Platform for stateful agents: AI with advanced memory that can learn and self-improve over time.
fix: set thinking budget for vertex tokens (#2367)
cthomasf9d2793cafec
diff --git a/letta/llm_api/google_ai_client.py b/letta/llm_api/google_ai_client.pyindex 47671398c..f1d8e091e 100644--- a/letta/llm_api/google_ai_client.py+++ b/letta/llm_api/google_ai_client.py@@ -7,7 +7,10 @@ from letta.errors import ErrorCode, LLMAuthenticationError, LLMErrorfrom letta.llm_api.google_constants import GOOGLE_MODEL_FOR_API_KEY_CHECKfrom letta.llm_api.google_vertex_client import GoogleVertexClientfrom letta.log import get_logger+from letta.schemas.llm_config import LLMConfig+from letta.schemas.message import Message as PydanticMessagefrom letta.settings import model_settings+from letta.tracing import trace_methodlogger = get_logger(__name__)@@ -17,6 +20,18 @@ class GoogleAIClient(GoogleVertexClient):def _get_client(self):return genai.Client(api_key=model_settings.gemini_api_key)+ @trace_method+ def build_request_data(+ self,+ messages: List[PydanticMessage],+ llm_config: LLMConfig,+ tools: List[dict],+ force_tool_call: Optional[str] = None,+ ) -> dict:+ request = super().build_request_data(messages, llm_config, tools, force_tool_call)+ del request["config"]["thinking_config"]+ return request+def get_gemini_endpoint_and_headers(base_url: str, model: Optional[str], api_key: str, key_in_header: bool = True, generate_content: bool = Falsediff --git a/letta/llm_api/google_vertex_client.py b/letta/llm_api/google_vertex_client.pyindex e8215813f..afc80ebde 100644--- a/letta/llm_api/google_vertex_client.py+++ b/letta/llm_api/google_vertex_client.py@@ -244,11 +244,10 @@ class GoogleVertexClient(LLMClientBase):# Add thinking_config# If enable_reasoner is False, set thinking_budget to 0# Otherwise, use the value from max_reasoning_tokens- if llm_config.enable_reasoner:- thinking_config = ThinkingConfig(- thinking_budget=llm_config.max_reasoning_tokens,- )- request_data["config"]["thinking_config"] = thinking_config.model_dump()+ thinking_config = ThinkingConfig(+ thinking_budget=llm_config.max_reasoning_tokens if llm_config.enable_reasoner else 0,+ )+ request_data["config"]["thinking_config"] = thinking_config.model_dump()return request_data