Fimeg/letta-server
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

feat: add compaction_settings to agents (#6625)

Browse Source 
* initial commit

* Add database migration for compaction_settings field

This migration adds the compaction_settings column to the agents table
to support customized summarization configuration for each agent.

🐾 Generated with [Letta Code](https://letta.com)

Co-Authored-By: Letta <noreply@letta.com>

* fix

* rename

* update apis

* fix tests

* update web test

---------

Co-authored-by: Letta <noreply@letta.com>
Co-authored-by: Kian Jones <kian@letta.com>
This commit is contained in:
Sarah Wooders
2025-12-11 14:14:04 -08:00
committed by Caren Thomas
parent 4309ecf606
commit 7ea297231a
13 changed files with 324 additions and 37 deletions
Download Patch File Download Diff File Expand all files Collapse all files

24
tests/integration_test_summarizer.py
View File

@@ -618,12 +618,12 @@ async def test_summarize_multiple_large_tool_calls(server: SyncServer, actor, ll
#
# ======================================================================================================================
# SummarizerConfig Mode Tests (with pytest.patch) - Using LettaAgentV3
# CompactionSettings Mode Tests (with pytest.patch) - Using LettaAgentV3
# ======================================================================================================================
from unittest.mock import patch
from letta.services.summarizer.summarizer_config import SummarizerConfig, get_default_summarizer_config
from letta.services.summarizer.summarizer_config import CompactionSettings, get_default_compaction_settings
# Test both summarizer modes: "all" summarizes entire history, "sliding_window" keeps recent messages
SUMMARIZER_CONFIG_MODES: list[Literal["all", "sliding_window"]] = ["all", "sliding_window"]
@@ -634,7 +634,7 @@ SUMMARIZER_CONFIG_MODES: list[Literal["all", "sliding_window"]] = ["all", "slidi
@pytest.mark.parametrize("llm_config", TESTED_LLM_CONFIGS, ids=[c.model for c in TESTED_LLM_CONFIGS])
async def test_summarize_with_mode(server: SyncServer, actor, llm_config: LLMConfig, mode: Literal["all", "sliding_window"]):
"""
Test summarization with different SummarizerConfig modes using LettaAgentV3.
Test summarization with different CompactionSettings modes using LettaAgentV3.
This test verifies that both summarization modes work correctly:
- "all": Summarizes the entire conversation history into a single summary
@@ -674,11 +674,11 @@ async def test_summarize_with_mode(server: SyncServer, actor, llm_config: LLMCon
# Persist the new messages
new_letta_messages = await server.message_manager.create_many_messages_async(new_letta_messages, actor=actor)
# Create a custom SummarizerConfig with the desired mode
def mock_get_default_summarizer_config(model_settings):
config = get_default_summarizer_config(model_settings)
# Create a custom CompactionSettings with the desired mode
def mock_get_default_compaction_settings(model_settings):
config = get_default_compaction_settings(model_settings)
# Override the mode
return SummarizerConfig(
return CompactionSettings(
model_settings=config.model_settings,
prompt=config.prompt,
prompt_acknowledgement=config.prompt_acknowledgement,
@@ -687,7 +687,7 @@ async def test_summarize_with_mode(server: SyncServer, actor, llm_config: LLMCon
sliding_window_percentage=config.sliding_window_percentage,
)
with patch("letta.agents.letta_agent_v3.get_default_summarizer_config", mock_get_default_summarizer_config):
with patch("letta.agents.letta_agent_v3.get_default_compaction_settings", mock_get_default_compaction_settings):
agent_loop = LettaAgentV3(agent_state=agent_state, actor=actor)
summary, result = await agent_loop.compact(messages=in_context_messages)
@@ -848,13 +848,13 @@ async def test_sliding_window_cutoff_index_does_not_exceed_message_count(server:
the sliding window logic works with actual token counting.
"""
from letta.schemas.model import ModelSettings
from letta.services.summarizer.summarizer_config import get_default_summarizer_config
from letta.services.summarizer.summarizer_config import get_default_compaction_settings
from letta.services.summarizer.summarizer_sliding_window import summarize_via_sliding_window
# Create a real summarizer config using the default factory
# Override sliding_window_percentage to 0.3 for this test
model_settings = ModelSettings() # Use defaults
summarizer_config = get_default_summarizer_config(model_settings)
summarizer_config = get_default_compaction_settings(model_settings)
summarizer_config.sliding_window_percentage = 0.3
# Create 65 messages (similar to the failing case in the bug report)
@@ -1401,11 +1401,11 @@ async def test_summarize_all(server: SyncServer, actor, llm_config: LLMConfig):
"""
from letta.schemas.model import ModelSettings
from letta.services.summarizer.summarizer_all import summarize_all
from letta.services.summarizer.summarizer_config import get_default_summarizer_config
from letta.services.summarizer.summarizer_config import get_default_compaction_settings
# Create a summarizer config with "all" mode
model_settings = ModelSettings()
summarizer_config = get_default_summarizer_config(model_settings)
summarizer_config = get_default_compaction_settings(model_settings)
summarizer_config.mode = "all"
# Create test messages - a simple conversation

89
tests/managers/test_agent_manager.py
View File

@@ -81,6 +81,7 @@ from letta.schemas.letta_stop_reason import LettaStopReason, StopReasonType
from letta.schemas.llm_batch_job import AgentStepState, LLMBatchItem
from letta.schemas.llm_config import LLMConfig
from letta.schemas.message import Message as PydanticMessage, MessageCreate, MessageUpdate
from letta.schemas.model import ModelSettings
from letta.schemas.openai.chat_completion_response import UsageStatistics
from letta.schemas.organization import Organization, Organization as PydanticOrganization, OrganizationUpdate
from letta.schemas.passage import Passage as PydanticPassage
@@ -96,6 +97,7 @@ from letta.server.server import SyncServer
from letta.services.block_manager import BlockManager
from letta.services.helpers.agent_manager_helper import calculate_base_tools, calculate_multi_agent_tools, validate_agent_exists_async
from letta.services.step_manager import FeedbackType
from letta.services.summarizer.summarizer_config import CompactionSettings
from letta.settings import settings, tool_settings
from letta.utils import calculate_file_defaults_based_on_context_window
from tests.helpers.utils import comprehensive_agent_checks, validate_context_window_overview
@@ -526,6 +528,91 @@ async def test_update_agent(server: SyncServer, comprehensive_test_agent_fixture
assert updated_agent.updated_at > last_updated_timestamp
@pytest.mark.asyncio
async def test_create_agent_with_compaction_settings(server: SyncServer, default_user, default_block):
"""Test that agents can be created with custom compaction_settings"""
# Upsert base tools
await server.tool_manager.upsert_base_tools_async(actor=default_user)
# Create custom compaction settings
llm_config = LLMConfig.default_config("gpt-4o-mini")
model_settings = llm_config._to_model_settings()
compaction_settings = CompactionSettings(
model_settings=model_settings,
prompt="Custom summarization prompt",
prompt_acknowledgement="Acknowledged",
clip_chars=1500,
mode="all",
sliding_window_percentage=0.5,
)
# Create agent with compaction settings
create_agent_request = CreateAgent(
name="test_compaction_agent",
agent_type="memgpt_v2_agent",
system="test system",
llm_config=llm_config,
embedding_config=EmbeddingConfig.default_config(provider="openai"),
block_ids=[default_block.id],
include_base_tools=True,
compaction_settings=compaction_settings,
)
created_agent = await server.agent_manager.create_agent_async(
create_agent_request,
actor=default_user,
)
# Verify compaction settings were stored correctly
assert created_agent.compaction_settings is not None
assert created_agent.compaction_settings.mode == "all"
assert created_agent.compaction_settings.clip_chars == 1500
assert created_agent.compaction_settings.sliding_window_percentage == 0.5
assert created_agent.compaction_settings.prompt == "Custom summarization prompt"
assert created_agent.compaction_settings.prompt_acknowledgement == "Acknowledged"
# Clean up
await server.agent_manager.delete_agent_async(agent_id=created_agent.id, actor=default_user)
@pytest.mark.asyncio
async def test_update_agent_compaction_settings(server: SyncServer, comprehensive_test_agent_fixture, default_user):
"""Test that an agent's compaction_settings can be updated"""
agent, _ = comprehensive_test_agent_fixture
# Verify initial state (should be None or default)
assert agent.compaction_settings is None
# Create new compaction settings
llm_config = LLMConfig.default_config("gpt-4o-mini")
model_settings = llm_config._to_model_settings()
new_compaction_settings = CompactionSettings(
model_settings=model_settings,
prompt="Updated summarization prompt",
prompt_acknowledgement="Updated acknowledgement",
clip_chars=3000,
mode="sliding_window",
sliding_window_percentage=0.4,
)
# Update agent with compaction settings
update_agent_request = UpdateAgent(
compaction_settings=new_compaction_settings,
)
updated_agent = await server.agent_manager.update_agent_async(agent.id, update_agent_request, actor=default_user)
# Verify compaction settings were updated correctly
assert updated_agent.compaction_settings is not None
assert updated_agent.compaction_settings.mode == "sliding_window"
assert updated_agent.compaction_settings.clip_chars == 3000
assert updated_agent.compaction_settings.sliding_window_percentage == 0.4
assert updated_agent.compaction_settings.prompt == "Updated summarization prompt"
assert updated_agent.compaction_settings.prompt_acknowledgement == "Updated acknowledgement"
@pytest.mark.asyncio
async def test_agent_file_defaults_based_on_context_window(server: SyncServer, default_user, default_block):
"""Test that file-related defaults are set based on the model's context window size"""
@@ -1282,6 +1369,7 @@ async def test_agent_state_schema_unchanged(server: SyncServer):
from letta.schemas.source import Source
from letta.schemas.tool import Tool
from letta.schemas.tool_rule import ToolRule
from letta.services.summarizer.summarizer_config import CompactionSettings
# Define the expected schema structure
expected_schema = {
@@ -1298,6 +1386,7 @@ async def test_agent_state_schema_unchanged(server: SyncServer):
"agent_type": AgentType,
# LLM information
"llm_config": LLMConfig,
"compaction_settings": CompactionSettings,
"model": str,
"embedding": str,
"embedding_config": EmbeddingConfig,