Platform for stateful agents: AI with advanced memory that can learn and self-improve over time.
fix: Bulk insert base tools if using postgres (#3419)
Matthew Zhou04d1926c03cc
diff --git a/letta/services/tool_manager.py b/letta/services/tool_manager.pyindex 9f9c7701d..98861cec9 100644--- a/letta/services/tool_manager.py+++ b/letta/services/tool_manager.py@@ -1,4 +1,3 @@-import asyncioimport importlibimport osimport warnings@@ -32,6 +31,7 @@ from letta.schemas.user import User as PydanticUserfrom letta.server.db import db_registryfrom letta.services.helpers.agent_manager_helper import calculate_multi_agent_toolsfrom letta.services.mcp.types import SSEServerConfig, StdioServerConfig+from letta.settings import settingsfrom letta.utils import enforce_types, printdlogger = get_logger(__name__)@@ -467,7 +467,10 @@ class ToolManager:actor: PydanticUser,allowed_types: Optional[Set[ToolType]] = None,) -> List[PydanticTool]:- """Add default tools defined in the various function_sets modules, optionally filtered by ToolType."""+ """Add default tools defined in the various function_sets modules, optionally filtered by ToolType.++ Optimized bulk implementation using single database session and batch operations.+ """functions_to_schema = {}for module_name in LETTA_TOOL_MODULE_NAMES:@@ -479,7 +482,8 @@ class ToolManager:except Exception as e:raise e- tools = []+ # prepare tool data for bulk operations+ tool_data_list = []for name, schema in functions_to_schema.items():if name not in LETTA_TOOL_SET:continue@@ -505,17 +509,82 @@ class ToolManager:if allowed_types is not None and tool_type not in allowed_types:continue- tools.append(- self.create_or_update_tool_async(- PydanticTool(- name=name,- tags=[tool_type.value],- source_type="python",- tool_type=tool_type,- return_char_limit=BASE_FUNCTION_RETURN_CHAR_LIMIT,- ),- actor=actor,- )+ # create pydantic tool for validation and conversion+ pydantic_tool = PydanticTool(+ name=name,+ tags=[tool_type.value],+ source_type="python",+ tool_type=tool_type,+ return_char_limit=BASE_FUNCTION_RETURN_CHAR_LIMIT,+ organization_id=actor.organization_id,)- return await asyncio.gather(*tools)+ # auto-generate description if not provided+ if pydantic_tool.description is None:+ pydantic_tool.description = pydantic_tool.json_schema.get("description", None)++ tool_data_list.append(pydantic_tool)++ if not tool_data_list:+ return []++ if settings.letta_pg_uri_no_default:+ async with db_registry.async_session() as session:+ return await self._bulk_upsert_postgresql(session, tool_data_list, actor)+ else:+ return await self._upsert_tools_individually(tool_data_list, actor)++ @trace_method+ async def _bulk_upsert_postgresql(self, session, tool_data_list: List[PydanticTool], actor: PydanticUser) -> List[PydanticTool]:+ """hyper-optimized postgresql bulk upsert using on_conflict_do_update."""+ from sqlalchemy import func, select+ from sqlalchemy.dialects.postgresql import insert++ # prepare data for bulk insert+ table = ToolModel.__table__+ valid_columns = {col.name for col in table.columns}++ insert_data = []+ for tool in tool_data_list:+ tool_dict = tool.model_dump(to_orm=True)+ # set created/updated by fields+ if actor:+ tool_dict["_created_by_id"] = actor.id+ tool_dict["_last_updated_by_id"] = actor.id++ # filter to only include columns that exist in the table+ filtered_dict = {k: v for k, v in tool_dict.items() if k in valid_columns}+ insert_data.append(filtered_dict)++ # use postgresql's native bulk upsert+ stmt = insert(table).values(insert_data)++ # on conflict, update all columns except id, created_at, and _created_by_id+ excluded = stmt.excluded+ update_dict = {}+ for col in table.columns:+ if col.name not in ("id", "created_at", "_created_by_id"):+ if col.name == "updated_at":+ update_dict[col.name] = func.now()+ else:+ update_dict[col.name] = excluded[col.name]++ upsert_stmt = stmt.on_conflict_do_update(index_elements=["name", "organization_id"], set_=update_dict)++ await session.execute(upsert_stmt)+ await session.commit()++ # fetch results+ tool_names = [tool.name for tool in tool_data_list]+ result_query = select(ToolModel).where(ToolModel.name.in_(tool_names), ToolModel.organization_id == actor.organization_id)+ result = await session.execute(result_query)+ return [tool.to_pydantic() for tool in result.scalars()]++ @trace_method+ async def _upsert_tools_individually(self, tool_data_list: List[PydanticTool], actor: PydanticUser) -> List[PydanticTool]:+ """fallback to individual upserts for sqlite (original approach)."""+ tools = []+ for tool in tool_data_list:+ upserted_tool = await self.create_or_update_tool_async(tool, actor)+ tools.append(upserted_tool)+ return toolsdiff --git a/tests/integration_test_builtin_tools.py b/tests/integration_test_builtin_tools.pyindex 5ca33c05a..4e8538257 100644--- a/tests/integration_test_builtin_tools.py+++ b/tests/integration_test_builtin_tools.py@@ -300,20 +300,12 @@ def test_web_search(@pytest.mark.asyncio-async def test_web_search_uses_agent_env_var_model(agent_state_with_web_search_env_var):+async def test_web_search_uses_agent_env_var_model():"""Test that web search uses the model specified in agent tool exec env vars."""- # mock firecrawl response- mock_search_result = {- "data": [- {- "url": "https://example.com",- "title": "Example Title",- "description": "Example description",- "markdown": "Line 1: Test content\nLine 2: More content",- }- ]- }+ # create mock agent state with web search model env var+ mock_agent_state = MagicMock()+ mock_agent_state.get_agent_env_vars_as_dict.return_value = {WEB_SEARCH_MODEL_ENV_VAR_NAME: "gpt-4o"}# mock openai responsemock_openai_response = MagicMock()@@ -350,7 +342,7 @@ async def test_web_search_uses_agent_env_var_model(agent_state_with_web_search_etask = SearchTask(query="test query", question="test question")- await executor.web_search(agent_state=agent_state_with_web_search_env_var, tasks=[task], limit=1)+ await executor.web_search(agent_state=mock_agent_state, tasks=[task], limit=1)# verify correct model was usedmock_openai_client.beta.chat.completions.parse.assert_called_once()