotation_runtime.py
@@ -266,7 +267,9 @@ tests/test_actuator_prometheus_admin.py
 tests/test_ai_annotations.py
 tests/test_ai_comprehensive.py
 tests/test_ai_declarative_annotations.py
+tests/test_ai_enterprise_hardening.py
 tests/test_ai_module.py
+tests/test_ai_stream_hardening.py
 tests/test_annotation_combinations.py
 tests/test_annotations_contract.py
 tests/test_aop_annotations_full.py
@@ -278,9 +281,11 @@ tests/test_bus.py
 tests/test_cache_annotations.py
 tests/test_cli.py
 tests/test_cloud_embedded_full.py
+tests/test_cloud_log_safety.py
 tests/test_component_init_error.py
 tests/test_compositional_annotations.py
 tests/test_conditional_annotations.py
+tests/test_config_admin_hardening.py
 tests/test_config_binding.py
 tests/test_config_center.py
 tests/test_config_center_hardening.py
@@ -297,10 +302,12 @@ tests/test_datasource_routing.py
 tests/test_ddl_auto.py
 tests/test_declarative_aop_post_authorize_recover.py
 tests/test_devtools.py
+tests/test_devtools_lifecycle_hardening.py
 tests/test_di_config_event_full.py
 tests/test_excel_module.py
 tests/test_hateoas.py
 tests/test_i18n_module.py
+tests/test_io_http_logging_hardening.py
 tests/test_ioc_lifecycle_enhancements.py
 tests/test_jpa_version_transient.py
 tests/test_kafka.py

{springbootai-2.3.8 → springbootai-2.3.9}/springbootai/__init__.py

@@ -19,7 +19,7 @@ from .config import *
 from .utils import *
 from .main import create_app, run, SpringApplication, run_cli
 
-__version__ = "2.3.8"
+__version__ = "2.3.9"
 __author__ = "yuconggen"
 __license__ = "MIT"
 

{springbootai-2.3.8 → springbootai-2.3.9}/springbootai/ai/__init__.py

@@ -16,7 +16,7 @@ SpringBootAI AI 模块 - 对齐 Spring AI 的 ChatClient/Advisor/ETL 抽象，
 from springbootai.ai.core import (
     Advisor, AdvisorRequest, ChatClient, ChatClientBuilder, ChatModel,
     ChatResponse, EmbeddingModel, Generation, Message, MessageType,
-    PromptSpec,
+    PromptSpec, TokenBudgetExceededError, ToolLoopLimitExceededError,
 )
 from springbootai.ai.annotations import (
     Agent, AiAdvisor, AiCache, AiClient, AiMemory, AiRetry, ContentModeration,
@@ -34,7 +34,8 @@ from springbootai.ai.vectorstore import (
     SearchRequest, SimpleInMemoryVectorStore, VectorStore, cosine_similarity,
 )
 from springbootai.ai.etl import (
-    CharacterTextSplitter, DocumentReader, TextDocument, TextReader,
+    CharacterTextSplitter, DocumentReader, DocumentTooLargeError,
+    TextDocument, TextReader,
     TextSplitter, TokenTextSplitter,
 )
 from springbootai.ai.tools import (
@@ -42,10 +43,12 @@ from springbootai.ai.tools import (
     ToolDefinition,
     ToolExecutionPolicy,
     ToolRegistry,
+    ToolCancellationToken,
 )
 from springbootai.ai.providers import (
     FakeChatModel, FakeEmbeddingModel, OllamaChatModel, OllamaEmbeddingModel,
     OpenAICompatChatModel, OpenAIChatModel, OpenAIEmbeddingModel,
+    ProviderProtocolError, ProviderResponseTooLargeError, ProviderStreamError,
 )
 from springbootai.ai.resilience import (
     AICircuitBreaker, CircuitOpenError, TransientError, resilient_call,
@@ -54,13 +57,14 @@ from springbootai.ai.observability import AIMetrics, ai_metrics
 from springbootai.ai.autoconfig import AIProperties, bind_ai_config, configure_ai
 from springbootai.ai.annotation_runtime import ContentModerationError
 
-__version__ = "2.3.8"
+__version__ = "2.3.9"
 
 __all__ = [
     # core
     "Advisor", "AdvisorRequest", "ChatClient", "ChatClientBuilder",
     "ChatModel", "ChatResponse", "EmbeddingModel", "Generation", "Message",
-    "MessageType", "PromptSpec",
+    "MessageType", "PromptSpec", "TokenBudgetExceededError",
+    "ToolLoopLimitExceededError",
     # annotations
     "AiAdvisor", "AiClient", "AiMemory", "Tool", "Prompt", "RAG",
     "StructuredOutput", "Agent", "Embedding", "VectorStore", "AiRetry",
@@ -76,14 +80,18 @@ __all__ = [
     "SearchRequest", "SimpleInMemoryVectorStore", "VectorStore",
     "cosine_similarity",
     # etl
-    "CharacterTextSplitter", "DocumentReader", "TextDocument", "TextReader",
+    "CharacterTextSplitter", "DocumentReader", "DocumentTooLargeError",
+    "TextDocument", "TextReader",
     "TextSplitter", "TokenTextSplitter",
     # tools
-    "CompositeToolRegistry", "ToolDefinition", "ToolExecutionPolicy", "ToolRegistry",
+    "CompositeToolRegistry", "ToolCancellationToken", "ToolDefinition",
+    "ToolExecutionPolicy", "ToolRegistry",
     # providers
     "FakeChatModel", "FakeEmbeddingModel", "OllamaChatModel",
     "OllamaEmbeddingModel", "OpenAICompatChatModel", "OpenAIChatModel",
-    "OpenAIEmbeddingModel",
+    "OpenAIEmbeddingModel", "ProviderProtocolError",
+    "ProviderResponseTooLargeError",
+    "ProviderStreamError",
     # resilience
     "AICircuitBreaker", "CircuitOpenError", "TransientError", "resilient_call",
     # observability

{springbootai-2.3.8 → springbootai-2.3.9}/springbootai/ai/advisors.py

@@ -1,7 +1,9 @@
 """
 Advisor 实现 - QuestionAnswerAdvisor（RAG）与 MessageChatMemoryAdvisor（会话记忆）。
 """
-from typing import Any, Dict, List
+import inspect
+from urllib.parse import quote
+from typing import Any, Dict, List, Optional, Sequence
 
 from springbootai.ai.core import (
     Advisor, AdvisorRequest, ChatResponse, Message, MessageType,
@@ -25,11 +27,31 @@ class MessageChatMemoryAdvisor(Advisor):
 
     @staticmethod
     def _build_namespace(context: dict) -> str:
-        tenant = str(context.get("tenant_id", "")).strip()
-        user = str(context.get("user_id", "")).strip()
+        tenant = quote(str(context.get("tenant_id", "")).strip()[:128], safe="-_.~")
+        user = quote(str(context.get("user_id", "")).strip()[:128], safe="-_.~")
         parts = [p for p in (tenant, user) if p]
         return ":".join(parts) if parts else ""
 
+    @staticmethod
+    def _call_memory(method, *args, namespace: str, **kwargs):
+        """Call new namespaced memory APIs without silently weakening old backends."""
+        try:
+            parameters = inspect.signature(method).parameters.values()
+            supports_namespace = any(
+                p.name == "namespace" or p.kind == inspect.Parameter.VAR_KEYWORD
+                for p in parameters
+            )
+        except (TypeError, ValueError):
+            supports_namespace = False
+        if namespace and not supports_namespace:
+            raise RuntimeError(
+                "ChatMemory backend does not support request-scoped namespaces; "
+                "refusing a potentially cross-tenant memory access"
+            )
+        if supports_namespace:
+            kwargs["namespace"] = namespace
+        return method(*args, **kwargs)
+
     def __init__(self, memory: ChatMemory, max_messages: int = 20):
         self.memory = memory
         self.max_messages = max_messages
@@ -44,12 +66,15 @@ class MessageChatMemoryAdvisor(Advisor):
                 "跳过历史注入。请在上游设置 request.context['conversation_id']。")
             return request
 
-        # 如果 memory 支持 namespace 则注入用户/租户上下文
+        # Namespace is passed per operation. Never mutate a shared memory bean:
+        # concurrent requests may belong to different tenants.
         ns = self._build_namespace(request.context)
-        if ns and hasattr(self.memory, "_namespace"):
-            self.memory._namespace = ns
+        request.context["memory_namespace"] = ns
 
-        history = self.memory.get(str(conv_id), last_n=self.max_messages)
+        history = self._call_memory(
+            self.memory.get, str(conv_id), last_n=self.max_messages,
+            namespace=ns,
+        )
         # 历史 + 本次输入合并
         request.messages = history + request.messages
         return request
@@ -59,14 +84,18 @@ class MessageChatMemoryAdvisor(Advisor):
         conv_id = request.context.get("conversation_id")
         if not conv_id:
             return response
+        ns = str(request.context.get("memory_namespace") or
+                 self._build_namespace(request.context))
         # 保存用户输入（最后一条 user 消息）
         for msg in reversed(request.messages):
             if msg.type == MessageType.USER:
-                self.memory.add(conv_id, msg)
+                self._call_memory(
+                    self.memory.add, str(conv_id), msg, namespace=ns)
                 break
         # 保存模型回复
         if response.output:
-            self.memory.add(conv_id, response.output)
+            self._call_memory(
+                self.memory.add, str(conv_id), response.output, namespace=ns)
         return response
 
 
@@ -89,12 +118,16 @@ class QuestionAnswerAdvisor(Advisor):
                  prompt_template: str = "",
                  top_k: int = 4,
                  embedding_model=None,
-                 harden_injection: bool = True):
+                 harden_injection: bool = True,
+                 filter_metadata: Optional[Dict[str, Any]] = None,
+                 filter_context_keys: Sequence[str] = ("tenant_id",)):
         self.vector_store = vector_store
         self.prompt_template = prompt_template or self.DEFAULT_PROMPT_TEMPLATE
         self.top_k = top_k
         self.embedding_model = embedding_model
         self.harden_injection = harden_injection
+        self.filter_metadata = dict(filter_metadata or {})
+        self.filter_context_keys = tuple(filter_context_keys)
 
     def advise_request(self, request: AdvisorRequest) -> AdvisorRequest:
         # 取最后一条用户消息作为查询
@@ -110,9 +143,21 @@ class QuestionAnswerAdvisor(Advisor):
         emb = None
         if self.embedding_model:
             emb = self.embedding_model.embed_one(query)
+        filters = dict(self.filter_metadata)
+        for key in self.filter_context_keys:
+            value = request.context.get(key)
+            if value is not None and str(value).strip():
+                filters[key] = str(value).strip()
+        # When no tenant exists, a verified user identity can still isolate a
+        # personal knowledge base. Tenant scope takes precedence by default so
+        # shared tenant documents do not require a redundant user_id field.
+        if not filters and request.context.get("user_id"):
+            filters["user_id"] = str(request.context["user_id"]).strip()
+
         search_req = SearchRequest(
             query=query, embedding=emb, top_k=self.top_k,
             similarity_threshold=0.1,
+            filter_metadata=filters or None,
         )
         docs = self.vector_store.similarity_search(search_req)
         if not docs:

{springbootai-2.3.8 → springbootai-2.3.9}/springbootai/ai/annotation_runtime.py

@@ -15,7 +15,7 @@ import os
 import threading
 import time
 from collections import OrderedDict
-from typing import Any, Callable, Iterable
+from typing import Any, Callable, Optional
 
 from springbootai.ai.annotations import (
     Agent, AiCache, AiRetry, ContentModeration, Embedding, Prompt, RAG,
@@ -449,10 +449,15 @@ def _invoke_chat(factory: Any, annotation: Prompt, query: str, values: dict[str,
     system = _render(annotation.system, values)
     if system:
         spec.system(system)
-    return spec.user(query).call()
+    spec.user(query)
+    for key in ("tenant_id", "user_id", "conversation_id"):
+        if values.get(key) is not None:
+            spec.param(key, values[key])
+    return spec.call()
 
 
-def _invoke_rag(factory: Any, annotation: RAG, query: str) -> Any:
+def _invoke_rag(factory: Any, annotation: RAG, query: str,
+                values: Optional[dict[str, Any]] = None) -> Any:
     client = _resolve(factory, annotation.client)
     store = _resolve(factory, annotation.vector_store)
     embedding = _resolve(factory, annotation.embedding)
@@ -463,7 +468,14 @@ def _invoke_rag(factory: Any, annotation: RAG, query: str) -> Any:
         store, prompt_template=annotation.prompt_template,
         top_k=annotation.top_k, embedding_model=embedding,
     )
-    return client.prompt().advisors(advisor).user(query).call()
+    spec = client.prompt().advisors(advisor).user(query)
+    # Carry only identity/session fields into AdvisorRequest. Arbitrary method
+    # arguments must never become trusted authorization context implicitly.
+    values = values or {}
+    for key in ("tenant_id", "user_id", "conversation_id"):
+        if values.get(key) is not None:
+            spec.param(key, values[key])
+    return spec.call()
 
 
 def _invoke_agent(factory: Any, annotation: Agent, query: str) -> Any:
@@ -508,7 +520,7 @@ def apply_ai_annotations(factory: Any, instance: Any, method: Callable) -> Calla
                 if query is None:
                     original = method(*args, **kwargs)
                     query = _content(original)
-                result = _invoke_rag(factory, rag_ann, query)
+                result = _invoke_rag(factory, rag_ann, query, values)
             elif agent_ann:
                 original = method(*args, **kwargs)
                 result = _invoke_agent(factory, agent_ann, _content(original))
@@ -545,7 +557,7 @@ def apply_ai_annotations(factory: Any, instance: Any, method: Callable) -> Calla
                 result = _invoke_chat(factory, prompt_ann, _render(prompt_ann.template, values), values)
             elif rag_ann:
                 original = await method(*args, **kwargs)
-                result = _invoke_rag(factory, rag_ann, _content(original))
+                result = _invoke_rag(factory, rag_ann, _content(original), values)
             elif agent_ann:
                 original = await method(*args, **kwargs)
                 result = _invoke_agent(factory, agent_ann, _content(original))