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.

fix: Refresh blocks in between steps (#1335)

Matthew Zhou82f9377c8af5
diff --git a/letta/agents/voice_agent.py b/letta/agents/voice_agent.py
index 006b5827f..f60bb6612 100644
--- a/letta/agents/voice_agent.py
+++ b/letta/agents/voice_agent.py
@@ -5,11 +5,15 @@ from typing import Any, AsyncGenerator, Dict, List, Tuple
import openai
from letta.agents.base_agent import BaseAgent
-from letta.agents.ephemeral_agent import EphemeralAgent
from letta.agents.ephemeral_memory_agent import EphemeralMemoryAgent
from letta.constants import NON_USER_MSG_PREFIX
from letta.helpers.datetime_helpers import get_utc_time
-from letta.helpers.tool_execution_helper import add_pre_execution_message, execute_external_tool, remove_request_heartbeat
+from letta.helpers.tool_execution_helper import (
+ add_pre_execution_message,
+ enable_strict_mode,
+ execute_external_tool,
+ remove_request_heartbeat,
+)
from letta.interfaces.openai_chat_completions_streaming_interface import OpenAIChatCompletionsStreamingInterface
from letta.log import get_logger
from letta.orm.enums import ToolType
@@ -38,7 +42,6 @@ from letta.services.helpers.agent_manager_helper import compile_system_message
from letta.services.message_manager import MessageManager
from letta.services.passage_manager import PassageManager
from letta.services.summarizer.enums import SummarizationMode
-from letta.services.summarizer.summarizer import Summarizer
from letta.utils import united_diff
logger = get_logger(__name__)
@@ -75,16 +78,16 @@ class VoiceAgent(BaseAgent):
self.passage_manager = PassageManager() # TODO: pass this in
# TODO: This is not guaranteed to exist!
self.summary_block_label = "human"
- self.summarizer = Summarizer(
- mode=summarization_mode,
- summarizer_agent=EphemeralAgent(
- agent_id=agent_id, openai_client=openai_client, message_manager=message_manager, agent_manager=agent_manager, actor=actor
- ),
- message_buffer_limit=message_buffer_limit,
- message_buffer_min=message_buffer_min,
- )
+ # self.summarizer = Summarizer(
+ # mode=summarization_mode,
+ # summarizer_agent=EphemeralAgent(
+ # agent_id=agent_id, openai_client=openai_client, message_manager=message_manager, agent_manager=agent_manager, actor=actor
+ # ),
+ # message_buffer_limit=message_buffer_limit,
+ # message_buffer_min=message_buffer_min,
+ # )
self.message_buffer_limit = message_buffer_limit
- self.message_buffer_min = message_buffer_min
+ # self.message_buffer_min = message_buffer_min
self.offline_memory_agent = EphemeralMemoryAgent(
agent_id=agent_id, openai_client=openai_client, message_manager=message_manager, agent_manager=agent_manager, actor=actor
)
@@ -105,8 +108,6 @@ class VoiceAgent(BaseAgent):
# TODO: Define max steps here
while True:
- step_first = True
-
# Rebuild memory each loop
in_context_messages = self._rebuild_memory(in_context_messages, agent_state)
openai_messages = convert_letta_messages_to_openai(in_context_messages)
@@ -119,8 +120,6 @@ class VoiceAgent(BaseAgent):
# 1) Yield partial tokens from OpenAI
async for sse_chunk in streaming_interface.process(stream):
- if step_first:
- step_first = False
yield sse_chunk
# 2) Now handle the final AI response. This might yield more text (stalling, etc.)
@@ -241,6 +240,14 @@ class VoiceAgent(BaseAgent):
)
def _rebuild_memory(self, in_context_messages: List[Message], agent_state: AgentState) -> List[Message]:
+ # Refresh memory
+ # TODO: This only happens for the summary block
+ # TODO: We want to extend this refresh to be general, and stick it in agent_manager
+ for i, b in enumerate(agent_state.memory.blocks):
+ if b.label == self.summary_block_label:
+ agent_state.memory.blocks[i] = self.block_manager.get_block_by_id(block_id=b.id, actor=self.actor)
+ break
+
# TODO: This is a pretty brittle pattern established all over our code, need to get rid of this
curr_system_message = in_context_messages[0]
curr_memory_str = agent_state.memory.compile()
@@ -309,28 +316,31 @@ class VoiceAgent(BaseAgent):
)
recall_memory_json = Tool(
type="function",
- function=add_pre_execution_message(
- {
- "name": "recall_memory",
- "description": "Retrieve relevant information from memory based on a given query. Use when you don't remember the answer to a question.",
- "parameters": {
- "type": "object",
- "properties": {
- "query": {
- "type": "string",
- "description": "A description of what the model is trying to recall from memory.",
- }
+ function=enable_strict_mode(
+ add_pre_execution_message(
+ {
+ "name": "recall_memory",
+ "description": "Retrieve relevant information from memory based on a given query. Use when you don't remember the answer to a question.",
+ "parameters": {
+ "type": "object",
+ "properties": {
+ "query": {
+ "type": "string",
+ "description": "A description of what the model is trying to recall from memory.",
+ }
+ },
+ "required": ["query"],
},
- "required": ["query"],
},
- },
- description=recall_memory_utterance_description,
+ description=recall_memory_utterance_description,
+ )
),
)
# TODO: Customize whether or not to have heartbeats, pre_exec_message, etc.
return [recall_memory_json] + [
- Tool(type="function", function=add_pre_execution_message(remove_request_heartbeat(t.json_schema))) for t in tools
+ Tool(type="function", function=enable_strict_mode(add_pre_execution_message(remove_request_heartbeat(t.json_schema))))
+ for t in tools
]
async def _execute_tool(self, tool_name: str, tool_args: dict, agent_state: AgentState) -> Tuple[str, bool]:
diff --git a/letta/helpers/tool_execution_helper.py b/letta/helpers/tool_execution_helper.py
index 886cf9950..f307d730b 100644
--- a/letta/helpers/tool_execution_helper.py
+++ b/letta/helpers/tool_execution_helper.py
@@ -36,7 +36,6 @@ def enable_strict_mode(tool_schema: Dict[str, Any]) -> Dict[str, Any]:
# Set additionalProperties to False
parameters["additionalProperties"] = False
schema["parameters"] = parameters
-
return schema
diff --git a/letta/schemas/openai/chat_completion_request.py b/letta/schemas/openai/chat_completion_request.py
index 12486bca9..f8ad12644 100644
--- a/letta/schemas/openai/chat_completion_request.py
+++ b/letta/schemas/openai/chat_completion_request.py
@@ -82,6 +82,7 @@ class FunctionSchema(BaseModel):
name: str
description: Optional[str] = None
parameters: Optional[Dict[str, Any]] = None # JSON Schema for the parameters
+ strict: bool = False
class Tool(BaseModel):
diff --git a/letta/server/rest_api/routers/v1/voice.py b/letta/server/rest_api/routers/v1/voice.py
index 5f187c5ee..5e32b60f5 100644
--- a/letta/server/rest_api/routers/v1/voice.py
+++ b/letta/server/rest_api/routers/v1/voice.py
@@ -68,8 +68,8 @@ async def create_voice_chat_completions(
agent_manager=server.agent_manager,
block_manager=server.block_manager,
actor=actor,
- message_buffer_limit=100,
- message_buffer_min=4,
+ message_buffer_limit=50,
+ message_buffer_min=10,
)
# Return the streaming generator
diff --git a/letta/services/agent_manager.py b/letta/services/agent_manager.py
index 6bf4a5cfd..d9e3894aa 100644
--- a/letta/services/agent_manager.py
+++ b/letta/services/agent_manager.py
@@ -755,7 +755,7 @@ class AgentManager:
if updated_value != agent_state.memory.get_block(label).value:
# update the block if it's changed
block_id = agent_state.memory.get_block(label).id
- block = self.block_manager.update_block(block_id=block_id, block_update=BlockUpdate(value=updated_value), actor=actor)
+ self.block_manager.update_block(block_id=block_id, block_update=BlockUpdate(value=updated_value), actor=actor)
# refresh memory from DB (using block ids)
agent_state.memory = Memory(
diff --git a/tests/integration_test_chat_completions.py b/tests/integration_test_chat_completions.py
index c7e77a970..36c7b742d 100644
--- a/tests/integration_test_chat_completions.py
+++ b/tests/integration_test_chat_completions.py
@@ -110,11 +110,11 @@ def composio_gmail_get_profile_tool(default_user):
@pytest.fixture(scope="function")
-def agent(client, roll_dice_tool, weather_tool, composio_gmail_get_profile_tool):
+def agent(client, roll_dice_tool, weather_tool):
"""Creates an agent and ensures cleanup after tests."""
agent_state = client.create_agent(
name=f"test_compl_{str(uuid.uuid4())[5:]}",
- tool_ids=[roll_dice_tool.id, weather_tool.id, composio_gmail_get_profile_tool.id],
+ tool_ids=[roll_dice_tool.id, weather_tool.id],
include_base_tools=False,
)
yield agent_state