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

feat: agent generate endpoint (#9304)

Browse Source 
* base

* update

* clean up

* update
This commit is contained in:
jnjpng
2026-02-05 12:58:45 -08:00
committed by Caren Thomas
parent 5c6ca705f1
commit 0bdedb3c0f
4 changed files with 105 additions and 14 deletions
Download Patch File Download Diff File Expand all files Collapse all files

13
fern/openapi.json
View File

@@ -33671,6 +33671,19 @@
],
"title": "Override Model",
"description": "Model handle to use instead of agent's default (e.g., 'openai/gpt-4', 'anthropic/claude-3-5-sonnet')"
},
"response_schema": {
"anyOf": [
{
"additionalProperties": true,
"type": "object"
},
{
"type": "null"
}
],
"title": "Response Schema",
"description": "JSON schema for structured output. When provided, the LLM will be forced to return a response matching this schema via tool calling. The schema should follow JSON Schema format with 'properties' and optionally 'required' fields."
}
},
"type": "object",

16
letta/llm_api/anthropic_client.py
View File

@@ -496,7 +496,14 @@ class AnthropicClient(LLMClientBase):
}
# Extended Thinking
if self.is_reasoning_model(llm_config) and llm_config.enable_reasoner:
# Note: Anthropic does not allow thinking when forcing tool use with split_thread_agent
should_enable_thinking = (
self.is_reasoning_model(llm_config)
and llm_config.enable_reasoner
and not (agent_type == AgentType.split_thread_agent and force_tool_call is not None)
)
if should_enable_thinking:
# Opus 4.6 uses Auto Thinking (no budget tokens)
if llm_config.model.startswith("claude-opus-4-6"):
data["thinking"] = {
@@ -556,7 +563,12 @@ class AnthropicClient(LLMClientBase):
tool_choice = None
elif self.is_reasoning_model(llm_config) and llm_config.enable_reasoner or agent_type == AgentType.letta_v1_agent:
# NOTE: reasoning models currently do not allow for `any`
# NOTE: react agents should always have auto on, since the precense/absense of tool calls controls chaining
# NOTE: react agents should always have at least auto on, since the precense/absense of tool calls controls chaining
if agent_type == AgentType.split_thread_agent and force_tool_call is not None:
tool_choice = {"type": "tool", "name": force_tool_call, "disable_parallel_tool_use": True}
# When forcing a specific tool, only include that tool
tools_for_request = [OpenAITool(function=f) for f in tools if f["name"] == force_tool_call]
else:
tool_choice = {"type": "auto", "disable_parallel_tool_use": True}
tools_for_request = [OpenAITool(function=f) for f in tools]
elif force_tool_call is not None:

10
letta/server/rest_api/routers/v1/agents.py
View File

@@ -103,6 +103,15 @@ class GenerateRequest(BaseModel):
description="Model handle to use instead of agent's default (e.g., 'openai/gpt-4', 'anthropic/claude-3-5-sonnet')",
)
response_schema: Optional[Dict[str, Any]] = Field(
None,
description=(
"JSON schema for structured output. When provided, the LLM will be forced to return "
"a response matching this schema via tool calling. The schema should follow JSON Schema "
"format with 'properties' and optionally 'required' fields."
),
)
@field_validator("prompt")
@classmethod
def validate_prompt_not_empty(cls, v: str) -> str:
@@ -1876,6 +1885,7 @@ async def generate_completion(
system_prompt=request.system_prompt,
actor=actor,
override_model=request.override_model,
response_schema=request.response_schema,
)
except NoResultFound:
raise HTTPException(status_code=404, detail=f"Agent with ID {agent_id} not found")

72
letta/services/agent_generate_completion_manager.py
View File

@@ -1,16 +1,20 @@
"""Manager for handling direct LLM completions using agent configuration."""
from typing import TYPE_CHECKING, Optional
import json
from typing import TYPE_CHECKING, Any, Dict, Optional
from letta.errors import HandleNotFoundError, LLMError
from letta.llm_api.llm_client import LLMClient
from letta.log import get_logger
from letta.orm.errors import NoResultFound
from letta.schemas.enums import MessageRole
from letta.schemas.enums import AgentType, MessageRole
from letta.schemas.letta_message_content import TextContent
from letta.schemas.message import Message
from letta.schemas.usage import LettaUsageStatistics
# Tool name used for structured output via tool forcing
STRUCTURED_OUTPUT_TOOL_NAME = "structured_output"
if TYPE_CHECKING:
from letta.orm import User
from letta.schemas.llm_config import LLMConfig
@@ -19,6 +23,27 @@ if TYPE_CHECKING:
logger = get_logger(__name__)
def _schema_to_tool_definition(schema: Dict[str, Any]) -> Dict[str, Any]:
"""
Convert a JSON schema into a tool definition for forced tool calling.
Args:
schema: JSON schema object with 'properties' and optionally 'required'
Returns:
Tool definition dict compatible with OpenAI/Anthropic function calling format
"""
return {
"name": STRUCTURED_OUTPUT_TOOL_NAME,
"description": "Returns a structured response matching the requested schema.",
"parameters": {
"type": "object",
"properties": schema.get("properties", {}),
"required": schema.get("required", list(schema.get("properties", {}).keys())),
},
}
class GenerateResponse:
"""Response from direct LLM generation."""
@@ -49,13 +74,14 @@ class AgentGenerateCompletionManager:
actor: "User",
system_prompt: Optional[str] = None,
override_model: Optional[str] = None,
response_schema: Optional[Dict[str, Any]] = None,
) -> GenerateResponse:
"""
Generate a completion directly from the LLM provider using the agent's configuration.
This method makes a direct request to the LLM provider without any agent processing:
- No memory or context retrieval
- No tool calling
- No tool calling (unless response_schema is provided)
- No message persistence
- No agent state modification
@@ -66,9 +92,14 @@ class AgentGenerateCompletionManager:
system_prompt: Optional system prompt to prepend to the conversation
override_model: Optional model handle to override the agent's default
(e.g., 'openai/gpt-4', 'anthropic/claude-3-5-sonnet')
response_schema: Optional JSON schema for structured output. When provided,
the LLM will be forced to return a response matching this
schema via tool calling.
Returns:
GenerateResponse with content, model, and usage statistics
GenerateResponse with content, model, and usage statistics.
When response_schema is provided, content will be the JSON string
matching the schema.
Raises:
NoResultFound: If agent not found
@@ -99,6 +130,7 @@ class AgentGenerateCompletionManager:
"override_model": override_model,
"prompt_length": len(prompt),
"has_system_prompt": system_prompt is not None,
"has_response_schema": response_schema is not None,
"model": llm_config.model,
},
)
@@ -132,13 +164,23 @@ class AgentGenerateCompletionManager:
if llm_client is None:
raise LLMError(f"Unsupported provider type: {llm_config.model_endpoint_type}")
# 5. Build request data (no tools, no function calling)
# 5. Build request data
# If response_schema is provided, create a tool and force the model to call it
tools = None
force_tool_call = None
if response_schema:
tools = [_schema_to_tool_definition(response_schema)]
force_tool_call = STRUCTURED_OUTPUT_TOOL_NAME
# TODO: create a separate agent type
effective_agent_type = AgentType.split_thread_agent if response_schema else agent.agent_type
request_data = llm_client.build_request_data(
agent_type=agent.agent_type,
agent_type=effective_agent_type,
messages=letta_messages,
llm_config=llm_config,
tools=None, # No tools for direct generation
force_tool_call=None,
tools=tools,
force_tool_call=force_tool_call,
)
# 6. Make direct LLM request
@@ -155,6 +197,20 @@ class AgentGenerateCompletionManager:
content = ""
if chat_completion.choices and len(chat_completion.choices) > 0:
message = chat_completion.choices[0].message
if response_schema:
# When using structured output, extract from tool call arguments
if message.tool_calls and len(message.tool_calls) > 0:
# The tool call arguments contain the structured output as JSON string
content = message.tool_calls[0].function.arguments
else:
# Fallback: some providers may return in content even with tool forcing
content = message.content or ""
logger.warning(
"Expected tool call for structured output but got content response",
extra={"agent_id": str(agent_id), "content_length": len(content)},
)
else:
content = message.content or ""
# 9. Extract usage statistics