Platform for stateful agents: AI with advanced memory that can learn and self-improve over time.
feat: support filtering out messages when converting to openai dict (#4337)
cthomas1edcc1377810
diff --git a/letta/agent.py b/letta/agent.pyindex cbfa7f6cf..99b48f1f2 100644--- a/letta/agent.py+++ b/letta/agent.py@@ -1106,7 +1106,7 @@ class Agent(BaseAgent):def summarize_messages_inplace(self):in_context_messages = self.agent_manager.get_in_context_messages(agent_id=self.agent_state.id, actor=self.user)- 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)in_context_messages_openai_no_system = in_context_messages_openai[1:]token_counts = get_token_counts_for_messages(in_context_messages)logger.info(f"System message token count={token_counts[0]}")@@ -1212,7 +1212,7 @@ class Agent(BaseAgent):# Grab the in-context messages# conversion of messages to OpenAI dict format, which is passed to the token counterin_context_messages = self.agent_manager.get_in_context_messages(agent_id=self.agent_state.id, actor=self.user)- 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)# Check if there's a summary message in the message queueif (@@ -1312,7 +1312,7 @@ class Agent(BaseAgent):)# conversion of messages to OpenAI dict format, which is passed to the token counter- 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)# Extract system, memory and external summaryif (diff --git a/letta/interface.py b/letta/interface.pyindex 733aaddbc..7e146b07e 100644--- a/letta/interface.py+++ b/letta/interface.py@@ -248,7 +248,7 @@ class CLIInterface(AgentInterface):@staticmethoddef print_messages(message_sequence: List[Message], dump=False):# rewrite to dict format- message_sequence = [msg.to_openai_dict() for msg in message_sequence]+ message_sequence = Message.to_openai_dicts_from_list(message_sequence)idx = len(message_sequence)for msg in message_sequence:@@ -291,7 +291,7 @@ class CLIInterface(AgentInterface):@staticmethoddef print_messages_simple(message_sequence: List[Message]):# rewrite to dict format- message_sequence = [msg.to_openai_dict() for msg in message_sequence]+ message_sequence = Message.to_openai_dicts_from_list(message_sequence)for msg in message_sequence:role = msg["role"]@@ -309,7 +309,7 @@ class CLIInterface(AgentInterface):@staticmethoddef print_messages_raw(message_sequence: List[Message]):# rewrite to dict format- message_sequence = [msg.to_openai_dict() for msg in message_sequence]+ message_sequence = Message.to_openai_dicts_from_list(message_sequence)for msg in message_sequence:print(msg)diff --git a/letta/interfaces/openai_streaming_interface.py b/letta/interfaces/openai_streaming_interface.pyindex 7bf3a1a30..08f23e0fa 100644--- a/letta/interfaces/openai_streaming_interface.py+++ b/letta/interfaces/openai_streaming_interface.py@@ -113,6 +113,7 @@ class OpenAIStreamingInterface:if self.messages:# Convert messages to dict format for token countingmessage_dicts = [msg.to_openai_dict() if hasattr(msg, "to_openai_dict") else msg for msg in self.messages]+ message_dicts = [m for m in message_dicts if m is not None]self.fallback_input_tokens = num_tokens_from_messages(message_dicts) # fallback to gpt-4 cl100k-baseif self.tools:diff --git a/letta/llm_api/deepseek_client.py b/letta/llm_api/deepseek_client.pyindex 7d02fcd94..a0037b1e2 100644--- a/letta/llm_api/deepseek_client.py+++ b/letta/llm_api/deepseek_client.py@@ -11,7 +11,7 @@ from openai.types.chat.chat_completion_chunk import ChatCompletionChunkfrom letta.llm_api.openai_client import OpenAIClientfrom letta.otel.tracing import trace_methodfrom letta.schemas.llm_config import LLMConfig-from letta.schemas.message import Message as PydanticMessage, Message as _Message+from letta.schemas.message import Message as PydanticMessagefrom 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 nativelyadd_functions_to_system_message(diff --git a/letta/llm_api/helpers.py b/letta/llm_api/helpers.pyindex b6f3acb86..c87ec1886 100644--- a/letta/llm_api/helpers.py+++ b/letta/llm_api/helpers.py@@ -310,7 +310,7 @@ def calculate_summarizer_cutoff(in_context_messages: List[Message], token_countsf"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_countsdef 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_countsdiff --git a/letta/llm_api/llm_api_tools.py b/letta/llm_api/llm_api_tools.pyindex 3050bec6c..8a75bc7bf 100644--- a/letta/llm_api/llm_api_tools.py+++ b/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 0if prompt_tokens + function_tokens > llm_config.context_window:diff --git a/letta/llm_api/openai.py b/letta/llm_api/openai.pyindex b113b7cd6..da1e5a8d0 100644--- a/letta/llm_api/openai.py+++ b/letta/llm_api/openai.py@@ -21,7 +21,7 @@ from letta.local_llm.utils import num_tokens_from_functions, num_tokens_from_mesfrom letta.log import get_loggerfrom letta.otel.tracing import log_eventfrom 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 _MessageRolefrom 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="",diff --git a/letta/llm_api/openai_client.py b/letta/llm_api/openai_client.pyindex fb5412103..43ea6dc4e 100644--- a/letta/llm_api/openai_client.py+++ b/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:diff --git a/letta/local_llm/chat_completion_proxy.py b/letta/local_llm/chat_completion_proxy.pyindex ba0ab45a9..1129b1258 100644--- a/letta/local_llm/chat_completion_proxy.py+++ b/letta/local_llm/chat_completion_proxy.py@@ -22,6 +22,7 @@ from letta.local_llm.webui.api import get_webui_completionfrom letta.local_llm.webui.legacy_api import get_webui_completion as get_webui_completion_legacyfrom letta.otel.tracing import log_eventfrom letta.prompts.gpt_summarize import SYSTEM as SUMMARIZE_SYSTEM_MESSAGE+from letta.schemas.message import Message as PydanticMessagefrom letta.schemas.openai.chat_completion_response import ChatCompletionResponse, Choice, Message, ToolCall, UsageStatisticsfrom letta.utils import get_tool_call_id@@ -61,7 +62,7 @@ def get_chat_completion(# TODO: eventually just process Message objectif not isinstance(messages[0], dict):- messages = [m.to_openai_dict() for m in messages]+ messages = PydanticMessage.to_openai_dicts_from_list(messages)if function_call is not None and function_call != "auto":raise ValueError(f"function_call == {function_call} not supported (auto or None only)")diff --git a/letta/schemas/message.py b/letta/schemas/message.pyindex ed0af7394..ee5430d61 100644--- a/letta/schemas/message.py+++ b/letta/schemas/message.py@@ -730,7 +730,7 @@ class Message(BaseMessage):max_tool_id_length: int = TOOL_CALL_ID_MAX_LEN,put_inner_thoughts_in_kwargs: bool = False,use_developer_message: bool = False,- ) -> dict:+ ) -> dict | None:"""Go from Message class to ChatCompletion message object"""# TODO change to pydantic casting, eg `return SystemMessageModel(self)`@@ -822,6 +822,24 @@ class Message(BaseMessage):return openai_message+ @staticmethod+ def to_openai_dicts_from_list(+ messages: List[Message],+ max_tool_id_length: int = TOOL_CALL_ID_MAX_LEN,+ put_inner_thoughts_in_kwargs: bool = False,+ use_developer_message: bool = False,+ ) -> List[dict]:+ result = [+ m.to_openai_dict(+ max_tool_id_length=max_tool_id_length,+ put_inner_thoughts_in_kwargs=put_inner_thoughts_in_kwargs,+ use_developer_message=use_developer_message,+ )+ for m in messages+ ]+ result = [m for m in result if m is not None]+ return result+def to_anthropic_dict(self,inner_thoughts_xml_tag="thinking",diff --git a/letta/server/rest_api/utils.py b/letta/server/rest_api/utils.pyindex a193e57b6..c72b25136 100644--- a/letta/server/rest_api/utils.py+++ b/letta/server/rest_api/utils.py@@ -435,7 +435,9 @@ def convert_in_context_letta_messages_to_openai(in_context_messages: List[Messagpass # It's not JSON, leave as-is# Finally, convert to dict using your existing method- openai_messages.append(msg.to_openai_dict())+ m = msg.to_openai_dict()+ assert m is not None+ openai_messages.append(m)return openai_messagesdiff --git a/letta/services/context_window_calculator/token_counter.py b/letta/services/context_window_calculator/token_counter.pyindex 244a8a52a..c432d72e2 100644--- a/letta/services/context_window_calculator/token_counter.py+++ b/letta/services/context_window_calculator/token_counter.py@@ -6,6 +6,7 @@ from typing import Any, Dict, Listfrom letta.helpers.decorators import async_redis_cachefrom letta.llm_api.anthropic_client import AnthropicClientfrom letta.otel.tracing import trace_method+from letta.schemas.message import Messagefrom letta.schemas.openai.chat_completion_request import Tool as OpenAIToolfrom letta.utils import count_tokens@@ -124,4 +125,4 @@ class TiktokenCounter(TokenCounter):return num_tokens_from_functions(functions=functions, model=self.model)def convert_messages(self, messages: List[Any]) -> List[Dict[str, Any]]:- return [m.to_openai_dict() for m in messages]+ return Message.to_openai_dicts_from_list(messages)diff --git a/letta/services/summarizer/summarizer.py b/letta/services/summarizer/summarizer.pyindex 575bf351d..3e4d040a5 100644--- a/letta/services/summarizer/summarizer.py+++ b/letta/services/summarizer/summarizer.py@@ -295,7 +295,9 @@ class Summarizer:def simple_formatter(messages: List[Message], include_system: bool = False) -> str:"""Go from an OpenAI-style list of messages to a concatenated string"""- parsed_messages = [message.to_openai_dict() for message in messages if message.role != MessageRole.system or include_system]+ parsed_messages = Message.to_openai_dicts_from_list(+ [message for message in messages if message.role != MessageRole.system or include_system]+ )return "\n".join(json.dumps(msg) for msg in parsed_messages)