feat: support filtering out messages when converting to openai dict (#4337)

* feat: support filtering out messages when converting to openai dict

* fix imports

letta/llm_api/deepseek_client.py

@@ -11,7 +11,7 @@ from openai.types.chat.chat_completion_chunk import ChatCompletionChunk
 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, Message as _Message
+from letta.schemas.message import Message as PydanticMessage
 from letta.schemas.openai.chat_completion_request import (
     AssistantMessage,
     ChatCompletionRequest,
@@ -119,7 +119,9 @@ def build_deepseek_chat_completions_request(
     #         inner_thoughts_description=inner_thoughts_desc,
     #     )
 
-    openai_message_list = [cast_message_to_subtype(m.to_openai_dict(put_inner_thoughts_in_kwargs=False)) for m in messages]
+    openai_message_list = [
+        cast_message_to_subtype(m) for m in PydanticMessage.to_openai_dicts_from_list(messages, put_inner_thoughts_in_kwargs=False)
+    ]
 
     if llm_config.model:
         model = llm_config.model
@@ -343,7 +345,9 @@ class DeepseekClient(OpenAIClient):
             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]
+        openai_message_list = [
+            cast_message_to_subtype(m) for m in PydanticMessage.to_openai_dicts_from_list(messages, put_inner_thoughts_in_kwargs=False)
+        ]
 
         if llm_config.model == "deepseek-reasoner":  # R1 currently doesn't support function calling natively
             add_functions_to_system_message(

letta/llm_api/helpers.py

@@ -310,7 +310,7 @@ def calculate_summarizer_cutoff(in_context_messages: List[Message], token_counts
             f"Given in_context_messages has different length from given token_counts: {len(in_context_messages)} != {len(token_counts)}"
         )
 
-    in_context_messages_openai = [m.to_openai_dict() for m in in_context_messages]
+    in_context_messages_openai = Message.to_openai_dicts_from_list(in_context_messages)
 
     if summarizer_settings.evict_all_messages:
         logger.info("Evicting all messages...")
@@ -351,7 +351,7 @@ def calculate_summarizer_cutoff(in_context_messages: List[Message], token_counts
 
 
 def get_token_counts_for_messages(in_context_messages: List[Message]) -> List[int]:
-    in_context_messages_openai = [m.to_openai_dict() for m in in_context_messages]
+    in_context_messages_openai = Message.to_openai_dicts_from_list(in_context_messages)
     token_counts = [count_tokens(str(msg)) for msg in in_context_messages_openai]
     return token_counts
 

letta/llm_api/llm_api_tools.py

@@ -145,7 +145,7 @@ def create(
 
     # Count the tokens first, if there's an overflow exit early by throwing an error up the stack
     # NOTE: we want to include a specific substring in the error message to trigger summarization
-    messages_oai_format = [m.to_openai_dict() for m in messages]
+    messages_oai_format = Message.to_openai_dicts_from_list(messages)
     prompt_tokens = num_tokens_from_messages(messages=messages_oai_format, model=llm_config.model)
     function_tokens = num_tokens_from_functions(functions=functions, model=llm_config.model) if functions else 0
     if prompt_tokens + function_tokens > llm_config.context_window:

letta/llm_api/openai.py

@@ -21,7 +21,7 @@ from letta.local_llm.utils import num_tokens_from_functions, num_tokens_from_mes
 from letta.log import get_logger
 from letta.otel.tracing import log_event
 from letta.schemas.llm_config import LLMConfig
-from letta.schemas.message import Message as _Message, MessageRole as _MessageRole
+from letta.schemas.message import Message as PydanticMessage, MessageRole as _MessageRole
 from letta.schemas.openai.chat_completion_request import (
     ChatCompletionRequest,
     FunctionCall as ToolFunctionChoiceFunctionCall,
@@ -177,7 +177,7 @@ async def openai_get_model_list_async(
 
 def build_openai_chat_completions_request(
     llm_config: LLMConfig,
-    messages: List[_Message],
+    messages: List[PydanticMessage],
     user_id: Optional[str],
     functions: Optional[list],
     function_call: Optional[str],
@@ -201,13 +201,12 @@ def build_openai_chat_completions_request(
     use_developer_message = accepts_developer_role(llm_config.model)
 
     openai_message_list = [
-        cast_message_to_subtype(
-            m.to_openai_dict(
-                put_inner_thoughts_in_kwargs=llm_config.put_inner_thoughts_in_kwargs,
-                use_developer_message=use_developer_message,
-            )
+        cast_message_to_subtype(m)
+        for m in PydanticMessage.to_openai_dicts_from_list(
+            messages,
+            put_inner_thoughts_in_kwargs=llm_config.put_inner_thoughts_in_kwargs,
+            use_developer_message=use_developer_message,
         )
-        for m in messages
     ]
 
     if llm_config.model:
@@ -326,7 +325,7 @@ def openai_chat_completions_process_stream(
 
     # Create a dummy Message object to get an ID and date
     # TODO(sarah): add message ID generation function
-    dummy_message = _Message(
+    dummy_message = PydanticMessage(
         role=_MessageRole.assistant,
         content=[],
         agent_id="",

letta/llm_api/openai_client.py

@@ -179,13 +179,12 @@ class OpenAIClient(LLMClientBase):
         use_developer_message = accepts_developer_role(llm_config.model)
 
         openai_message_list = [
-            cast_message_to_subtype(
-                m.to_openai_dict(
-                    put_inner_thoughts_in_kwargs=llm_config.put_inner_thoughts_in_kwargs,
-                    use_developer_message=use_developer_message,
-                )
+            cast_message_to_subtype(m)
+            for m in PydanticMessage.to_openai_dicts_from_list(
+                messages,
+                put_inner_thoughts_in_kwargs=llm_config.put_inner_thoughts_in_kwargs,
+                use_developer_message=use_developer_message,
             )
-            for m in messages
         ]
 
         if llm_config.model: