MemGPT/letta/llm_api/anthropic_client.py

488 lines
19 KiB
Python

import json
import re
from typing import List, Optional, Union
import anthropic
from anthropic.types import Message as AnthropicMessage
from letta.helpers.datetime_helpers import get_utc_time
from letta.llm_api.helpers import add_inner_thoughts_to_functions, unpack_all_inner_thoughts_from_kwargs
from letta.llm_api.llm_api_tools import cast_message_to_subtype
from letta.llm_api.llm_client_base import LLMClientBase
from letta.local_llm.constants import INNER_THOUGHTS_KWARG, INNER_THOUGHTS_KWARG_DESCRIPTION
from letta.log import get_logger
from letta.schemas.message import Message as PydanticMessage
from letta.schemas.openai.chat_completion_request import ChatCompletionRequest, Tool
from letta.schemas.openai.chat_completion_response import ChatCompletionResponse, Choice, FunctionCall
from letta.schemas.openai.chat_completion_response import Message as ChoiceMessage
from letta.schemas.openai.chat_completion_response import ToolCall, UsageStatistics
from letta.services.provider_manager import ProviderManager
DUMMY_FIRST_USER_MESSAGE = "User initializing bootup sequence."
logger = get_logger(__name__)
class AnthropicClient(LLMClientBase):
def request(self, request_data: dict) -> dict:
try:
client = self._get_anthropic_client(async_client=False)
response = client.beta.messages.create(**request_data, betas=["tools-2024-04-04"])
return response.model_dump()
except Exception as e:
self._handle_anthropic_error(e)
async def request_async(self, request_data: dict) -> dict:
try:
client = self._get_anthropic_client(async_client=True)
response = await client.beta.messages.create(**request_data, betas=["tools-2024-04-04"])
return response.model_dump()
except Exception as e:
self._handle_anthropic_error(e)
def _get_anthropic_client(self, async_client: bool = False) -> Union[anthropic.AsyncAnthropic, anthropic.Anthropic]:
override_key = ProviderManager().get_anthropic_override_key()
if async_client:
return anthropic.AsyncAnthropic(api_key=override_key) if override_key else anthropic.AsyncAnthropic()
return anthropic.Anthropic(api_key=override_key) if override_key else anthropic.Anthropic()
def _handle_anthropic_error(self, e: Exception):
if isinstance(e, anthropic.APIConnectionError):
logger.warning(f"[Anthropic] API connection error: {e.__cause__}")
elif isinstance(e, anthropic.RateLimitError):
logger.warning("[Anthropic] Rate limited (429). Consider backoff.")
elif isinstance(e, anthropic.APIStatusError):
logger.warning(f"[Anthropic] API status error: {e.status_code}, {e.response}")
raise e
def build_request_data(
self,
messages: List[PydanticMessage],
tools: List[dict],
tool_call: Optional[str],
force_tool_call: Optional[str] = None,
) -> dict:
if not self.use_tool_naming:
raise NotImplementedError("Only tool calling supported on Anthropic API requests")
if tools is None:
# Special case for summarization path
available_tools = None
tool_choice = None
elif force_tool_call is not None:
assert tools is not None
tool_choice = {"type": "tool", "name": force_tool_call}
available_tools = [{"type": "function", "function": f} for f in tools if f["name"] == force_tool_call]
# need to have this setting to be able to put inner thoughts in kwargs
self.llm_config.put_inner_thoughts_in_kwargs = True
else:
if self.llm_config.put_inner_thoughts_in_kwargs:
# tool_choice_type other than "auto" only plays nice if thinking goes inside the tool calls
tool_choice = {"type": "any", "disable_parallel_tool_use": True}
else:
tool_choice = {"type": "auto", "disable_parallel_tool_use": True}
available_tools = [{"type": "function", "function": f} for f in tools]
chat_completion_request = ChatCompletionRequest(
model=self.llm_config.model,
messages=[cast_message_to_subtype(m.to_openai_dict()) for m in messages],
tools=available_tools,
tool_choice=tool_choice,
max_tokens=self.llm_config.max_tokens, # Note: max_tokens is required for Anthropic API
temperature=self.llm_config.temperature,
)
return _prepare_anthropic_request(
data=chat_completion_request,
put_inner_thoughts_in_kwargs=self.llm_config.put_inner_thoughts_in_kwargs,
extended_thinking=self.llm_config.enable_reasoner,
max_reasoning_tokens=self.llm_config.max_reasoning_tokens,
)
def convert_response_to_chat_completion(
self,
response_data: dict,
input_messages: List[PydanticMessage],
) -> ChatCompletionResponse:
"""
Example response from Claude 3:
response.json = {
'id': 'msg_01W1xg9hdRzbeN2CfZM7zD2w',
'type': 'message',
'role': 'assistant',
'content': [
{
'type': 'text',
'text': "<thinking>Analyzing user login event. This is Chad's first
interaction with me. I will adjust my personality and rapport accordingly.</thinking>"
},
{
'type':
'tool_use',
'id': 'toolu_01Ka4AuCmfvxiidnBZuNfP1u',
'name': 'core_memory_append',
'input': {
'name': 'human',
'content': 'Chad is logging in for the first time. I will aim to build a warm
and welcoming rapport.',
'request_heartbeat': True
}
}
],
'model': 'claude-3-haiku-20240307',
'stop_reason': 'tool_use',
'stop_sequence': None,
'usage': {
'input_tokens': 3305,
'output_tokens': 141
}
}
"""
response = AnthropicMessage(**response_data)
prompt_tokens = response.usage.input_tokens
completion_tokens = response.usage.output_tokens
finish_reason = remap_finish_reason(response.stop_reason)
content = None
reasoning_content = None
reasoning_content_signature = None
redacted_reasoning_content = None
tool_calls = None
if len(response.content) > 0:
for content_part in response.content:
if content_part.type == "text":
content = strip_xml_tags(string=content_part.text, tag="thinking")
if content_part.type == "tool_use":
tool_calls = [
ToolCall(
id=content_part.id,
type="function",
function=FunctionCall(
name=content_part.name,
arguments=json.dumps(content_part.input, indent=2),
),
)
]
if content_part.type == "thinking":
reasoning_content = content_part.thinking
reasoning_content_signature = content_part.signature
if content_part.type == "redacted_thinking":
redacted_reasoning_content = content_part.data
else:
raise RuntimeError("Unexpected empty content in response")
assert response.role == "assistant"
choice = Choice(
index=0,
finish_reason=finish_reason,
message=ChoiceMessage(
role=response.role,
content=content,
reasoning_content=reasoning_content,
reasoning_content_signature=reasoning_content_signature,
redacted_reasoning_content=redacted_reasoning_content,
tool_calls=tool_calls,
),
)
chat_completion_response = ChatCompletionResponse(
id=response.id,
choices=[choice],
created=get_utc_time(),
model=response.model,
usage=UsageStatistics(
prompt_tokens=prompt_tokens,
completion_tokens=completion_tokens,
total_tokens=prompt_tokens + completion_tokens,
),
)
if self.llm_config.put_inner_thoughts_in_kwargs:
chat_completion_response = unpack_all_inner_thoughts_from_kwargs(
response=chat_completion_response, inner_thoughts_key=INNER_THOUGHTS_KWARG
)
return chat_completion_response
def _prepare_anthropic_request(
data: ChatCompletionRequest,
inner_thoughts_xml_tag: Optional[str] = "thinking",
# if true, prefix fill the generation with the thinking tag
prefix_fill: bool = True,
# if true, put COT inside the tool calls instead of inside the content
put_inner_thoughts_in_kwargs: bool = False,
bedrock: bool = False,
# extended thinking related fields
# https://docs.anthropic.com/en/docs/build-with-claude/extended-thinking
extended_thinking: bool = False,
max_reasoning_tokens: Optional[int] = None,
) -> dict:
"""Prepare the request data for Anthropic API format."""
if extended_thinking:
assert (
max_reasoning_tokens is not None and max_reasoning_tokens < data.max_tokens
), "max tokens must be greater than thinking budget"
assert not put_inner_thoughts_in_kwargs, "extended thinking not compatible with put_inner_thoughts_in_kwargs"
# assert not prefix_fill, "extended thinking not compatible with prefix_fill"
# Silently disable prefix_fill for now
prefix_fill = False
# if needed, put inner thoughts as a kwarg for all tools
if data.tools and put_inner_thoughts_in_kwargs:
functions = add_inner_thoughts_to_functions(
functions=[t.function.model_dump() for t in data.tools],
inner_thoughts_key=INNER_THOUGHTS_KWARG,
inner_thoughts_description=INNER_THOUGHTS_KWARG_DESCRIPTION,
)
data.tools = [Tool(function=f) for f in functions]
# convert the tools to Anthropic's payload format
anthropic_tools = None if data.tools is None else convert_tools_to_anthropic_format(data.tools)
# pydantic -> dict
data = data.model_dump(exclude_none=True)
if extended_thinking:
data["thinking"] = {
"type": "enabled",
"budget_tokens": max_reasoning_tokens,
}
# `temperature` may only be set to 1 when thinking is enabled. Please consult our documentation at https://docs.anthropic.com/en/docs/build-with-claude/extended-thinking#important-considerations-when-using-extended-thinking'
data["temperature"] = 1.0
if "functions" in data:
raise ValueError(f"'functions' unexpected in Anthropic API payload")
# Handle tools
if "tools" in data and data["tools"] is None:
data.pop("tools")
data.pop("tool_choice", None)
elif anthropic_tools is not None:
# TODO eventually enable parallel tool use
data["tools"] = anthropic_tools
# Move 'system' to the top level
assert data["messages"][0]["role"] == "system", f"Expected 'system' role in messages[0]:\n{data['messages'][0]}"
data["system"] = data["messages"][0]["content"]
data["messages"] = data["messages"][1:]
# Process messages
for message in data["messages"]:
if "content" not in message:
message["content"] = None
# Convert to Anthropic format
msg_objs = [
PydanticMessage.dict_to_message(
user_id=None,
agent_id=None,
openai_message_dict=m,
)
for m in data["messages"]
]
data["messages"] = [
m.to_anthropic_dict(
inner_thoughts_xml_tag=inner_thoughts_xml_tag,
put_inner_thoughts_in_kwargs=put_inner_thoughts_in_kwargs,
)
for m in msg_objs
]
# Ensure first message is user
if data["messages"][0]["role"] != "user":
data["messages"] = [{"role": "user", "content": DUMMY_FIRST_USER_MESSAGE}] + data["messages"]
# Handle alternating messages
data["messages"] = merge_tool_results_into_user_messages(data["messages"])
# Handle prefix fill (not compatible with inner-thouguhts-in-kwargs)
# https://docs.anthropic.com/en/api/messages#body-messages
# NOTE: cannot prefill with tools for opus:
# Your API request included an `assistant` message in the final position, which would pre-fill the `assistant` response. When using tools with "claude-3-opus-20240229"
if prefix_fill and not put_inner_thoughts_in_kwargs and "opus" not in data["model"]:
if not bedrock: # not support for bedrock
data["messages"].append(
# Start the thinking process for the assistant
{"role": "assistant", "content": f"<{inner_thoughts_xml_tag}>"},
)
# Validate max_tokens
assert "max_tokens" in data, data
# Remove OpenAI-specific fields
for field in ["frequency_penalty", "logprobs", "n", "top_p", "presence_penalty", "user", "stream"]:
data.pop(field, None)
return data
def convert_tools_to_anthropic_format(tools: List[Tool]) -> List[dict]:
"""See: https://docs.anthropic.com/claude/docs/tool-use
OpenAI style:
"tools": [{
"type": "function",
"function": {
"name": "find_movies",
"description": "find ....",
"parameters": {
"type": "object",
"properties": {
PARAM: {
"type": PARAM_TYPE, # eg "string"
"description": PARAM_DESCRIPTION,
},
...
},
"required": List[str],
}
}
}
]
Anthropic style:
"tools": [{
"name": "find_movies",
"description": "find ....",
"input_schema": {
"type": "object",
"properties": {
PARAM: {
"type": PARAM_TYPE, # eg "string"
"description": PARAM_DESCRIPTION,
},
...
},
"required": List[str],
}
}
]
Two small differences:
- 1 level less of nesting
- "parameters" -> "input_schema"
"""
formatted_tools = []
for tool in tools:
formatted_tool = {
"name": tool.function.name,
"description": tool.function.description,
"input_schema": tool.function.parameters or {"type": "object", "properties": {}, "required": []},
}
formatted_tools.append(formatted_tool)
return formatted_tools
def merge_tool_results_into_user_messages(messages: List[dict]):
"""Anthropic API doesn't allow role 'tool'->'user' sequences
Example HTTP error:
messages: roles must alternate between "user" and "assistant", but found multiple "user" roles in a row
From: https://docs.anthropic.com/claude/docs/tool-use
You may be familiar with other APIs that return tool use as separate from the model's primary output,
or which use a special-purpose tool or function message role.
In contrast, Anthropic's models and API are built around alternating user and assistant messages,
where each message is an array of rich content blocks: text, image, tool_use, and tool_result.
"""
# TODO walk through the messages list
# When a dict (dict_A) with 'role' == 'user' is followed by a dict with 'role' == 'user' (dict B), do the following
# dict_A["content"] = dict_A["content"] + dict_B["content"]
# The result should be a new merged_messages list that doesn't have any back-to-back dicts with 'role' == 'user'
merged_messages = []
if not messages:
return merged_messages
# Start with the first message in the list
current_message = messages[0]
for next_message in messages[1:]:
if current_message["role"] == "user" and next_message["role"] == "user":
# Merge contents of the next user message into current one
current_content = (
current_message["content"]
if isinstance(current_message["content"], list)
else [{"type": "text", "text": current_message["content"]}]
)
next_content = (
next_message["content"]
if isinstance(next_message["content"], list)
else [{"type": "text", "text": next_message["content"]}]
)
merged_content = current_content + next_content
current_message["content"] = merged_content
else:
# Append the current message to result as it's complete
merged_messages.append(current_message)
# Move on to the next message
current_message = next_message
# Append the last processed message to the result
merged_messages.append(current_message)
return merged_messages
def remap_finish_reason(stop_reason: str) -> str:
"""Remap Anthropic's 'stop_reason' to OpenAI 'finish_reason'
OpenAI: 'stop', 'length', 'function_call', 'content_filter', null
see: https://platform.openai.com/docs/guides/text-generation/chat-completions-api
From: https://docs.anthropic.com/claude/reference/migrating-from-text-completions-to-messages#stop-reason
Messages have a stop_reason of one of the following values:
"end_turn": The conversational turn ended naturally.
"stop_sequence": One of your specified custom stop sequences was generated.
"max_tokens": (unchanged)
"""
if stop_reason == "end_turn":
return "stop"
elif stop_reason == "stop_sequence":
return "stop"
elif stop_reason == "max_tokens":
return "length"
elif stop_reason == "tool_use":
return "function_call"
else:
raise ValueError(f"Unexpected stop_reason: {stop_reason}")
def strip_xml_tags(string: str, tag: Optional[str]) -> str:
if tag is None:
return string
# Construct the regular expression pattern to find the start and end tags
tag_pattern = f"<{tag}.*?>|</{tag}>"
# Use the regular expression to replace the tags with an empty string
return re.sub(tag_pattern, "", string)
def strip_xml_tags_streaming(string: str, tag: Optional[str]) -> str:
if tag is None:
return string
# Handle common partial tag cases
parts_to_remove = [
"<", # Leftover start bracket
f"<{tag}", # Opening tag start
f"</{tag}", # Closing tag start
f"/{tag}>", # Closing tag end
f"{tag}>", # Opening tag end
f"/{tag}", # Partial closing tag without >
">", # Leftover end bracket
]
result = string
for part in parts_to_remove:
result = result.replace(part, "")
return result