Skip to content

armgit / letta

Public Imported from GitHub
0

Platform for stateful agents: AI with advanced memory that can learn and self-improve over time.

feat: move deepseek to new agent loop [LET-4081] (#4266)

cthomas27458ffad5fe
diff --git a/letta/llm_api/deepseek_client.py b/letta/llm_api/deepseek_client.py
index 1d3828c0c..7c4d9d748 100644
--- a/letta/llm_api/deepseek_client.py
+++ b/letta/llm_api/deepseek_client.py
@@ -1,3 +1,4 @@
+import json
import os
from typing import List, Optional
@@ -10,6 +11,7 @@ from letta.llm_api.openai_client import OpenAIClient
from letta.otel.tracing import trace_method
from letta.schemas.llm_config import LLMConfig
from letta.schemas.message import Message as PydanticMessage
+from letta.schemas.openai.chat_completion_request import ChatMessage, cast_message_to_subtype
from letta.schemas.openai.chat_completion_response import ChatCompletionResponse
from letta.settings import model_settings
@@ -36,15 +38,19 @@ class DeepseekClient(OpenAIClient):
data = super().build_request_data(messages, llm_config, tools, force_tool_call)
def add_functions_to_system_message(system_message: ChatMessage):
- system_message.content += f"<available functions> {''.join(json.dumps(f) for f in functions)} </available functions>"
+ system_message.content += f"<available functions> {''.join(json.dumps(f) for f in tools)} </available functions>"
system_message.content += 'Select best function to call simply respond with a single json block with the fields "name" and "arguments". Use double quotes around the arguments.'
+ openai_message_list = [cast_message_to_subtype(m.to_openai_dict(put_inner_thoughts_in_kwargs=False)) for m in messages]
+
if llm_config.model == "deepseek-reasoner": # R1 currently doesn't support function calling natively
add_functions_to_system_message(
- data["messages"][0]
+ openai_message_list[0]
) # Inject additional instructions to the system prompt with the available functions
- data["messages"] = map_messages_to_deepseek_format(data["messages"])
+ openai_message_list = map_messages_to_deepseek_format(openai_message_list)
+
+ data["messages"] = [m.dict() for m in openai_message_list]
return data
@@ -94,4 +100,6 @@ class DeepseekClient(OpenAIClient):
Handles potential extraction of inner thoughts if they were added via kwargs.
"""
response = ChatCompletionResponse(**response_data)
+ if response.choices[0].message.tool_calls:
+ return super().convert_response_to_chat_completion(response_data, input_messages, llm_config)
return convert_deepseek_response_to_chatcompletion(response)
diff --git a/letta/server/rest_api/routers/v1/agents.py b/letta/server/rest_api/routers/v1/agents.py
index 987e1dc55..df34961ce 100644
--- a/letta/server/rest_api/routers/v1/agents.py
+++ b/letta/server/rest_api/routers/v1/agents.py
@@ -1078,6 +1078,7 @@ async def send_message(
"azure",
"xai",
"groq",
+ "deepseek",
]
# Create a new run for execution tracking
@@ -1223,8 +1224,9 @@ async def send_message_streaming(
"azure",
"xai",
"groq",
+ "deepseek",
]
- model_compatible_token_streaming = agent.llm_config.model_endpoint_type in ["anthropic", "openai", "bedrock"]
+ model_compatible_token_streaming = agent.llm_config.model_endpoint_type in ["anthropic", "openai", "bedrock", "deepseek"]
# Create a new job for execution tracking
if settings.track_agent_run:
@@ -1480,7 +1482,10 @@ async def _process_message_background(
"google_vertex",
"bedrock",
"ollama",
+ "azure",
+ "xai",
"groq",
+ "deepseek",
]
if agent_eligible and model_compatible:
if agent.enable_sleeptime and agent.agent_type != AgentType.voice_convo_agent:
@@ -1671,6 +1676,7 @@ async def preview_raw_payload(
"azure",
"xai",
"groq",
+ "deepseek",
]
if agent_eligible and model_compatible:
@@ -1742,6 +1748,7 @@ async def summarize_agent_conversation(
"azure",
"xai",
"groq",
+ "deepseek",
]
if agent_eligible and model_compatible: