Platform for stateful agents: AI with advanced memory that can learn and self-improve over time.
feat: add tracing to llm clients (#2340)
cthomasb554171d41c4
diff --git a/letta/llm_api/anthropic_client.py b/letta/llm_api/anthropic_client.pyindex f7509b037..f131e776c 100644--- a/letta/llm_api/anthropic_client.py+++ b/letta/llm_api/anthropic_client.py@@ -45,11 +45,13 @@ logger = get_logger(__name__)class AnthropicClient(LLMClientBase):+ @trace_methoddef request(self, request_data: dict, llm_config: LLMConfig) -> dict:client = self._get_anthropic_client(llm_config, async_client=False)response = client.beta.messages.create(**request_data, betas=["tools-2024-04-04"])return response.model_dump()+ @trace_methodasync def request_async(self, request_data: dict, llm_config: LLMConfig) -> dict:client = self._get_anthropic_client(llm_config, async_client=True)response = await client.beta.messages.create(**request_data, betas=["tools-2024-04-04"])@@ -339,6 +341,7 @@ class AnthropicClient(LLMClientBase):# TODO: Input messages doesn't get used here# TODO: Clean up this interface+ @trace_methoddef convert_response_to_chat_completion(self,response_data: dict,diff --git a/letta/llm_api/google_vertex_client.py b/letta/llm_api/google_vertex_client.pyindex 2874b62a9..e8215813f 100644--- a/letta/llm_api/google_vertex_client.py+++ b/letta/llm_api/google_vertex_client.py@@ -17,6 +17,7 @@ from letta.schemas.message import Message as PydanticMessagefrom letta.schemas.openai.chat_completion_request import Toolfrom letta.schemas.openai.chat_completion_response import ChatCompletionResponse, Choice, FunctionCall, Message, ToolCall, UsageStatisticsfrom letta.settings import model_settings, settings+from letta.tracing import trace_methodfrom letta.utils import get_tool_call_idlogger = get_logger(__name__)@@ -32,6 +33,7 @@ class GoogleVertexClient(LLMClientBase):http_options={"api_version": "v1"},)+ @trace_methoddef request(self, request_data: dict, llm_config: LLMConfig) -> dict:"""Performs underlying request to llm and returns raw response.@@ -44,6 +46,7 @@ class GoogleVertexClient(LLMClientBase):)return response.model_dump()+ @trace_methodasync def request_async(self, request_data: dict, llm_config: LLMConfig) -> dict:"""Performs underlying request to llm and returns raw response.@@ -189,6 +192,7 @@ class GoogleVertexClient(LLMClientBase):return [{"functionDeclarations": function_list}]+ @trace_methoddef build_request_data(self,messages: List[PydanticMessage],@@ -248,6 +252,7 @@ class GoogleVertexClient(LLMClientBase):return request_data+ @trace_methoddef convert_response_to_chat_completion(self,response_data: dict,diff --git a/letta/llm_api/openai_client.py b/letta/llm_api/openai_client.pyindex e6ac37a22..d144d03cc 100644--- a/letta/llm_api/openai_client.py+++ b/letta/llm_api/openai_client.py@@ -32,6 +32,7 @@ from letta.schemas.openai.chat_completion_request import Tool as OpenAIToolfrom letta.schemas.openai.chat_completion_request import ToolFunctionChoice, cast_message_to_subtypefrom letta.schemas.openai.chat_completion_response import ChatCompletionResponsefrom letta.settings import model_settings+from letta.tracing import trace_methodlogger = get_logger(__name__)@@ -124,6 +125,7 @@ class OpenAIClient(LLMClientBase):return kwargs+ @trace_methoddef build_request_data(self,messages: List[PydanticMessage],@@ -213,6 +215,7 @@ class OpenAIClient(LLMClientBase):return data.model_dump(exclude_unset=True)+ @trace_methoddef request(self, request_data: dict, llm_config: LLMConfig) -> dict:"""Performs underlying synchronous request to OpenAI API and returns raw response dict.@@ -222,6 +225,7 @@ class OpenAIClient(LLMClientBase):response: ChatCompletion = client.chat.completions.create(**request_data)return response.model_dump()+ @trace_methodasync def request_async(self, request_data: dict, llm_config: LLMConfig) -> dict:"""Performs underlying asynchronous request to OpenAI API and returns raw response dict.@@ -230,6 +234,7 @@ class OpenAIClient(LLMClientBase):response: ChatCompletion = await client.chat.completions.create(**request_data)return response.model_dump()+ @trace_methoddef convert_response_to_chat_completion(self,response_data: dict,