"""Wrapper around Google VertexAI chat-based models."""
from __future__ import annotations # noqa
import ast
import json
import logging
from dataclasses import dataclass, field
from operator import itemgetter
import uuid
from typing import (
Any,
AsyncIterator,
Callable,
Dict,
Iterator,
List,
Optional,
Sequence,
Type,
Union,
cast,
Literal,
Tuple,
TypedDict,
overload,
)
import proto # type: ignore[import-untyped]
from google.cloud.aiplatform import telemetry
from langchain_core.callbacks import (
AsyncCallbackManagerForLLMRun,
CallbackManagerForLLMRun,
)
from langchain_core.language_models import LanguageModelInput
from langchain_core.language_models.chat_models import (
BaseChatModel,
LangSmithParams,
generate_from_stream,
agenerate_from_stream,
)
from langchain_core.messages import (
AIMessage,
AIMessageChunk,
BaseMessage,
FunctionMessage,
HumanMessage,
SystemMessage,
ToolCall,
ToolMessage,
)
from langchain_core.messages.ai import UsageMetadata
from langchain_core.messages.tool import (
tool_call_chunk,
tool_call as create_tool_call,
invalid_tool_call,
)
from langchain_core.output_parsers.base import OutputParserLike
from langchain_core.output_parsers import JsonOutputParser, PydanticOutputParser
from langchain_core.output_parsers.openai_tools import (
JsonOutputKeyToolsParser,
PydanticToolsParser,
)
from langchain_core.output_parsers.openai_tools import parse_tool_calls
from langchain_core.outputs import ChatGeneration, ChatGenerationChunk, ChatResult
from pydantic import BaseModel, Field, model_validator
from langchain_core.runnables import Runnable, RunnablePassthrough
from langchain_core.utils.function_calling import convert_to_openai_tool
from langchain_core.utils.pydantic import is_basemodel_subclass
from vertexai.generative_models import ( # type: ignore
Tool as VertexTool,
)
from vertexai.generative_models._generative_models import ( # type: ignore
ToolConfig,
SafetySettingsType,
GenerationConfigType,
GenerationResponse,
_convert_schema_dict_to_gapic,
)
from vertexai.language_models import ( # type: ignore
ChatMessage,
ChatModel,
ChatSession,
CodeChatModel,
CodeChatSession,
InputOutputTextPair,
)
from vertexai.preview.language_models import ( # type: ignore
ChatModel as PreviewChatModel,
)
from vertexai.preview.language_models import (
CodeChatModel as PreviewCodeChatModel,
)
from google.cloud.aiplatform_v1beta1.types import (
Blob,
Candidate,
Part,
HarmCategory,
Content,
FileData,
FunctionCall,
FunctionResponse,
GenerateContentRequest,
GenerationConfig,
SafetySetting,
Tool as GapicTool,
ToolConfig as GapicToolConfig,
VideoMetadata,
)
from langchain_google_vertexai._base import _VertexAICommon, GoogleModelFamily
from langchain_google_vertexai._image_utils import ImageBytesLoader
from langchain_google_vertexai._utils import (
create_retry_decorator,
get_generation_info,
_format_model_name,
is_gemini_model,
replace_defs_in_schema,
)
from langchain_google_vertexai.functions_utils import (
_format_tool_config,
_ToolConfigDict,
_tool_choice_to_tool_config,
_ToolChoiceType,
_ToolsType,
_format_to_gapic_tool,
_ToolType,
)
from pydantic import ConfigDict
from typing_extensions import Self
logger = logging.getLogger(__name__)
_allowed_params = [
"temperature",
"top_k",
"top_p",
"response_mime_type",
"response_schema",
"max_output_tokens",
"presence_penalty",
"frequency_penalty",
"candidate_count",
"seed",
"response_logprobs",
"logprobs",
]
_allowed_params_prediction_service = ["request", "timeout", "metadata"]
@dataclass
class _ChatHistory:
"""Represents a context and a history of messages."""
history: List[ChatMessage] = field(default_factory=list)
context: Optional[str] = None
class _GeminiGenerateContentKwargs(TypedDict):
generation_config: Optional[GenerationConfigType]
safety_settings: Optional[SafetySettingsType]
tools: Optional[List[VertexTool]]
tool_config: Optional[ToolConfig]
def _parse_chat_history(history: List[BaseMessage]) -> _ChatHistory:
"""Parse a sequence of messages into history.
Args:
history: The list of messages to re-create the history of the chat.
Returns:
A parsed chat history.
Raises:
ValueError: If a sequence of message has a SystemMessage not at the
first place.
"""
vertex_messages, context = [], None
for i, message in enumerate(history):
content = cast(str, message.content)
if i == 0 and isinstance(message, SystemMessage):
context = content
elif isinstance(message, AIMessage):
vertex_message = ChatMessage(content=message.content, author="bot")
vertex_messages.append(vertex_message)
elif isinstance(message, HumanMessage):
vertex_message = ChatMessage(content=message.content, author="user")
vertex_messages.append(vertex_message)
else:
raise ValueError(
f"Unexpected message with type {type(message)} at the position {i}."
)
chat_history = _ChatHistory(context=context, history=vertex_messages)
return chat_history
def _parse_chat_history_gemini(
history: List[BaseMessage],
project: Optional[str] = None,
convert_system_message_to_human: Optional[bool] = False,
) -> tuple[Content | None, list[Content]]:
def _convert_to_prompt(part: Union[str, Dict]) -> Optional[Part]:
if isinstance(part, str):
return Part(text=part)
if not isinstance(part, Dict):
raise ValueError(
f"Message's content is expected to be a dict, got {type(part)}!"
)
if part["type"] == "text":
return Part(text=part["text"])
if part["type"] == "tool_use":
if part.get("text"):
return Part(text=part["text"])
else:
return None
if part["type"] == "image_url":
path = part["image_url"]["url"]
return ImageBytesLoader(project=project).load_gapic_part(path)
# Handle media type like LangChain.js
# https://github.com/langchain-ai/langchainjs/blob/e536593e2585f1dd7b0afc187de4d07cb40689ba/libs/langchain-google-common/src/utils/gemini.ts#L93-L106
if part["type"] == "media":
if "mime_type" not in part:
raise ValueError(f"Missing mime_type in media part: {part}")
mime_type = part["mime_type"]
proto_part = Part()
if "data" in part:
proto_part.inline_data = Blob(data=part["data"], mime_type=mime_type)
elif "file_uri" in part:
proto_part.file_data = FileData(
file_uri=part["file_uri"], mime_type=mime_type
)
else:
raise ValueError(
f"Media part must have either data or file_uri: {part}"
)
if "video_metadata" in part:
metadata = VideoMetadata(part["video_metadata"])
proto_part.video_metadata = metadata
return proto_part
raise ValueError("Only text, image_url, and media types are supported!")
def _convert_to_parts(message: BaseMessage) -> List[Part]:
raw_content = message.content
# If a user sends a multimodal request with agents, then the full input
# will be sent as a string due to the ChatPromptTemplate formatting.
# Because of this, we need to first try to convert the string to its
# native type (such as list or dict) so that results can be properly
# appended to the prompt, otherwise they will all be parsed as Text
# rather than `inline_data`.
if isinstance(raw_content, str):
try:
raw_content = ast.literal_eval(raw_content)
except SyntaxError:
pass
except ValueError:
pass
# A linting error is thrown here because it does not think this line is
# reachable due to typing, but mypy is wrong so we ignore the lint
# error.
if isinstance(raw_content, int): # type: ignore
raw_content = str(raw_content) # type: ignore
if isinstance(raw_content, str):
raw_content = [raw_content]
result = []
for raw_part in raw_content:
part = _convert_to_prompt(raw_part)
if part:
result.append(part)
return result
vertex_messages: List[Content] = []
system_parts: List[Part] | None = None
system_instruction = None
# the last AI Message before a sequence of tool calls
prev_ai_message: Optional[AIMessage] = None
for i, message in enumerate(history):
if isinstance(message, SystemMessage):
prev_ai_message = None
system_parts = _convert_to_parts(message)
if convert_system_message_to_human:
logger.warning(
"gemini models released from April 2024 support"
"SystemMessages natively. For best performances,"
"when working with these models,"
"set convert_system_message_to_human to False"
)
continue
if system_instruction is not None:
system_instruction.parts.extend(system_parts) # type: ignore[unreachable]
else:
system_instruction = Content(role="system", parts=system_parts)
system_parts = None
elif isinstance(message, HumanMessage):
prev_ai_message = None
role = "user"
parts = _convert_to_parts(message)
if system_parts is not None:
parts = system_parts + parts
system_parts = None
if vertex_messages and vertex_messages[-1].role == "user":
prev_parts = list(vertex_messages[-1].parts)
vertex_messages[-1] = Content(role=role, parts=prev_parts + parts)
else:
vertex_messages.append(Content(role=role, parts=parts))
elif isinstance(message, AIMessage):
prev_ai_message = message
role = "model"
parts = []
if message.content:
parts = _convert_to_parts(message)
for tc in message.tool_calls:
function_call = FunctionCall({"name": tc["name"], "args": tc["args"]})
parts.append(Part(function_call=function_call))
if len(vertex_messages):
prev_content = vertex_messages[-1]
prev_content_is_model = prev_content and prev_content.role == "model"
if prev_content_is_model:
prev_parts = list(prev_content.parts)
prev_parts.extend(parts)
vertex_messages[-1] = Content(role=role, parts=prev_parts)
continue
vertex_messages.append(Content(role=role, parts=parts))
elif isinstance(message, FunctionMessage):
prev_ai_message = None
role = "function"
part = Part(
function_response=FunctionResponse(
name=message.name, response={"content": message.content}
)
)
parts = [part]
if len(vertex_messages):
prev_content = vertex_messages[-1]
prev_content_is_function = (
prev_content and prev_content.role == "function"
)
if prev_content_is_function:
prev_parts = list(prev_content.parts)
prev_parts.extend(parts)
# replacing last message
vertex_messages[-1] = Content(role=role, parts=prev_parts)
continue
vertex_messages.append(Content(role=role, parts=parts))
elif isinstance(message, ToolMessage):
role = "function"
# message.name can be null for ToolMessage
name = message.name
if name is None:
if prev_ai_message:
tool_call_id = message.tool_call_id
tool_call: ToolCall | None = next(
(
t
for t in prev_ai_message.tool_calls
if t["id"] == tool_call_id
),
None,
)
if tool_call is None:
raise ValueError(
(
"Message name is empty and can't find"
+ f"corresponding tool call for id: '${tool_call_id}'"
)
)
name = tool_call["name"]
def _parse_content(raw_content: str | Dict[Any, Any]) -> Dict[Any, Any]:
if isinstance(raw_content, dict):
return raw_content
if isinstance(raw_content, str):
try:
content = json.loads(raw_content)
# json.loads("2") returns 2 since it's a valid json
if isinstance(content, dict):
return content
except json.JSONDecodeError:
pass
return {"content": raw_content}
if isinstance(message.content, list):
parsed_content = [_parse_content(c) for c in message.content]
if len(parsed_content) > 1:
merged_content: Dict[Any, Any] = {}
for content_piece in parsed_content:
for key, value in content_piece.items():
if key not in merged_content:
merged_content[key] = []
merged_content[key].append(value)
logger.warning(
"Expected content to be a str, got a list with > 1 element."
"Merging values together"
)
content = {k: "".join(v) for k, v in merged_content.items()}
else:
content = parsed_content[0]
else:
content = _parse_content(message.content)
part = Part(
function_response=FunctionResponse(
name=name,
response=content,
)
)
parts = [part]
prev_content = vertex_messages[-1]
prev_content_is_function = prev_content and prev_content.role == "function"
if prev_content_is_function:
prev_parts = list(prev_content.parts)
prev_parts.extend(parts)
# replacing last message
vertex_messages[-1] = Content(role=role, parts=prev_parts)
continue
vertex_messages.append(Content(role=role, parts=parts))
else:
raise ValueError(
f"Unexpected message with type {type(message)} at the position {i}."
)
return system_instruction, vertex_messages
def _parse_examples(examples: List[BaseMessage]) -> List[InputOutputTextPair]:
if len(examples) % 2 != 0:
raise ValueError(
f"Expect examples to have an even amount of messages, got {len(examples)}."
)
example_pairs = []
input_text = None
for i, example in enumerate(examples):
if i % 2 == 0:
if not isinstance(example, HumanMessage):
raise ValueError(
f"Expected the first message in a part to be from human, got "
f"{type(example)} for the {i}th message."
)
input_text = example.content
if i % 2 == 1:
if not isinstance(example, AIMessage):
raise ValueError(
f"Expected the second message in a part to be from AI, got "
f"{type(example)} for the {i}th message."
)
pair = InputOutputTextPair(
input_text=input_text, output_text=example.content
)
example_pairs.append(pair)
return example_pairs
def _get_question(messages: List[BaseMessage]) -> HumanMessage:
"""Get the human message at the end of a list of input messages to a chat model."""
if not messages:
raise ValueError("You should provide at least one message to start the chat!")
question = messages[-1]
if not isinstance(question, HumanMessage):
raise ValueError(
f"Last message in the list should be from human, got {question.type}."
)
return question
@overload
def _parse_response_candidate(
response_candidate: "Candidate", streaming: Literal[False] = False
) -> AIMessage:
...
@overload
def _parse_response_candidate(
response_candidate: "Candidate", streaming: Literal[True]
) -> AIMessageChunk:
...
def _parse_response_candidate(
response_candidate: "Candidate", streaming: bool = False
) -> AIMessage:
content: Union[None, str, List[str]] = None
additional_kwargs = {}
tool_calls = []
invalid_tool_calls = []
tool_call_chunks = []
for part in response_candidate.content.parts:
try:
text: Optional[str] = part.text
except AttributeError:
text = None
if text:
if not content:
content = text
elif isinstance(content, str):
content = [content, text]
elif isinstance(content, list):
content.append(text)
else:
raise Exception("Unexpected content type")
if part.function_call:
if "function_call" in additional_kwargs:
logger.warning(
(
"This model can reply with multiple "
"function calls in one response. "
"Please don't rely on `additional_kwargs.function_call` "
"as only the last one will be saved."
"Use `tool_calls` instead."
)
)
function_call = {"name": part.function_call.name}
# dump to match other function calling llm for now
function_call_args_dict = proto.Message.to_dict(part.function_call)["args"]
function_call["arguments"] = json.dumps(
{k: function_call_args_dict[k] for k in function_call_args_dict}
)
additional_kwargs["function_call"] = function_call
if streaming:
index = function_call.get("index")
tool_call_chunks.append(
tool_call_chunk(
name=function_call.get("name"),
args=function_call.get("arguments"),
id=function_call.get("id", str(uuid.uuid4())),
index=int(index) if index else None,
)
)
else:
try:
tool_calls_dicts = parse_tool_calls(
[{"function": function_call}],
return_id=False,
)
tool_calls.extend(
[
create_tool_call(
name=tool_call["name"],
args=tool_call["args"],
id=tool_call.get("id", str(uuid.uuid4())),
)
for tool_call in tool_calls_dicts
]
)
except Exception as e:
invalid_tool_calls.append(
invalid_tool_call(
name=function_call.get("name"),
args=function_call.get("arguments"),
id=function_call.get("id", str(uuid.uuid4())),
error=str(e),
)
)
if content is None:
content = ""
if streaming:
return AIMessageChunk(
content=cast(Union[str, List[Union[str, Dict[Any, Any]]]], content),
additional_kwargs=additional_kwargs,
tool_call_chunks=tool_call_chunks,
)
return AIMessage(
content=cast(Union[str, List[Union[str, Dict[Any, Any]]]], content),
tool_calls=tool_calls,
additional_kwargs=additional_kwargs,
invalid_tool_calls=invalid_tool_calls,
)
def _completion_with_retry(
generation_method: Callable,
*,
max_retries: int,
run_manager: Optional[CallbackManagerForLLMRun] = None,
**kwargs: Any,
) -> Any:
"""Use tenacity to retry the completion call."""
retry_decorator = create_retry_decorator(
max_retries=max_retries, run_manager=run_manager
)
@retry_decorator
def _completion_with_retry_inner(generation_method: Callable, **kwargs: Any) -> Any:
return generation_method(**kwargs)
params = (
{k: v for k, v in kwargs.items() if k in _allowed_params_prediction_service}
if kwargs.get("is_gemini")
else kwargs
)
return _completion_with_retry_inner(
generation_method,
**params,
)
async def _acompletion_with_retry(
generation_method: Callable,
*,
max_retries: int,
run_manager: Optional[CallbackManagerForLLMRun] = None,
**kwargs: Any,
) -> Any:
"""Use tenacity to retry the completion call."""
retry_decorator = create_retry_decorator(
max_retries=max_retries, run_manager=run_manager
)
@retry_decorator
async def _completion_with_retry_inner(
generation_method: Callable, **kwargs: Any
) -> Any:
return await generation_method(**kwargs)
params = (
{k: v for k, v in kwargs.items() if k in _allowed_params_prediction_service}
if kwargs.get("is_gemini")
else kwargs
)
return await _completion_with_retry_inner(
generation_method,
**params,
)
[docs]
class ChatVertexAI(_VertexAICommon, BaseChatModel):
"""Google Cloud Vertex AI chat model integration.
Setup:
You must have the langchain-google-vertexai Python package installed
.. code-block:: bash
pip install -U langchain-google-vertexai
And either:
- Have credentials configured for your environment (gcloud, workload identity, etc...)
- Store the path to a service account JSON file as the GOOGLE_APPLICATION_CREDENTIALS environment variable
This codebase uses the google.auth library which first looks for the application
credentials variable mentioned above, and then looks for system-level auth.
For more information, see:
https://cloud.google.com/docs/authentication/application-default-credentials#GAC
and https://googleapis.dev/python/google-auth/latest/reference/google.auth.html#module-google.auth.
Key init args — completion params:
model: str
Name of ChatVertexAI model to use. e.g. "gemini-1.5-flash-001",
"gemini-1.5-pro-001", etc.
temperature: Optional[float]
Sampling temperature.
seed: Optional[int]
Sampling integer to use.
max_tokens: Optional[int]
Max number of tokens to generate.
stop: Optional[List[str]]
Default stop sequences.
safety_settings: Optional[Dict[vertexai.generative_models.HarmCategory, vertexai.generative_models.HarmBlockThreshold]]
The default safety settings to use for all generations.
Key init args — client params:
max_retries: int
Max number of retries.
credentials: Optional[google.auth.credentials.Credentials]
The default custom credentials to use when making API calls. If not
provided, credentials will be ascertained from the environment.
project: Optional[str]
The default GCP project to use when making Vertex API calls.
location: str = "us-central1"
The default location to use when making API calls.
request_parallelism: int = 5
The amount of parallelism allowed for requests issued to VertexAI models.
Default is 5.
base_url: Optional[str]
Base URL for API requests.
See full list of supported init args and their descriptions in the params section.
Instantiate:
.. code-block:: python
from langchain_google_vertexai import ChatVertexAI
llm = ChatVertexAI(
model="gemini-1.5-flash-001",
temperature=0,
max_tokens=None,
max_retries=6,
stop=None,
# other params...
)
Invoke:
.. code-block:: python
messages = [
("system", "You are a helpful translator. Translate the user sentence to French."),
("human", "I love programming."),
]
llm.invoke(messages)
.. code-block:: python
AIMessage(content="J'adore programmer. \n", response_metadata={'is_blocked': False, 'safety_ratings': [{'category': 'HARM_CATEGORY_HATE_SPEECH', 'probability_label': 'NEGLIGIBLE', 'blocked': False}, {'category': 'HARM_CATEGORY_DANGEROUS_CONTENT', 'probability_label': 'NEGLIGIBLE', 'blocked': False}, {'category': 'HARM_CATEGORY_HARASSMENT', 'probability_label': 'NEGLIGIBLE', 'blocked': False}, {'category': 'HARM_CATEGORY_SEXUALLY_EXPLICIT', 'probability_label': 'NEGLIGIBLE', 'blocked': False}], 'citation_metadata': None, 'usage_metadata': {'prompt_token_count': 17, 'candidates_token_count': 7, 'total_token_count': 24}}, id='run-925ce305-2268-44c4-875f-dde9128520ad-0')
Stream:
.. code-block:: python
for chunk in llm.stream(messages):
print(chunk)
.. code-block:: python
AIMessageChunk(content='J', response_metadata={'is_blocked': False, 'safety_ratings': [], 'citation_metadata': None}, id='run-9df01d73-84d9-42db-9d6b-b1466a019e89')
AIMessageChunk(content="'adore programmer. \n", response_metadata={'is_blocked': False, 'safety_ratings': [{'category': 'HARM_CATEGORY_HATE_SPEECH', 'probability_label': 'NEGLIGIBLE', 'blocked': False}, {'category': 'HARM_CATEGORY_DANGEROUS_CONTENT', 'probability_label': 'NEGLIGIBLE', 'blocked': False}, {'category': 'HARM_CATEGORY_HARASSMENT', 'probability_label': 'NEGLIGIBLE', 'blocked': False}, {'category': 'HARM_CATEGORY_SEXUALLY_EXPLICIT', 'probability_label': 'NEGLIGIBLE', 'blocked': False}], 'citation_metadata': None}, id='run-9df01d73-84d9-42db-9d6b-b1466a019e89')
AIMessageChunk(content='', response_metadata={'is_blocked': False, 'safety_ratings': [], 'citation_metadata': None, 'usage_metadata': {'prompt_token_count': 17, 'candidates_token_count': 7, 'total_token_count': 24}}, id='run-9df01d73-84d9-42db-9d6b-b1466a019e89')
.. code-block:: python
stream = llm.stream(messages)
full = next(stream)
for chunk in stream:
full += chunk
full
.. code-block:: python
AIMessageChunk(content="J'adore programmer. \n", response_metadata={'is_blocked': False, 'safety_ratings': [{'category': 'HARM_CATEGORY_HATE_SPEECH', 'probability_label': 'NEGLIGIBLE', 'blocked': False}, {'category': 'HARM_CATEGORY_DANGEROUS_CONTENT', 'probability_label': 'NEGLIGIBLE', 'blocked': False}, {'category': 'HARM_CATEGORY_HARASSMENT', 'probability_label': 'NEGLIGIBLE', 'blocked': False}, {'category': 'HARM_CATEGORY_SEXUALLY_EXPLICIT', 'probability_label': 'NEGLIGIBLE', 'blocked': False}], 'citation_metadata': None, 'usage_metadata': {'prompt_token_count': 17, 'candidates_token_count': 7, 'total_token_count': 24}}, id='run-b7f7492c-4cb5-42d0-8fc3-dce9b293b0fb')
Async:
.. code-block:: python
await llm.ainvoke(messages)
# stream:
# async for chunk in (await llm.astream(messages))
# batch:
# await llm.abatch([messages])
.. code-block:: python
AIMessage(content="J'adore programmer. \n", response_metadata={'is_blocked': False, 'safety_ratings': [{'category': 'HARM_CATEGORY_HATE_SPEECH', 'probability_label': 'NEGLIGIBLE', 'blocked': False}, {'category': 'HARM_CATEGORY_DANGEROUS_CONTENT', 'probability_label': 'NEGLIGIBLE', 'blocked': False}, {'category': 'HARM_CATEGORY_HARASSMENT', 'probability_label': 'NEGLIGIBLE', 'blocked': False}, {'category': 'HARM_CATEGORY_SEXUALLY_EXPLICIT', 'probability_label': 'NEGLIGIBLE', 'blocked': False}], 'citation_metadata': None, 'usage_metadata': {'prompt_token_count': 17, 'candidates_token_count': 7, 'total_token_count': 24}}, id='run-925ce305-2268-44c4-875f-dde9128520ad-0')
Tool calling:
.. code-block:: python
from pydantic import BaseModel, Field
class GetWeather(BaseModel):
'''Get the current weather in a given location'''
location: str = Field(..., description="The city and state, e.g. San Francisco, CA")
class GetPopulation(BaseModel):
'''Get the current population in a given location'''
location: str = Field(..., description="The city and state, e.g. San Francisco, CA")
llm_with_tools = llm.bind_tools([GetWeather, GetPopulation])
ai_msg = llm_with_tools.invoke("Which city is hotter today and which is bigger: LA or NY?")
ai_msg.tool_calls
.. code-block:: python
[{'name': 'GetWeather',
'args': {'location': 'Los Angeles, CA'},
'id': '2a2401fa-40db-470d-83ce-4e52de910d9e'},
{'name': 'GetWeather',
'args': {'location': 'New York City, NY'},
'id': '96761deb-ab7f-4ef9-b4b4-6d44562fc46e'},
{'name': 'GetPopulation',
'args': {'location': 'Los Angeles, CA'},
'id': '9147d532-abee-43a2-adb5-12f164300484'},
{'name': 'GetPopulation',
'args': {'location': 'New York City, NY'},
'id': 'c43374ea-bde5-49ca-8487-5b83ebeea1e6'}]
See ``ChatVertexAI.bind_tools()`` method for more.
Structured output:
.. code-block:: python
from typing import Optional
from pydantic import BaseModel, Field
class Joke(BaseModel):
'''Joke to tell user.'''
setup: str = Field(description="The setup of the joke")
punchline: str = Field(description="The punchline to the joke")
rating: Optional[int] = Field(description="How funny the joke is, from 1 to 10")
structured_llm = llm.with_structured_output(Joke)
structured_llm.invoke("Tell me a joke about cats")
.. code-block:: python
Joke(setup='What do you call a cat that loves to bowl?', punchline='An alley cat!', rating=None)
See ``ChatVertexAI.with_structured_output()`` for more.
Image input:
.. code-block:: python
import base64
import httpx
from langchain_core.messages import HumanMessage
image_url = "https://upload.wikimedia.org/wikipedia/commons/thumb/d/dd/Gfp-wisconsin-madison-the-nature-boardwalk.jpg/2560px-Gfp-wisconsin-madison-the-nature-boardwalk.jpg"
image_data = base64.b64encode(httpx.get(image_url).content).decode("utf-8")
message = HumanMessage(
content=[
{"type": "text", "text": "describe the weather in this image"},
{
"type": "image_url",
"image_url": {"url": f"data:image/jpeg;base64,{image_data}"},
},
],
)
ai_msg = llm.invoke([message])
ai_msg.content
.. code-block:: python
'The weather in this image appears to be sunny and pleasant. The sky is a bright blue with scattered white clouds, suggesting a clear and mild day. The lush green grass indicates recent rainfall or sufficient moisture. The absence of strong shadows suggests that the sun is high in the sky, possibly late afternoon. Overall, the image conveys a sense of tranquility and warmth, characteristic of a beautiful summer day. \n'
You can also point to GCS files which is faster / more efficient because bytes are transferred back and forth.
.. code-block:: python
llm.invoke(
[
HumanMessage(
[
"What's in the image?",
{
"type": "media",
"file_uri": "gs://cloud-samples-data/generative-ai/image/scones.jpg",
"mime_type": "image/jpeg",
},
]
)
]
).content
.. code-block:: python
'The image is of five blueberry scones arranged on a piece of baking paper. \n\nHere is a list of what is in the picture:\n* **Five blueberry scones:** They are scattered across the parchment paper, dusted with powdered sugar. \n* **Two cups of coffee:** Two white cups with saucers. One appears full, the other partially drunk.\n* **A bowl of blueberries:** A brown bowl is filled with fresh blueberries, placed near the scones.\n* **A spoon:** A silver spoon with the words "Let\'s Jam" rests on the paper.\n* **Pink peonies:** Several pink peonies lie beside the scones, adding a touch of color.\n* **Baking paper:** The scones, cups, bowl, and spoon are arranged on a piece of white baking paper, splattered with purple. The paper is crinkled and sits on a dark surface. \n\nThe image has a rustic and delicious feel, suggesting a cozy and enjoyable breakfast or brunch setting. \n'
Video input:
**NOTE**: Currently only supported for ``gemini-...-vision`` models.
.. code-block:: python
llm = ChatVertexAI(model="gemini-1.0-pro-vision")
llm.invoke(
[
HumanMessage(
[
"What's in the video?",
{
"type": "media",
"file_uri": "gs://cloud-samples-data/video/animals.mp4",
"mime_type": "video/mp4",
},
]
)
]
).content
.. code-block:: python
'The video is about a new feature in Google Photos called "Zoomable Selfies". The feature allows users to take selfies with animals at the zoo. The video shows several examples of people taking selfies with animals, including a tiger, an elephant, and a sea otter. The video also shows how the feature works. Users simply need to open the Google Photos app and select the "Zoomable Selfies" option. Then, they need to choose an animal from the list of available animals. The app will then guide the user through the process of taking the selfie.'
Audio input:
.. code-block:: python
from langchain_core.messages import HumanMessage
llm = ChatVertexAI(model="gemini-1.5-flash-001")
llm.invoke(
[
HumanMessage(
[
"What's this audio about?",
{
"type": "media",
"file_uri": "gs://cloud-samples-data/generative-ai/audio/pixel.mp3",
"mime_type": "audio/mpeg",
},
]
)
]
).content
.. code-block:: python
"This audio is an interview with two product managers from Google who work on Pixel feature drops. They discuss how feature drops are important for showcasing how Google devices are constantly improving and getting better. They also discuss some of the highlights of the January feature drop and the new features coming in the March drop for Pixel phones and Pixel watches. The interview concludes with discussion of how user feedback is extremely important to them in deciding which features to include in the feature drops. "
Token usage:
.. code-block:: python
ai_msg = llm.invoke(messages)
ai_msg.usage_metadata
.. code-block:: python
{'input_tokens': 17, 'output_tokens': 7, 'total_tokens': 24}
Logprobs:
.. code-block:: python
llm = ChatVertexAI(model="gemini-1.5-flash-001", logprobs=True)
ai_msg = llm.invoke(messages)
ai_msg.response_metadata["logprobs_result"]
.. code-block:: python
[
{'token': 'J', 'logprob': -1.549651415189146e-06, 'top_logprobs': []},
{'token': "'", 'logprob': -1.549651415189146e-06, 'top_logprobs': []},
{'token': 'adore', 'logprob': 0.0, 'top_logprobs': []},
{'token': ' programmer', 'logprob': -1.1922384146600962e-07, 'top_logprobs': []},
{'token': '.', 'logprob': -4.827636439586058e-05, 'top_logprobs': []},
{'token': ' ', 'logprob': -0.018011733889579773, 'top_logprobs': []},
{'token': '\n', 'logprob': -0.0008687592926435173, 'top_logprobs': []}
]
Response metadata
.. code-block:: python
ai_msg = llm.invoke(messages)
ai_msg.response_metadata
.. code-block:: python
{'is_blocked': False,
'safety_ratings': [{'category': 'HARM_CATEGORY_HATE_SPEECH',
'probability_label': 'NEGLIGIBLE',
'blocked': False},
{'category': 'HARM_CATEGORY_DANGEROUS_CONTENT',
'probability_label': 'NEGLIGIBLE',
'blocked': False},
{'category': 'HARM_CATEGORY_HARASSMENT',
'probability_label': 'NEGLIGIBLE',
'blocked': False},
{'category': 'HARM_CATEGORY_SEXUALLY_EXPLICIT',
'probability_label': 'NEGLIGIBLE',
'blocked': False}],
'usage_metadata': {'prompt_token_count': 17,
'candidates_token_count': 7,
'total_token_count': 24}}
Safety settings
.. code-block:: python
from langchain_google_vertexai import HarmBlockThreshold, HarmCategory
llm = ChatVertexAI(
model="gemini-1.5-pro",
safety_settings={
HarmCategory.HARM_CATEGORY_HATE_SPEECH: HarmBlockThreshold.BLOCK_LOW_AND_ABOVE,
HarmCategory.HARM_CATEGORY_DANGEROUS_CONTENT: HarmBlockThreshold.BLOCK_MEDIUM_AND_ABOVE,
HarmCategory.HARM_CATEGORY_HARASSMENT: HarmBlockThreshold.BLOCK_LOW_AND_ABOVE,
HarmCategory.HARM_CATEGORY_SEXUALLY_EXPLICIT: HarmBlockThreshold.BLOCK_ONLY_HIGH,
},
)
llm.invoke(messages).response_metadata
.. code-block:: python
{'is_blocked': False,
'safety_ratings': [{'category': 'HARM_CATEGORY_HATE_SPEECH',
'probability_label': 'NEGLIGIBLE',
'blocked': False},
{'category': 'HARM_CATEGORY_DANGEROUS_CONTENT',
'probability_label': 'NEGLIGIBLE',
'blocked': False},
{'category': 'HARM_CATEGORY_HARASSMENT',
'probability_label': 'NEGLIGIBLE',
'blocked': False},
{'category': 'HARM_CATEGORY_SEXUALLY_EXPLICIT',
'probability_label': 'NEGLIGIBLE',
'blocked': False}],
'usage_metadata': {'prompt_token_count': 17,
'candidates_token_count': 7,
'total_token_count': 24}}
""" # noqa: E501
model_name: str = Field(default="chat-bison-default", alias="model")
"Underlying model name."
examples: Optional[List[BaseMessage]] = None
convert_system_message_to_human: bool = False
"""[Deprecated] Since new Gemini models support setting a System Message,
setting this parameter to True is discouraged.
"""
response_mime_type: Optional[str] = None
"""Optional. Output response mimetype of the generated candidate text. Only
supported in Gemini 1.5 and later models. Supported mimetype:
* "text/plain": (default) Text output.
* "application/json": JSON response in the candidates.
* "text/x.enum": Enum in plain text.
The model also needs to be prompted to output the appropriate response
type, otherwise the behavior is undefined. This is a preview feature.
"""
response_schema: Optional[Dict[str, Any]] = None
""" Optional. Enforce an schema to the output.
The format of the dictionary should follow Open API schema.
"""
cached_content: Optional[str] = None
""" Optional. Use the model in cache mode. Only supported in Gemini 1.5 and later
models. Must be a string containing the cache name (A sequence of numbers)
"""
logprobs: Union[bool, int] = False
"""Whether to return logprobs as part of AIMessage.response_metadata.
If False, don't return logprobs. If True, return logprobs for top candidate.
If int, return logprobs for top ``logprobs`` candidates.
**NOTE**: As of 10.28.24 this is only supported for gemini-1.5-flash models.
.. versionadded: 2.0.6
"""
def __init__(self, *, model_name: Optional[str] = None, **kwargs: Any) -> None:
"""Needed for mypy typing to recognize model_name as a valid arg."""
if model_name:
kwargs["model_name"] = model_name
super().__init__(**kwargs)
model_config = ConfigDict(
populate_by_name=True,
arbitrary_types_allowed=True,
)
@classmethod
def is_lc_serializable(self) -> bool:
return True
@classmethod
def get_lc_namespace(cls) -> List[str]:
"""Get the namespace of the langchain object."""
return ["langchain", "chat_models", "vertexai"]
@model_validator(mode="after")
def validate_environment(self) -> Self:
"""Validate that the python package exists in environment."""
safety_settings = self.safety_settings
tuned_model_name = self.tuned_model_name
self.model_family = GoogleModelFamily(self.model_name)
if self.model_name == "chat-bison-default":
logger.warning(
"Model_name will become a required arg for VertexAIEmbeddings "
"starting from Sep-01-2024. Currently the default is set to "
"chat-bison"
)
self.model_name = "chat-bison"
if self.full_model_name is not None:
pass
elif self.tuned_model_name is not None:
self.full_model_name = _format_model_name(
self.tuned_model_name,
location=self.location,
project=cast(str, self.project),
)
else:
self.full_model_name = _format_model_name(
self.model_name,
location=self.location,
project=cast(str, self.project),
)
if safety_settings and not is_gemini_model(self.model_family):
raise ValueError("Safety settings are only supported for Gemini models")
if tuned_model_name:
generative_model_name = self.tuned_model_name
else:
generative_model_name = self.model_name
if not is_gemini_model(self.model_family):
logger.warning(
"Non-Gemini models are deprecated. "
"They will be remoced starting from Dec-01-2024. "
)
values = {
"project": self.project,
"location": self.location,
"credentials": self.credentials,
"api_transport": self.api_transport,
"api_endpoint": self.api_endpoint,
"default_metadata": self.default_metadata,
}
self._init_vertexai(values)
if self.model_family == GoogleModelFamily.CODEY:
model_cls = CodeChatModel
model_cls_preview = PreviewCodeChatModel
else:
model_cls = ChatModel
model_cls_preview = PreviewChatModel
self.client = model_cls.from_pretrained(generative_model_name)
self.client_preview = model_cls_preview.from_pretrained(
generative_model_name
)
return self
@property
def _is_gemini_advanced(self) -> bool:
return self.model_family == GoogleModelFamily.GEMINI_ADVANCED
@property
def _default_params(self) -> Dict[str, Any]:
updated_params = super()._default_params
if self.response_mime_type is not None:
updated_params["response_mime_type"] = self.response_mime_type
if self.response_schema is not None:
allowed_mime_types = ("application/json", "text/x.enum")
if self.response_mime_type not in allowed_mime_types:
error_message = (
"`response_schema` is only supported when "
f"`response_mime_type` is set to one of {allowed_mime_types}"
)
raise ValueError(error_message)
gapic_response_schema = _convert_schema_dict_to_gapic(self.response_schema)
updated_params["response_schema"] = gapic_response_schema
return updated_params
def _get_ls_params(
self, stop: Optional[List[str]] = None, **kwargs: Any
) -> LangSmithParams:
"""Get standard params for tracing."""
params = self._prepare_params(stop=stop, **kwargs)
ls_params = LangSmithParams(
ls_provider="google_vertexai",
ls_model_name=self.model_name,
ls_model_type="chat",
ls_temperature=params.get("temperature", self.temperature),
)
if ls_max_tokens := params.get("max_output_tokens", self.max_output_tokens):
ls_params["ls_max_tokens"] = ls_max_tokens
if ls_stop := stop or params.get("stop", None) or self.stop:
ls_params["ls_stop"] = ls_stop
return ls_params
def _generate(
self,
messages: List[BaseMessage],
stop: Optional[List[str]] = None,
run_manager: Optional[CallbackManagerForLLMRun] = None,
stream: Optional[bool] = None,
**kwargs: Any,
) -> ChatResult:
"""Generate next turn in the conversation.
Args:
messages: The history of the conversation as a list of messages. Code chat
does not support context.
stop: The list of stop words (optional).
run_manager: The CallbackManager for LLM run, it's not used at the moment.
stream: Whether to use the streaming endpoint.
Returns:
The ChatResult that contains outputs generated by the model.
Raises:
ValueError: if the last message in the list is not from human.
"""
if stream is True or (stream is None and self.streaming):
stream_iter = self._stream(
messages, stop=stop, run_manager=run_manager, **kwargs
)
return generate_from_stream(stream_iter)
if not self._is_gemini_model:
return self._generate_non_gemini(messages, stop=stop, **kwargs)
return self._generate_gemini(
messages=messages,
stop=stop,
run_manager=run_manager,
is_gemini=True,
**kwargs,
)
def _generation_config_gemini(
self,
stop: Optional[List[str]] = None,
stream: bool = False,
*,
logprobs: int | bool = False,
**kwargs: Any,
) -> GenerationConfig:
"""Prepares GenerationConfig part of the request.
https://cloud.google.com/vertex-ai/docs/reference/rpc/google.cloud.aiplatform.v1beta1#generationconfig
"""
if logprobs and isinstance(logprobs, bool):
kwargs["response_logprobs"] = logprobs
elif logprobs and isinstance(logprobs, int):
kwargs["response_logprobs"] = True
kwargs["logprobs"] = logprobs
else:
pass
return GenerationConfig(
**self._prepare_params(
stop=stop,
stream=stream,
**{k: v for k, v in kwargs.items() if k in _allowed_params},
)
)
def _safety_settings_gemini(
self, safety_settings: Optional[SafetySettingsType]
) -> Optional[Sequence[SafetySetting]]:
"""Prepares SafetySetting part of the request.
https://cloud.google.com/vertex-ai/docs/reference/rpc/google.cloud.aiplatform.v1beta1#safetysetting
"""
if safety_settings is None:
if self.safety_settings:
return self._safety_settings_gemini(self.safety_settings)
return None
if isinstance(safety_settings, list):
return safety_settings
if isinstance(safety_settings, dict):
formatted_safety_settings = []
for category, threshold in safety_settings.items():
if isinstance(category, str):
category = HarmCategory[category] # type: ignore[misc]
if isinstance(threshold, str):
threshold = SafetySetting.HarmBlockThreshold[threshold] # type: ignore[misc]
formatted_safety_settings.append(
SafetySetting(
category=HarmCategory(category),
threshold=SafetySetting.HarmBlockThreshold(threshold),
)
)
return formatted_safety_settings
raise ValueError("safety_settings should be either")
def _prepare_request_gemini(
self,
messages: List[BaseMessage],
stop: Optional[List[str]] = None,
stream: bool = False,
tools: Optional[_ToolsType] = None,
functions: Optional[_ToolsType] = None,
tool_config: Optional[Union[_ToolConfigDict, ToolConfig]] = None,
safety_settings: Optional[SafetySettingsType] = None,
cached_content: Optional[str] = None,
*,
tool_choice: Optional[_ToolChoiceType] = None,
logprobs: Optional[Union[int, bool]] = None,
**kwargs,
) -> GenerateContentRequest:
system_instruction, contents = _parse_chat_history_gemini(messages)
formatted_tools = self._tools_gemini(tools=tools, functions=functions)
if tool_config:
tool_config = self._tool_config_gemini(tool_config=tool_config)
elif tool_choice:
all_names = [
f.name
for tool in (formatted_tools or [])
for f in tool.function_declarations
]
tool_config = _tool_choice_to_tool_config(tool_choice, all_names)
else:
pass
safety_settings = self._safety_settings_gemini(safety_settings)
logprobs = logprobs if logprobs is not None else self.logprobs
logprobs = logprobs if isinstance(logprobs, (int, bool)) else False
generation_config = self._generation_config_gemini(
stream=stream, stop=stop, logprobs=logprobs, **kwargs
)
if (self.cached_content is not None) or (cached_content is not None):
selected_cached_content = self.cached_content or cached_content
return self._request_from_cached_content(
cached_content=selected_cached_content, # type: ignore
contents=contents,
system_instruction=system_instruction,
tools=formatted_tools,
tool_config=tool_config,
safety_settings=safety_settings,
generation_config=generation_config,
model=self.full_model_name,
)
return GenerateContentRequest(
contents=contents,
system_instruction=system_instruction,
tools=formatted_tools,
tool_config=tool_config,
safety_settings=safety_settings,
generation_config=generation_config,
model=self.full_model_name,
)
def _request_from_cached_content(
self,
cached_content: str,
system_instruction: Optional[Content],
tools: Optional[Sequence[GapicTool]],
tool_config: Optional[Union[_ToolConfigDict, ToolConfig]],
contents: list[Content],
safety_settings: Optional[Sequence[SafetySetting]],
generation_config: GenerationConfig,
model: Optional[str],
) -> GenerateContentRequest:
not_allowed_parameters = [
("system_instructions", system_instruction),
("tools", tools),
("tool_config", tool_config),
]
for param_name, parameter in not_allowed_parameters:
if parameter:
message = (
f"Using cached content. Parameter `{param_name}` will be ignored. "
)
logger.warning(message)
full_cache_name = (
f"projects/{self.project}/locations/{self.location}/"
f"cachedContents/{cached_content}"
)
return GenerateContentRequest(
contents=contents,
model=model,
safety_settings=safety_settings,
generation_config=generation_config,
cached_content=full_cache_name,
)
def _generate_gemini(
self,
messages: List[BaseMessage],
stop: Optional[List[str]] = None,
run_manager: Optional[CallbackManagerForLLMRun] = None,
**kwargs: Any,
) -> ChatResult:
request = self._prepare_request_gemini(messages=messages, stop=stop, **kwargs)
response = _completion_with_retry(
self.prediction_client.generate_content,
max_retries=self.max_retries,
request=request,
metadata=self.default_metadata,
**kwargs,
)
return self._gemini_response_to_chat_result(response)
async def _agenerate_gemini(
self,
messages: List[BaseMessage],
stop: Optional[List[str]] = None,
run_manager: Optional[AsyncCallbackManagerForLLMRun] = None,
**kwargs: Any,
) -> ChatResult:
response = await _acompletion_with_retry(
self.async_prediction_client.generate_content,
max_retries=self.max_retries,
request=self._prepare_request_gemini(
messages=messages, stop=stop, **kwargs
),
is_gemini=True,
metadata=self.default_metadata,
**kwargs,
)
return self._gemini_response_to_chat_result(response)
[docs]
def get_num_tokens(self, text: str) -> int:
"""Get the number of tokens present in the text."""
if self._is_gemini_model:
# https://cloud.google.com/vertex-ai/docs/reference/rpc/google.cloud.aiplatform.v1beta1#counttokensrequest
_, contents = _parse_chat_history_gemini([HumanMessage(content=text)])
response = self.prediction_client.count_tokens(
{
"endpoint": self.full_model_name,
"model": self.full_model_name,
"contents": contents,
}
)
return response.total_tokens
else:
return super().get_num_tokens(text=text)
def _tools_gemini(
self,
tools: Optional[_ToolsType] = None,
functions: Optional[_ToolsType] = None,
) -> Optional[List[GapicTool]]:
if tools and functions:
logger.warning(
"Binding tools and functions together is not supported.",
"Only tools will be used",
)
if tools:
return [_format_to_gapic_tool(tools)]
if functions:
return [_format_to_gapic_tool(functions)]
return None
def _tool_config_gemini(
self, tool_config: Optional[Union[_ToolConfigDict, ToolConfig]] = None
) -> Optional[GapicToolConfig]:
if tool_config and not isinstance(tool_config, ToolConfig):
return _format_tool_config(cast(_ToolConfigDict, tool_config))
return None
def _generate_non_gemini(
self,
messages: List[BaseMessage],
stop: Optional[List[str]] = None,
**kwargs: Any,
) -> ChatResult:
kwargs.pop("safety_settings", None)
params = self._prepare_params(stop=stop, stream=False, **kwargs)
question = _get_question(messages)
history = _parse_chat_history(messages[:-1])
examples = kwargs.get("examples") or self.examples
msg_params = {}
if "candidate_count" in params:
msg_params["candidate_count"] = params.pop("candidate_count")
if examples:
params["examples"] = _parse_examples(examples)
with telemetry.tool_context_manager(self._user_agent):
chat = self._start_chat(history, **params)
response = _completion_with_retry(
chat.send_message,
max_retries=self.max_retries,
message=question.content,
**msg_params,
)
usage_metadata = response.raw_prediction_response.metadata
lc_usage = _get_usage_metadata_non_gemini(usage_metadata)
generations = [
ChatGeneration(
message=AIMessage(content=candidate.text, usage_metadata=lc_usage),
generation_info=get_generation_info(
candidate,
self._is_gemini_model,
usage_metadata=usage_metadata,
),
)
for candidate in response.candidates
]
return ChatResult(generations=generations)
async def _agenerate(
self,
messages: List[BaseMessage],
stop: Optional[List[str]] = None,
run_manager: Optional[AsyncCallbackManagerForLLMRun] = None,
stream: Optional[bool] = None,
**kwargs: Any,
) -> ChatResult:
"""Asynchronously generate next turn in the conversation.
Args:
messages: The history of the conversation as a list of messages. Code chat
does not support context.
stop: The list of stop words (optional).
run_manager: The CallbackManager for LLM run, it's not used at the moment.
Returns:
The ChatResult that contains outputs generated by the model.
Raises:
ValueError: if the last message in the list is not from human.
"""
should_stream = stream is True or (stream is None and self.streaming)
if not self._is_gemini_model:
if should_stream:
logger.warning(
"ChatVertexAI does not currently support async streaming."
)
return await self._agenerate_non_gemini(messages, stop=stop, **kwargs)
if should_stream:
stream_iter = self._astream(
messages, stop=stop, run_manager=run_manager, **kwargs
)
return await agenerate_from_stream(stream_iter)
return await self._agenerate_gemini(
messages=messages,
stop=stop,
run_manager=run_manager,
**kwargs,
)
async def _agenerate_non_gemini(
self,
messages: List[BaseMessage],
stop: Optional[List[str]] = None,
**kwargs: Any,
) -> ChatResult:
kwargs.pop("safety_settings", None)
params = self._prepare_params(stop=stop, stream=False, **kwargs)
question = _get_question(messages)
history = _parse_chat_history(messages[:-1])
examples = kwargs.get("examples") or self.examples
msg_params = {}
if "candidate_count" in params:
msg_params["candidate_count"] = params.pop("candidate_count")
if examples:
params["examples"] = _parse_examples(examples)
with telemetry.tool_context_manager(self._user_agent):
chat = self._start_chat(history, **params)
response = await _acompletion_with_retry(
chat.send_message_async,
message=question.content,
max_retries=self.max_retries,
**msg_params,
)
usage_metadata = response.raw_prediction_response.metadata
lc_usage = _get_usage_metadata_non_gemini(usage_metadata)
generations = [
ChatGeneration(
message=AIMessage(content=candidate.text, usage_metadata=lc_usage),
generation_info=get_generation_info(
candidate,
self._is_gemini_model,
usage_metadata=usage_metadata,
),
)
for candidate in response.candidates
]
return ChatResult(generations=generations)
def _stream(
self,
messages: List[BaseMessage],
stop: Optional[List[str]] = None,
run_manager: Optional[CallbackManagerForLLMRun] = None,
**kwargs: Any,
) -> Iterator[ChatGenerationChunk]:
if not self._is_gemini_model:
yield from self._stream_non_gemini(
messages, stop=stop, run_manager=run_manager, **kwargs
)
return
yield from self._stream_gemini(
messages=messages, stop=stop, run_manager=run_manager, **kwargs
)
return
def _stream_gemini(
self,
messages: List[BaseMessage],
stop: Optional[List[str]] = None,
run_manager: Optional[CallbackManagerForLLMRun] = None,
**kwargs: Any,
) -> Iterator[ChatGenerationChunk]:
request = self._prepare_request_gemini(messages=messages, stop=stop, **kwargs)
response_iter = _completion_with_retry(
self.prediction_client.stream_generate_content,
max_retries=self.max_retries,
request=request,
is_gemini=True,
metadata=self.default_metadata,
**kwargs,
)
total_lc_usage = None
for response_chunk in response_iter:
chunk, total_lc_usage = self._gemini_chunk_to_generation_chunk(
response_chunk, prev_total_usage=total_lc_usage
)
if run_manager and isinstance(chunk.message.content, str):
run_manager.on_llm_new_token(chunk.message.content)
yield chunk
def _stream_non_gemini(
self,
messages: List[BaseMessage],
stop: Optional[List[str]] = None,
run_manager: Optional[CallbackManagerForLLMRun] = None,
**kwargs: Any,
) -> Iterator[ChatGenerationChunk]:
params = self._prepare_params(stop=stop, stream=True, **kwargs)
question = _get_question(messages)
history = _parse_chat_history(messages[:-1])
examples = kwargs.get("examples", None)
if examples:
params["examples"] = _parse_examples(examples)
with telemetry.tool_context_manager(self._user_agent):
chat = self._start_chat(history, **params)
responses = chat.send_message_streaming(question.content, **params)
for response in responses:
if run_manager:
run_manager.on_llm_new_token(response.text)
yield ChatGenerationChunk(
message=AIMessageChunk(content=response.text),
generation_info=get_generation_info(
response,
self._is_gemini_model,
usage_metadata=response.raw_prediction_response.metadata,
),
)
async def _astream(
self,
messages: List[BaseMessage],
stop: Optional[List[str]] = None,
run_manager: Optional[AsyncCallbackManagerForLLMRun] = None,
**kwargs: Any,
) -> AsyncIterator[ChatGenerationChunk]:
# TODO: Update to properly support async streaming from gemini.
if not self._is_gemini_model:
async for chunk in super()._astream(
messages, stop=stop, run_manager=run_manager, **kwargs
):
yield chunk
return
request = self._prepare_request_gemini(messages=messages, stop=stop, **kwargs)
response_iter = _acompletion_with_retry(
self.async_prediction_client.stream_generate_content,
max_retries=self.max_retries,
request=request,
is_gemini=True,
**kwargs,
)
total_lc_usage = None
async for response_chunk in await response_iter:
chunk, total_lc_usage = self._gemini_chunk_to_generation_chunk(
response_chunk, prev_total_usage=total_lc_usage
)
if run_manager and isinstance(chunk.message.content, str):
await run_manager.on_llm_new_token(chunk.message.content)
yield chunk
[docs]
def with_structured_output(
self,
schema: Union[Dict, Type[BaseModel]],
*,
include_raw: bool = False,
method: Optional[Literal["json_mode"]] = None,
**kwargs: Any,
) -> Runnable[LanguageModelInput, Union[Dict, BaseModel]]:
"""Model wrapper that returns outputs formatted to match the given schema.
.. versionchanged:: 1.1.0
Return type corrected in version 1.1.0. Previously if a dict schema
was provided then the output had the form
``[{"args": {}, "name": "schema_name"}]`` where the output was a list with
a single dict and the "args" of the one dict corresponded to the schema.
As of `1.1.0` this has been fixed so that the schema (the value
corresponding to the old "args" key) is returned directly.
Args:
schema: The output schema as a dict or a Pydantic class. If a Pydantic class
then the model output will be an object of that class. If a dict then
the model output will be a dict. With a Pydantic class the returned
attributes will be validated, whereas with a dict they will not be. If
`method` is "function_calling" and `schema` is a dict, then the dict
must match the OpenAI function-calling spec.
include_raw: If False then only the parsed structured output is returned. If
an error occurs during model output parsing it will be raised. If True
then both the raw model response (a BaseMessage) and the parsed model
response will be returned. If an error occurs during output parsing it
will be caught and returned as well. The final output is always a dict
with keys "raw", "parsed", and "parsing_error".
method: If set to 'json_schema' it will use controlled genetration to
generate the response rather than function calling. Does not work with
schemas with references or Pydantic models with self-references.
Returns:
A Runnable that takes any ChatModel input. If include_raw is True then a
dict with keys — raw: BaseMessage, parsed: Optional[_DictOrPydantic],
parsing_error: Optional[BaseException]. If include_raw is False then just
_DictOrPydantic is returned, where _DictOrPydantic depends on the schema.
If schema is a Pydantic class then _DictOrPydantic is the Pydantic class.
If schema is a dict then _DictOrPydantic is a dict.
Example: Pydantic schema, exclude raw:
.. code-block:: python
from pydantic import BaseModel
from langchain_google_vertexai import ChatVertexAI
class AnswerWithJustification(BaseModel):
'''An answer to the user question along with justification for the answer.'''
answer: str
justification: str
llm = ChatVertexAI(model_name="gemini-pro", temperature=0)
structured_llm = llm.with_structured_output(AnswerWithJustification)
structured_llm.invoke("What weighs more a pound of bricks or a pound of feathers")
# -> AnswerWithJustification(
# answer='They weigh the same.', justification='A pound is a pound.'
# )
Example: Pydantic schema, include raw:
.. code-block:: python
from pydantic import BaseModel
from langchain_google_vertexai import ChatVertexAI
class AnswerWithJustification(BaseModel):
'''An answer to the user question along with justification for the answer.'''
answer: str
justification: str
llm = ChatVertexAI(model_name="gemini-pro", temperature=0)
structured_llm = llm.with_structured_output(AnswerWithJustification, include_raw=True)
structured_llm.invoke("What weighs more a pound of bricks or a pound of feathers")
# -> {
# 'raw': AIMessage(content='', additional_kwargs={'tool_calls': [{'id': 'call_Ao02pnFYXD6GN1yzc0uXPsvF', 'function': {'arguments': '{"answer":"They weigh the same.","justification":"Both a pound of bricks and a pound of feathers weigh one pound. The weight is the same, but the volume or density of the objects may differ."}', 'name': 'AnswerWithJustification'}, 'type': 'function'}]}),
# 'parsed': AnswerWithJustification(answer='They weigh the same.', justification='Both a pound of bricks and a pound of feathers weigh one pound. The weight is the same, but the volume or density of the objects may differ.'),
# 'parsing_error': None
# }
Example: Dict schema, exclude raw:
.. code-block:: python
from pydantic import BaseModel
from langchain_core.utils.function_calling import convert_to_openai_function
from langchain_google_vertexai import ChatVertexAI
class AnswerWithJustification(BaseModel):
'''An answer to the user question along with justification for the answer.'''
answer: str
justification: str
dict_schema = convert_to_openai_function(AnswerWithJustification)
llm = ChatVertexAI(model_name="gemini-pro", temperature=0)
structured_llm = llm.with_structured_output(dict_schema)
structured_llm.invoke("What weighs more a pound of bricks or a pound of feathers")
# -> {
# 'answer': 'They weigh the same',
# 'justification': 'Both a pound of bricks and a pound of feathers weigh one pound. The weight is the same, but the volume and density of the two substances differ.'
# }
""" # noqa: E501
if kwargs:
raise ValueError(f"Received unsupported arguments {kwargs}")
parser: OutputParserLike
if method == "json_mode":
if isinstance(schema, type):
# TODO: This gets the json schema of a pydantic model. It fails for
# nested models because the generated schema contains $refs that the
# gemini api doesn't support. We can implement a postprocessing function
# that takes care of this if necessary.
schema_json = schema.model_json_schema()
schema_json = replace_defs_in_schema(schema_json)
self.response_schema = schema_json
parser = PydanticOutputParser(pydantic_object=schema)
else:
parser = JsonOutputParser()
self.response_schema = schema
self.response_mime_type = "application/json"
llm: Runnable = self
else:
tool_name = _get_tool_name(schema)
if isinstance(schema, type) and is_basemodel_subclass(schema):
parser = PydanticToolsParser(tools=[schema], first_tool_only=True)
else:
parser = JsonOutputKeyToolsParser(
key_name=tool_name, first_tool_only=True
)
tool_choice = tool_name if self._is_gemini_advanced else None
llm = self.bind_tools([schema], tool_choice=tool_choice)
if include_raw:
parser_with_fallback = RunnablePassthrough.assign(
parsed=itemgetter("raw") | parser, parsing_error=lambda _: None
).with_fallbacks(
[RunnablePassthrough.assign(parsed=lambda _: None)],
exception_key="parsing_error",
)
return {"raw": llm} | parser_with_fallback
else:
return llm | parser
def _start_chat(
self, history: _ChatHistory, **kwargs: Any
) -> Union[ChatSession, CodeChatSession]:
if self.model_family == GoogleModelFamily.CODEY:
return self.client.start_chat(
context=history.context, message_history=history.history, **kwargs
)
else:
return self.client.start_chat(message_history=history.history, **kwargs)
def _gemini_response_to_chat_result(
self, response: GenerationResponse
) -> ChatResult:
generations = []
usage = proto.Message.to_dict(response.usage_metadata)
lc_usage = _get_usage_metadata_gemini(usage)
logprobs = self.logprobs if isinstance(self.logprobs, (int, bool)) else False
for candidate in response.candidates:
info = get_generation_info(
candidate, is_gemini=True, usage_metadata=usage, logprobs=logprobs
)
message = _parse_response_candidate(candidate)
if isinstance(message, AIMessage):
message.usage_metadata = lc_usage
generations.append(ChatGeneration(message=message, generation_info=info))
if not response.candidates:
message = AIMessage(content="")
if usage:
generation_info = {"usage_metadata": usage}
message.usage_metadata = lc_usage
else:
generation_info = {}
generations.append(
ChatGeneration(message=message, generation_info=generation_info)
)
return ChatResult(generations=generations)
def _gemini_chunk_to_generation_chunk(
self,
response_chunk: GenerationResponse,
prev_total_usage: Optional[UsageMetadata] = None,
) -> Tuple[ChatGenerationChunk, Optional[UsageMetadata]]:
# return an empty completion message if there's no candidates
usage_metadata = proto.Message.to_dict(response_chunk.usage_metadata)
# Gather langchain (standard) usage metadata
# Note: some models (e.g., gemini-1.5-pro with image inputs) return
# cumulative sums of token counts.
total_lc_usage = _get_usage_metadata_gemini(usage_metadata)
if total_lc_usage and prev_total_usage:
lc_usage: Optional[UsageMetadata] = UsageMetadata(
input_tokens=total_lc_usage["input_tokens"]
- prev_total_usage["input_tokens"],
output_tokens=total_lc_usage["output_tokens"]
- prev_total_usage["output_tokens"],
total_tokens=total_lc_usage["total_tokens"]
- prev_total_usage["total_tokens"],
)
else:
lc_usage = total_lc_usage
if not response_chunk.candidates:
message = AIMessageChunk(content="")
if lc_usage:
message.usage_metadata = lc_usage
generation_info = {}
else:
top_candidate = response_chunk.candidates[0]
message = _parse_response_candidate(top_candidate, streaming=True)
if lc_usage:
message.usage_metadata = lc_usage
generation_info = get_generation_info(
top_candidate,
is_gemini=True,
)
# is_blocked is part of "safety_ratings" list
# but if it's True/False then chunks can't be marged
generation_info.pop("is_blocked", None)
return ChatGenerationChunk(
message=message,
generation_info=generation_info,
), total_lc_usage
def _get_usage_metadata_gemini(raw_metadata: dict) -> Optional[UsageMetadata]:
"""Get UsageMetadata from raw response metadata."""
input_tokens = raw_metadata.get("prompt_token_count", 0)
output_tokens = raw_metadata.get("candidates_token_count", 0)
total_tokens = raw_metadata.get("total_token_count", 0)
if all(count == 0 for count in [input_tokens, output_tokens, total_tokens]):
return None
else:
return UsageMetadata(
input_tokens=input_tokens,
output_tokens=output_tokens,
total_tokens=total_tokens,
)
def _get_usage_metadata_non_gemini(raw_metadata: dict) -> Optional[UsageMetadata]:
"""Get UsageMetadata from raw response metadata."""
token_usage = raw_metadata.get("tokenMetadata", {})
input_tokens = token_usage.get("inputTokenCount", {}).get("totalTokens", 0)
output_tokens = token_usage.get("outputTokenCount", {}).get("totalTokens", 0)
if input_tokens == 0 and output_tokens == 0:
return None
else:
return UsageMetadata(
input_tokens=input_tokens,
output_tokens=output_tokens,
total_tokens=input_tokens + output_tokens,
)
def _get_tool_name(tool: _ToolType) -> str:
vertexai_tool = _format_to_gapic_tool([tool])
return [f.name for f in vertexai_tool.function_declarations][0]
