The agent engineering platform.
Python agentsaiai-agentsanthropicchatgptdeepagentsenterpriseframeworkgeminigenerative-ailangchainlanggraphllmmultiagentopen-sourceopenaipydanticpythonragtypescript
"""Module contains utility functions for working with messages.
Some examples of what you can do with these functions include:
* Convert messages to strings (serialization)
* Convert messages from dicts to Message objects (deserialization)
* Filter messages from a list of messages based on name, type or id etc.
"""
from __future__ import annotations
import base64
import inspect
import json
import logging
import math
from collections.abc import Callable, Iterable, Sequence
from functools import partial, wraps
from typing import (
TYPE_CHECKING,
Annotated,
Any,
Concatenate,
Literal,
ParamSpec,
Protocol,
TypeVar,
cast,
overload,
)
from xml.sax.saxutils import escape, quoteattr
from pydantic import Discriminator, Field, Tag
from langchain_core.exceptions import ErrorCode, create_message
from langchain_core.messages.ai import AIMessage, AIMessageChunk
from langchain_core.messages.base import BaseMessage, BaseMessageChunk
from langchain_core.messages.block_translators.openai import (
convert_to_openai_data_block,
)
from langchain_core.messages.chat import ChatMessage, ChatMessageChunk
from langchain_core.messages.content import (
is_data_content_block,
)
from langchain_core.messages.function import FunctionMessage, FunctionMessageChunk
from langchain_core.messages.human import HumanMessage, HumanMessageChunk
from langchain_core.messages.modifier import RemoveMessage
from langchain_core.messages.system import SystemMessage, SystemMessageChunk
from langchain_core.messages.tool import ToolCall, ToolMessage, ToolMessageChunk
from langchain_core.utils.pydantic import model_json_schema as get_model_json_schema
if TYPE_CHECKING:
from langchain_core.language_models import BaseLanguageModel
from langchain_core.prompt_values import PromptValue
from langchain_core.runnables.base import Runnable
from langchain_core.tools import BaseTool
try:
from langchain_text_splitters import TextSplitter
_HAS_LANGCHAIN_TEXT_SPLITTERS = True
except ImportError:
_HAS_LANGCHAIN_TEXT_SPLITTERS = False
logger = logging.getLogger(__name__)
def _get_type(v: Any) -> str:
"""Get the type associated with the object for serialization purposes."""
if isinstance(v, dict) and "type" in v:
result = v["type"]
elif hasattr(v, "type"):
result = v.type
else:
msg = (
f"Expected either a dictionary with a 'type' key or an object "
f"with a 'type' attribute. Instead got type {type(v)}."
)
raise TypeError(msg)
if not isinstance(result, str):
msg = f"Expected 'type' to be a str, got {type(result).__name__}"
raise TypeError(msg)
return result
AnyMessage = Annotated[
Annotated[AIMessage, Tag(tag="ai")]
| Annotated[HumanMessage, Tag(tag="human")]
| Annotated[ChatMessage, Tag(tag="chat")]
| Annotated[SystemMessage, Tag(tag="system")]
| Annotated[FunctionMessage, Tag(tag="function")]
| Annotated[ToolMessage, Tag(tag="tool")]
| Annotated[AIMessageChunk, Tag(tag="AIMessageChunk")]
| Annotated[HumanMessageChunk, Tag(tag="HumanMessageChunk")]
| Annotated[ChatMessageChunk, Tag(tag="ChatMessageChunk")]
| Annotated[SystemMessageChunk, Tag(tag="SystemMessageChunk")]
| Annotated[FunctionMessageChunk, Tag(tag="FunctionMessageChunk")]
| Annotated[ToolMessageChunk, Tag(tag="ToolMessageChunk")],
Field(discriminator=Discriminator(_get_type)),
]
"""A type representing any defined `Message` or `MessageChunk` type."""
def _has_base64_data(block: dict[str, Any]) -> bool:
"""Check if a content block contains base64 encoded data.
Args:
block: A content block dictionary.
Returns:
Whether the block contains base64 data.
"""
# Check for explicit base64 field (standard content blocks)
if block.get("base64"):
return True
# Check for data: URL in url field
url = block.get("url", "")
if isinstance(url, str) and url.startswith("data:"):
return True
# Check for OpenAI-style image_url with data: URL
image_url = block.get("image_url", {})
if isinstance(image_url, dict):
url = image_url.get("url", "")
if isinstance(url, str) and url.startswith("data:"):
return True
return False
_XML_CONTENT_BLOCK_MAX_LEN = 500
def _truncate(text: str, max_len: int = _XML_CONTENT_BLOCK_MAX_LEN) -> str:
"""Truncate text to `max_len` characters, adding ellipsis if truncated."""
if len(text) <= max_len:
return text
return text[:max_len] + "..."
def _format_content_block_xml(block: dict[str, Any]) -> str | None:
"""Format a content block as XML.
Args:
block: A LangChain content block.
Returns:
XML string representation of the block, or `None` if the block should be
skipped.
Note:
Plain text document content, server tool call arguments, and server tool
result outputs are truncated to 500 characters.
"""
block_type = block.get("type", "")
# Skip blocks with base64 encoded data
if _has_base64_data(block):
return None
# Text blocks
if block_type == "text":
text = block.get("text", "")
return escape(text) if text else None
# Reasoning blocks
if block_type == "reasoning":
reasoning = block.get("reasoning", "")
if reasoning:
return f"<reasoning>{escape(reasoning)}</reasoning>"
return None
# Image blocks (URL only, base64 already filtered)
if block_type == "image":
url = block.get("url")
file_id = block.get("file_id")
if url:
return f"<image url={quoteattr(url)} />"
if file_id:
return f"<image file_id={quoteattr(file_id)} />"
return None
# OpenAI-style image_url blocks
if block_type == "image_url":
image_url = block.get("image_url", {})
if isinstance(image_url, dict):
url = image_url.get("url", "")
if url and not url.startswith("data:"):
return f"<image url={quoteattr(url)} />"
return None
# Audio blocks (URL only)
if block_type == "audio":
url = block.get("url")
file_id = block.get("file_id")
if url:
return f"<audio url={quoteattr(url)} />"
if file_id:
return f"<audio file_id={quoteattr(file_id)} />"
return None
# Video blocks (URL only)
if block_type == "video":
url = block.get("url")
file_id = block.get("file_id")
if url:
return f"<video url={quoteattr(url)} />"
if file_id:
return f"<video file_id={quoteattr(file_id)} />"
return None
# Plain text document blocks
if block_type == "text-plain":
text = block.get("text", "")
return escape(_truncate(text)) if text else None
# Server tool call blocks (from AI messages)
if block_type == "server_tool_call":
tc_id = quoteattr(str(block.get("id") or ""))
tc_name = quoteattr(str(block.get("name") or ""))
tc_args_json = json.dumps(block.get("args", {}), ensure_ascii=False)
tc_args = escape(_truncate(tc_args_json))
return (
f"<server_tool_call id={tc_id} name={tc_name}>{tc_args}</server_tool_call>"
)
# Server tool result blocks
if block_type == "server_tool_result":
tool_call_id = quoteattr(str(block.get("tool_call_id") or ""))
status = quoteattr(str(block.get("status") or ""))
output = block.get("output")
if output:
output_json = json.dumps(output, ensure_ascii=False)
output_str = escape(_truncate(output_json))
else:
output_str = ""
return (
f"<server_tool_result tool_call_id={tool_call_id} status={status}>"
f"{output_str}</server_tool_result>"
)
# Unknown block type - skip silently
return None
def _get_message_type_str(
m: BaseMessage,
human_prefix: str,
ai_prefix: str,
system_prefix: str,
function_prefix: str,
tool_prefix: str,
) -> str:
"""Get the type string for XML message element.
Args:
m: The message to get the type string for.
human_prefix: The prefix to use for `HumanMessage`.
ai_prefix: The prefix to use for `AIMessage`.
system_prefix: The prefix to use for `SystemMessage`.
function_prefix: The prefix to use for `FunctionMessage`.
tool_prefix: The prefix to use for `ToolMessage`.
Returns:
The type string for the message element.
Raises:
ValueError: If an unsupported message type is encountered.
"""
if isinstance(m, HumanMessage):
return human_prefix.lower()
if isinstance(m, AIMessage):
return ai_prefix.lower()
if isinstance(m, SystemMessage):
return system_prefix.lower()
if isinstance(m, FunctionMessage):
return function_prefix.lower()
if isinstance(m, ToolMessage):
return tool_prefix.lower()
if isinstance(m, ChatMessage):
return m.role
msg = f"Got unsupported message type: {m}"
raise ValueError(msg)
_TOOL_CALL_ID_DISPLAY_LIMIT = 64
def _display_tool_call_id(tool_call_id: str) -> str:
"""Abbreviate long tool-call IDs in formatted message strings."""
if len(tool_call_id) > _TOOL_CALL_ID_DISPLAY_LIMIT:
return f"{tool_call_id[:_TOOL_CALL_ID_DISPLAY_LIMIT]}..."
return tool_call_id
def get_buffer_string(
messages: Sequence[BaseMessage],
human_prefix: str = "Human",
ai_prefix: str = "AI",
*,
system_prefix: str = "System",
function_prefix: str = "Function",
tool_prefix: str = "Tool",
message_separator: str = "\n",
format: Literal["prefix", "xml"] = "prefix", # noqa: A002
) -> str:
r"""Convert a sequence of messages to strings and concatenate them into one string.
Args:
messages: Messages to be converted to strings.
human_prefix: The prefix to prepend to contents of `HumanMessage`s.
ai_prefix: The prefix to prepend to contents of `AIMessage`.
system_prefix: The prefix to prepend to contents of `SystemMessage`s.
function_prefix: The prefix to prepend to contents of `FunctionMessage`s.
tool_prefix: The prefix to prepend to contents of `ToolMessage`s.
message_separator: The separator to use between messages.
format: The output format. `'prefix'` uses `Role: content` format (default).
For multimodal messages, only string content and `text` blocks are
included; non-text blocks such as images, audio, and video are omitted.
`'xml'` uses XML-style `<message type='role'>` format with proper
character escaping, which is useful when message content may
contain role-like prefixes that could cause ambiguity.
Use `'xml'` when you need a structured representation of supported
multimodal content blocks.
Returns:
A single string concatenation of all input messages.
Raises:
ValueError: If an unsupported message type is encountered.
!!! warning
If a message is an `AIMessage` and contains both tool calls under `tool_calls`
and a function call under `additional_kwargs["function_call"]`, only the tool
calls will be appended to the string representation.
In XML format, tool-call IDs longer than 64 characters are displayed
as the first 64 characters followed by `...`; the original messages
are not changed.
!!! note "XML format"
When using `format='xml'`:
- All messages use uniform `<message type="role">content</message>` format.
- The `type` attribute uses `human_prefix` (lowercased) for `HumanMessage`,
`ai_prefix` (lowercased) for `AIMessage`, `system_prefix` (lowercased)
for `SystemMessage`, `function_prefix` (lowercased) for `FunctionMessage`,
`tool_prefix` (lowercased) for `ToolMessage`, and the original role
(unchanged) for `ChatMessage`.
- Message content is escaped using `xml.sax.saxutils.escape()`.
- Attribute values are escaped using `xml.sax.saxutils.quoteattr()`.
- AI messages with tool calls use nested structure with `<content>` and
`<tool_call>` elements.
- For multi-modal content (list of content blocks), supported block types
are: `text`, `reasoning`, `image` (URL/file_id only), `image_url`
(OpenAI-style, URL only), `audio` (URL/file_id only), `video` (URL/file_id
only), `text-plain`, `server_tool_call`, and `server_tool_result`.
- Content blocks with base64-encoded data are skipped (including blocks
with `base64` field or `data:` URLs).
- Unknown block types are skipped.
- Plain text document content (`text-plain`), server tool call arguments,
and server tool result outputs are truncated to 500 characters.
Example:
Default prefix format:
```python
from langchain_core.messages import AIMessage, HumanMessage, get_buffer_string
messages = [
HumanMessage(content="Hi, how are you?"),
AIMessage(content="Good, how are you?"),
]
get_buffer_string(messages)
# -> "Human: Hi, how are you?\nAI: Good, how are you?"
```
XML format (useful when content contains role-like prefixes):
```python
messages = [
HumanMessage(content="Example: Human: some text"),
AIMessage(content="I see the example."),
]
get_buffer_string(messages, format="xml")
# -> '<message type="human">Example: Human: some text</message>\\n'
# -> '<message type="ai">I see the example.</message>'
```
XML format with special characters (automatically escaped):
```python
messages = [
HumanMessage(content="Is 5 < 10 & 10 > 5?"),
]
get_buffer_string(messages, format="xml")
# -> '<message type="human">Is 5 < 10 & 10 > 5?</message>'
```
XML format with tool calls:
```python
messages = [
AIMessage(
content="I'll search for that.",
tool_calls=[
{"id": "call_123", "name": "search", "args": {"query": "weather"}}
],
),
]
get_buffer_string(messages, format="xml")
# -> '<message type="ai">\\n'
# -> ' <content>I\\'ll search for that.</content>\\n'
# -> ' <tool_call id="call_123" name="search">'
# -> '{"query": "weather"}</tool_call>\\n'
# -> '</message>'
```
"""
if format not in {"prefix", "xml"}:
msg = (
f"Unrecognized format={format!r}. Supported formats are 'prefix' and 'xml'."
)
raise ValueError(msg)
string_messages = []
for m in messages:
if isinstance(m, HumanMessage):
role = human_prefix
elif isinstance(m, AIMessage):
role = ai_prefix
elif isinstance(m, SystemMessage):
role = system_prefix
elif isinstance(m, FunctionMessage):
role = function_prefix
elif isinstance(m, ToolMessage):
role = tool_prefix
elif isinstance(m, ChatMessage):
role = m.role
else:
msg = f"Got unsupported message type: {m}"
raise ValueError(msg) # noqa: TRY004
if format == "xml":
msg_type = _get_message_type_str(
m, human_prefix, ai_prefix, system_prefix, function_prefix, tool_prefix
)
# Format content blocks
if isinstance(m.content, str):
content_parts = [escape(m.content)] if m.content else []
else:
# List of content blocks
content_parts = []
for block in m.content:
if isinstance(block, str):
if block:
content_parts.append(escape(block))
else:
formatted = _format_content_block_xml(block)
if formatted:
content_parts.append(formatted)
# Check if this is an AIMessage with tool calls
has_tool_calls = isinstance(m, AIMessage) and m.tool_calls
has_function_call = (
isinstance(m, AIMessage)
and not m.tool_calls
and "function_call" in m.additional_kwargs
)
if has_tool_calls or has_function_call:
# Use nested structure for AI messages with tool calls
# Type narrowing: at this point m is AIMessage (verified above)
ai_msg = cast("AIMessage", m)
parts = [f"<message type={quoteattr(msg_type)}>"]
if content_parts:
parts.append(f" <content>{' '.join(content_parts)}</content>")
if has_tool_calls:
for tc in ai_msg.tool_calls:
tc_id = quoteattr(_display_tool_call_id(tc.get("id") or ""))
tc_name = quoteattr(str(tc.get("name") or ""))
tc_args = escape(
json.dumps(tc.get("args", {}), ensure_ascii=False)
)
parts.append(
f" <tool_call id={tc_id} name={tc_name}>"
f"{tc_args}</tool_call>"
)
elif has_function_call:
fc = ai_msg.additional_kwargs["function_call"]
fc_name = quoteattr(str(fc.get("name") or ""))
fc_args = escape(str(fc.get("arguments") or "{}"))
parts.append(
f" <function_call name={fc_name}>{fc_args}</function_call>"
)
parts.append("</message>")
message = "\n".join(parts)
else:
# Simple structure for messages without tool calls
joined_content = " ".join(content_parts)
message = (
f"<message type={quoteattr(msg_type)}>{joined_content}</message>"
)
else: # format == "prefix"
content = m.text
message = f"{role}: {content}"
tool_info = ""
if isinstance(m, AIMessage):
if m.tool_calls:
tool_info = str(m.tool_calls)
elif "function_call" in m.additional_kwargs:
# Legacy behavior assumes only one function call per message
tool_info = str(m.additional_kwargs["function_call"])
if tool_info:
message += tool_info # Preserve original behavior
string_messages.append(message)
return message_separator.join(string_messages)
def _message_from_dict(message: dict[str, Any]) -> BaseMessage:
type_ = message["type"]
if type_ == "human":
return HumanMessage(**message["data"])
if type_ == "ai":
return AIMessage(**message["data"])
if type_ == "system":
return SystemMessage(**message["data"])
if type_ == "chat":
return ChatMessage(**message["data"])
if type_ == "function":
return FunctionMessage(**message["data"])
if type_ == "tool":
return ToolMessage(**message["data"])
if type_ == "remove":
return RemoveMessage(**message["data"])
if type_ == "AIMessageChunk":
return AIMessageChunk(**message["data"])
if type_ == "HumanMessageChunk":
return HumanMessageChunk(**message["data"])
if type_ == "FunctionMessageChunk":
return FunctionMessageChunk(**message["data"])
if type_ == "ToolMessageChunk":
return ToolMessageChunk(**message["data"])
if type_ == "SystemMessageChunk":
return SystemMessageChunk(**message["data"])
if type_ == "ChatMessageChunk":
return ChatMessageChunk(**message["data"])
msg = f"Got unexpected message type: {type_}"
raise ValueError(msg)
def messages_from_dict(messages: Sequence[dict[str, Any]]) -> list[BaseMessage]:
"""Convert a sequence of messages from dicts to `Message` objects.
Args:
messages: Sequence of messages (as dicts) to convert.
Returns:
list of messages (BaseMessages).
"""
return [_message_from_dict(m) for m in messages]
def message_chunk_to_message(chunk: BaseMessage) -> BaseMessage:
"""Convert a message chunk to a `Message`.
Args:
chunk: Message chunk to convert.
Returns:
Message.
"""
if not isinstance(chunk, BaseMessageChunk):
return chunk
# chunk classes always have the equivalent non-chunk class as their first parent
ignore_keys = ["type"]
if isinstance(chunk, AIMessageChunk):
ignore_keys.extend(["tool_call_chunks", "chunk_position"])
return cast(
"BaseMessage",
chunk.__class__.__mro__[1](
**{k: v for k, v in chunk.__dict__.items() if k not in ignore_keys}
),
)
MessageLikeRepresentation = (
BaseMessage
| list[str]
| tuple[str, str | list[str | dict[str, Any]]]
| str
| dict[str, Any]
)
"""A type representing the various ways a message can be represented."""
def _create_message_from_message_type(
message_type: str,
content: str | list[str | dict[str, Any]],
name: str | None = None,
tool_call_id: str | None = None,
tool_calls: list[dict[str, Any]] | None = None,
id: str | None = None,
**additional_kwargs: Any,
) -> BaseMessage:
"""Create a message from a `Message` type and content string.
Args:
message_type: the type of the message (e.g., `'human'`, `'ai'`, etc.).
content: the content string.
name: the name of the message.
tool_call_id: the tool call id.
tool_calls: the tool calls.
id: the id of the message.
additional_kwargs: additional keyword arguments.
Returns:
a message of the appropriate type.
Raises:
ValueError: if the message type is not one of `'human'`, `'user'`, `'ai'`,
`'assistant'`, `'function'`, `'tool'`, `'system'`, or
`'developer'`.
"""
kwargs: dict[str, Any] = {}
if name is not None:
kwargs["name"] = name
if tool_call_id is not None:
kwargs["tool_call_id"] = tool_call_id
if additional_kwargs:
if response_metadata := additional_kwargs.pop("response_metadata", None):
kwargs["response_metadata"] = response_metadata
kwargs["additional_kwargs"] = additional_kwargs
additional_kwargs.update(additional_kwargs.pop("additional_kwargs", {}))
if id is not None:
kwargs["id"] = id
if tool_calls is not None:
kwargs["tool_calls"] = []
for tool_call in tool_calls:
# Convert OpenAI-format tool call to LangChain format.
if "function" in tool_call:
args = tool_call["function"]["arguments"]
if isinstance(args, str):
args = json.loads(args, strict=False)
kwargs["tool_calls"].append(
{
"name": tool_call["function"]["name"],
"args": args,
"id": tool_call["id"],
"type": "tool_call",
}
)
else:
kwargs["tool_calls"].append(tool_call)
if message_type in {"human", "user"}:
if example := kwargs.get("additional_kwargs", {}).pop("example", False):
kwargs["example"] = example
message: BaseMessage = HumanMessage(content=content, **kwargs)
elif message_type in {"ai", "assistant"}:
if example := kwargs.get("additional_kwargs", {}).pop("example", False):
kwargs["example"] = example
message = AIMessage(content=content, **kwargs)
elif message_type in {"system", "developer"}:
if message_type == "developer":
kwargs["additional_kwargs"] = kwargs.get("additional_kwargs") or {}
kwargs["additional_kwargs"]["__openai_role__"] = "developer"
message = SystemMessage(content=content, **kwargs)
elif message_type == "function":
message = FunctionMessage(content=content, **kwargs)
elif message_type == "tool":
artifact = kwargs.get("additional_kwargs", {}).pop("artifact", None)
status = kwargs.get("additional_kwargs", {}).pop("status", None)
if status is not None:
kwargs["status"] = status
message = ToolMessage(content=content, artifact=artifact, **kwargs)
elif message_type == "remove":
message = RemoveMessage(**kwargs)
else:
msg = (
f"Unexpected message type: '{message_type}'. Use one of 'human',"
f" 'user', 'ai', 'assistant', 'function', 'tool', 'system', or 'developer'."
)
msg = create_message(message=msg, error_code=ErrorCode.MESSAGE_COERCION_FAILURE)
raise ValueError(msg)
return message
# Map of class names emitted in the `Serializable` constructor-envelope
# (`{"lc": 1, "type": "constructor", "id": [..., "<ClassName>"],
# "kwargs": {...}}`) to the message-type strings
# `_create_message_from_message_type` accepts. Read by
# `_convert_to_message`'s dict branch when unpacking that wire shape.
# Kept as a hardcoded allowlist of strings rather than a class registry
# lookup so dispatch never resolves to a class chosen by the caller.
_LC_CONSTRUCTOR_NAME_TO_TYPE: dict[str, str] = {
"HumanMessage": "human",
"HumanMessageChunk": "human",
"AIMessage": "ai",
"AIMessageChunk": "ai",
"SystemMessage": "system",
"SystemMessageChunk": "system",
"FunctionMessage": "function",
"FunctionMessageChunk": "function",
"ToolMessage": "tool",
"ToolMessageChunk": "tool",
"RemoveMessage": "remove",
}
def _convert_to_message(message: MessageLikeRepresentation) -> BaseMessage:
"""Instantiate a `Message` from a variety of message formats.
The message format can be one of the following:
- `BaseMessagePromptTemplate`
- `BaseMessage`
- 2-tuple of (role string, template); e.g., (`'human'`, `'{user_input}'`)
- dict: a message dict with role and content keys
- dict: the `Serializable` constructor-envelope wire shape
`{"lc": 1, "type": "constructor", "id": [..., "<ClassName>"],
"kwargs": {...}}` — unpacked structurally and routed through the
standard dict-with-type dispatch.
- string: shorthand for (`'human'`, template); e.g., `'{user_input}'`
Args:
message: a representation of a message in one of the supported formats.
Returns:
An instance of a message or a message template.
Raises:
NotImplementedError: if the message type is not supported.
ValueError: if the message dict does not contain the required keys.
"""
if isinstance(message, BaseMessage):
message_ = message
elif isinstance(message, Sequence):
if isinstance(message, str):
message_ = _create_message_from_message_type("human", message)
else:
try:
message_type_str, template = message
except ValueError as e:
msg = "Message as a sequence must be (role string, template)"
raise NotImplementedError(msg) from e
message_ = _create_message_from_message_type(message_type_str, template)
elif isinstance(message, dict):
# `Serializable` constructor-envelope wire shape. Detect structurally, map
# the class name to a known message-type string via a hardcoded
# allowlist, and recurse with the canonical
# `{"type": ..., **kwargs}` shape — no `load()`, no dynamic
# class instantiation.
if (
message.get("lc") == 1
and message.get("type") == "constructor"
and isinstance(message.get("id"), list)
and message["id"]
and isinstance(message.get("kwargs"), dict)
):
mapped = _LC_CONSTRUCTOR_NAME_TO_TYPE.get(message["id"][-1])
if mapped is not None:
return _convert_to_message({"type": mapped, **message["kwargs"]})
msg_kwargs = message.copy()
try:
try:
msg_type = msg_kwargs.pop("role")
except KeyError:
msg_type = msg_kwargs.pop("type")
# None msg content is not allowed
msg_content = msg_kwargs.pop("content") or ""
except KeyError as e:
msg = f"Message dict must contain 'role' and 'content' keys, got {message}"
msg = create_message(
message=msg, error_code=ErrorCode.MESSAGE_COERCION_FAILURE
)
raise ValueError(msg) from e
message_ = _create_message_from_message_type(
msg_type, msg_content, **msg_kwargs
)
else:
msg = f"Unsupported message type: {type(message)}" # type: ignore[unreachable]
msg = create_message(message=msg, error_code=ErrorCode.MESSAGE_COERCION_FAILURE)
raise NotImplementedError(msg)
return message_
def convert_to_messages(
messages: Iterable[MessageLikeRepresentation] | PromptValue,
) -> list[BaseMessage]:
"""Convert a sequence of messages to a list of messages.
Args:
messages: Sequence of messages to convert.
Returns:
list of messages (BaseMessages).
"""
# Import here to avoid circular imports
from langchain_core.prompt_values import PromptValue # noqa: PLC0415
if isinstance(messages, PromptValue):
return messages.to_messages()
return [_convert_to_message(m) for m in messages]
_P = ParamSpec("_P")
_R_co = TypeVar("_R_co", covariant=True)
class _RunnableSupportCallable(Protocol[_P, _R_co]):
@overload
def __call__(
self,
messages: None = None,
*args: _P.args,
**kwargs: _P.kwargs,
) -> Runnable[Sequence[MessageLikeRepresentation], _R_co]: ...
@overload
def __call__(
self,
messages: Sequence[MessageLikeRepresentation] | PromptValue,
*args: _P.args,
**kwargs: _P.kwargs,
) -> _R_co: ...
def __call__(
self,
messages: Sequence[MessageLikeRepresentation] | PromptValue | None = None,
*args: _P.args,
**kwargs: _P.kwargs,
) -> _R_co | Runnable[Sequence[MessageLikeRepresentation], _R_co]: ...
def _runnable_support(
func: Callable[
Concatenate[Sequence[MessageLikeRepresentation] | PromptValue, _P], _R_co
],
) -> _RunnableSupportCallable[_P, _R_co]:
@wraps(func)
def wrapped(
messages: Sequence[MessageLikeRepresentation] | PromptValue | None = None,
*args: _P.args,
**kwargs: _P.kwargs,
) -> _R_co | Runnable[Sequence[MessageLikeRepresentation], _R_co]:
# Import locally to prevent circular import.
from langchain_core.runnables.base import RunnableLambda # noqa: PLC0415
if messages is not None:
return func(messages, *args, **kwargs)
return RunnableLambda(partial(func, **kwargs), name=func.__name__)
return cast("_RunnableSupportCallable[_P, _R_co]", wrapped)
@_runnable_support
def filter_messages(
messages: Iterable[MessageLikeRepresentation] | PromptValue,
*,
include_names: Sequence[str] | None = None,
exclude_names: Sequence[str] | None = None,
include_types: Sequence[str | type[BaseMessage]] | None = None,
exclude_types: Sequence[str | type[BaseMessage]] | None = None,
include_ids: Sequence[str] | None = None,
exclude_ids: Sequence[str] | None = None,
exclude_tool_calls: Sequence[str] | bool | None = None,
) -> list[BaseMessage]:
"""Filter messages based on `name`, `type` or `id`.
Args:
messages: Sequence Message-like objects to filter.
include_names: Message names to include.
exclude_names: Messages names to exclude.
include_types: Message types to include. Can be specified as string names
(e.g. `'system'`, `'human'`, `'ai'`, ...) or as `BaseMessage`
classes (e.g. `SystemMessage`, `HumanMessage`, `AIMessage`, ...).
exclude_types: Message types to exclude. Can be specified as string names
(e.g. `'system'`, `'human'`, `'ai'`, ...) or as `BaseMessage`
classes (e.g. `SystemMessage`, `HumanMessage`, `AIMessage`, ...).
include_ids: Message IDs to include.
exclude_ids: Message IDs to exclude.
exclude_tool_calls: Tool call IDs to exclude.
Can be one of the following:
- `True`: All `AIMessage` objects with tool calls and all `ToolMessage`
objects will be excluded.
- a sequence of tool call IDs to exclude:
- `ToolMessage` objects with the corresponding tool call ID will be
excluded.
- The `tool_calls` in the AIMessage will be updated to exclude
matching tool calls. If all `tool_calls` are filtered from an
AIMessage, the whole message is excluded.
Returns:
A list of Messages that meets at least one of the `incl_*` conditions and none
of the `excl_*` conditions. If not `incl_*` conditions are specified then
anything that is not explicitly excluded will be included.
Raises:
ValueError: If two incompatible arguments are provided.
Example:
```python
from langchain_core.messages import (
filter_messages,
AIMessage,
HumanMessage,
SystemMessage,
)
messages = [
SystemMessage("you're a good assistant."),
HumanMessage("what's your name", id="foo", name="example_user"),
AIMessage("steve-o", id="bar", name="example_assistant"),
HumanMessage(
"what's your favorite color",
id="baz",
),
AIMessage(
"silicon blue",
id="blah",
),
]
filter_messages(
messages,
include_names=("example_user", "example_assistant"),
include_types=("system",),
exclude_ids=("bar",),
)
```
```python
[
SystemMessage("you're a good assistant."),
HumanMessage("what's your name", id="foo", name="example_user"),
]
```
"""
messages = convert_to_messages(messages)
filtered: list[BaseMessage] = []
for msg in messages:
if (
(exclude_names and msg.name in exclude_names)
or (exclude_types and _is_message_type(msg, exclude_types))
or (exclude_ids and msg.id in exclude_ids)
):
continue
if exclude_tool_calls is True and (
(isinstance(msg, AIMessage) and msg.tool_calls)
or isinstance(msg, ToolMessage)
):
continue
new_msg = msg
if isinstance(exclude_tool_calls, (list, tuple, set)):
if isinstance(msg, AIMessage) and msg.tool_calls:
tool_calls = [
tool_call
for tool_call in msg.tool_calls
if tool_call["id"] not in exclude_tool_calls
]
if not tool_calls:
continue
content = msg.content
# handle Anthropic content blocks
if isinstance(msg.content, list):
content = [
content_block
for content_block in msg.content
if (
not isinstance(content_block, dict)
or content_block.get("type") != "tool_use"
or content_block.get("id") not in exclude_tool_calls
)
]
new_msg = msg.model_copy(
update={"tool_calls": tool_calls, "content": content}
)
elif (
isinstance(msg, ToolMessage) and msg.tool_call_id in exclude_tool_calls
):
continue
# default to inclusion when no inclusion criteria given.
if (
not (include_types or include_ids or include_names)
or (include_names and new_msg.name in include_names)
or (include_types and _is_message_type(new_msg, include_types))
or (include_ids and new_msg.id in include_ids)
):
filtered.append(new_msg)
return filtered
@_runnable_support
def merge_message_runs(
messages: Iterable[MessageLikeRepresentation] | PromptValue,
*,
chunk_separator: str = "\n",
) -> list[BaseMessage]:
r"""Merge consecutive Messages of the same type.
!!! note
`ToolMessage` objects are not merged, as each has a distinct tool call id that
can't be merged.
Args:
messages: Sequence Message-like objects to merge.
chunk_separator: Specify the string to be inserted between message chunks.
Returns:
list of BaseMessages with consecutive runs of message types merged into single
messages. By default, if two messages being merged both have string contents,
the merged content is a concatenation of the two strings with a new-line
separator.
The separator inserted between message chunks can be controlled by specifying
any string with `chunk_separator`. If at least one of the messages has a list
of content blocks, the merged content is a list of content blocks.
Example:
```python
from langchain_core.messages import (
merge_message_runs,
AIMessage,
HumanMessage,
SystemMessage,
ToolCall,
)
messages = [
SystemMessage("you're a good assistant."),
HumanMessage(
"what's your favorite color",
id="foo",
),
HumanMessage(
"wait your favorite food",
id="bar",
),
AIMessage(
"my favorite colo",
tool_calls=[
ToolCall(
name="blah_tool", args={"x": 2}, id="123", type="tool_call"
)
],
id="baz",
),
AIMessage(
[{"type": "text", "text": "my favorite dish is lasagna"}],
tool_calls=[
ToolCall(
name="blah_tool",
args={"x": -10},
id="456",
type="tool_call",
)
],
id="blur",
),
]
merge_message_runs(messages)
```
```python
[
SystemMessage("you're a good assistant."),
HumanMessage(
"what's your favorite color\\n"
"wait your favorite food", id="foo",
),
AIMessage(
[
"my favorite colo",
{"type": "text", "text": "my favorite dish is lasagna"}
],
tool_calls=[
ToolCall({
"name": "blah_tool",
"args": {"x": 2},
"id": "123",
"type": "tool_call"
}),
ToolCall({
"name": "blah_tool",
"args": {"x": -10},
"id": "456",
"type": "tool_call"
})
]
id="baz"
),
]
```
"""
if not messages:
return []
messages = convert_to_messages(messages)
merged: list[BaseMessage] = []
for msg in messages:
last = merged.pop() if merged else None
if not last:
merged.append(msg)
elif isinstance(msg, ToolMessage) or not isinstance(msg, last.__class__):
merged.extend([last, msg])
else:
last_chunk = _msg_to_chunk(last)
curr_chunk = _msg_to_chunk(msg)
if curr_chunk.response_metadata:
curr_chunk.response_metadata.clear()
if (
isinstance(last_chunk.content, str)
and isinstance(curr_chunk.content, str)
and last_chunk.content
and curr_chunk.content
):
last_chunk.content += chunk_separator
merged.append(_chunk_to_msg(last_chunk + curr_chunk))
return merged
# TODO: Update so validation errors (for token_counter, for example) are raised on
# init not at runtime.
@_runnable_support
def trim_messages(
messages: Iterable[MessageLikeRepresentation] | PromptValue,
*,
max_tokens: int,
token_counter: Callable[[list[BaseMessage]], int]
| Callable[[BaseMessage], int]
| BaseLanguageModel[Any]
| Literal["approximate"],
strategy: Literal["first", "last"] = "last",
allow_partial: bool = False,
end_on: str | type[BaseMessage] | Sequence[str | type[BaseMessage]] | None = None,
start_on: str | type[BaseMessage] | Sequence[str | type[BaseMessage]] | None = None,
include_system: bool = False,
text_splitter: Callable[[str], list[str]] | TextSplitter | None = None,
) -> list[BaseMessage]:
r"""Trim messages to be below a token count.
`trim_messages` can be used to reduce the size of a chat history to a specified
token or message count.
In either case, if passing the trimmed chat history back into a chat model
directly, the resulting chat history should usually satisfy the following
properties:
1. The resulting chat history should be valid. Most chat models expect that chat
history starts with either (1) a `HumanMessage` or (2) a `SystemMessage`
followed by a `HumanMessage`. To achieve this, set `start_on='human'`.
In addition, generally a `ToolMessage` can only appear after an `AIMessage`
that involved a tool call.
2. It includes recent messages and drops old messages in the chat history.
To achieve this set the `strategy='last'`.
3. Usually, the new chat history should include the `SystemMessage` if it
was present in the original chat history since the `SystemMessage` includes
special instructions to the chat model. The `SystemMessage` is almost always
the first message in the history if present. To achieve this set the
`include_system=True`.
!!! note
The examples below show how to configure `trim_messages` to achieve a behavior
consistent with the above properties.
Args:
messages: Sequence of Message-like objects to trim.
max_tokens: Max token count of trimmed messages.
token_counter: Function or llm for counting tokens in a `BaseMessage` or a
list of `BaseMessage`.
If a `BaseLanguageModel` is passed in then
`BaseLanguageModel.get_num_tokens_from_messages()` will be used. Set to
`len` to count the number of **messages** in the chat history.
You can also use string shortcuts for convenience:
- `'approximate'`: Uses `count_tokens_approximately` for fast, approximate
token counts.
!!! note
`count_tokens_approximately` (or the shortcut `'approximate'`) is
recommended for using `trim_messages` on the hot path, where exact token
counting is not necessary.
strategy: Strategy for trimming.
- `'first'`: Keep the first `<= n_count` tokens of the messages.
- `'last'`: Keep the last `<= n_count` tokens of the messages.
allow_partial: Whether to split a message if only part of the message can be
included.
If `strategy='last'` then the last partial contents of a message are
included. If `strategy='first'` then the first partial contents of a
message are included.
end_on: The message type to end on.
If specified then every message after the last occurrence of this type is
ignored. If `strategy='last'` then this is done before we attempt to get the
last `max_tokens`. If `strategy='first'` then this is done after we get the
first `max_tokens`. Can be specified as string names (e.g. `'system'`,
`'human'`, `'ai'`, ...) or as `BaseMessage` classes (e.g. `SystemMessage`,
`HumanMessage`, `AIMessage`, ...). Can be a single type or a list of types.
start_on: The message type to start on.
Should only be specified if `strategy='last'`. If specified then every
message before the first occurrence of this type is ignored. This is done
after we trim the initial messages to the last `max_tokens`. Does not apply
to a `SystemMessage` at index 0 if `include_system=True`. Can be specified
as string names (e.g. `'system'`, `'human'`, `'ai'`, ...) or as
`BaseMessage` classes (e.g. `SystemMessage`, `HumanMessage`, `AIMessage`,
...). Can be a single type or a list of types.
include_system: Whether to keep the `SystemMessage` if there is one at index
`0`.
Should only be specified if `strategy="last"`.
text_splitter: Function or `langchain_text_splitters.TextSplitter` for
splitting the string contents of a message.
Only used if `allow_partial=True`. If `strategy='last'` then the last split
tokens from a partial message will be included. if `strategy='first'` then
the first split tokens from a partial message will be included. Token
splitter assumes that separators are kept, so that split contents can be
directly concatenated to recreate the original text. Defaults to splitting
on newlines.
Returns:
List of trimmed `BaseMessage`.
Raises:
ValueError: if two incompatible arguments are specified or an unrecognized
`strategy` is specified.
Example:
Trim chat history based on token count, keeping the `SystemMessage` if
present, and ensuring that the chat history starts with a `HumanMessage` (or a
`SystemMessage` followed by a `HumanMessage`).
```python
from langchain_core.messages import (
AIMessage,
HumanMessage,
BaseMessage,
SystemMessage,
trim_messages,
)
messages = [
SystemMessage("you're a good assistant, you always respond with a joke."),
HumanMessage("i wonder why it's called langchain"),
AIMessage(
'Well, I guess they thought "WordRope" and "SentenceString" just '
"didn't have the same ring to it!"
),
HumanMessage("and who is harrison chasing anyways"),
AIMessage(
"Hmmm let me think.\n\nWhy, he's probably chasing after the last "
"cup of coffee in the office!"
),
HumanMessage("what do you call a speechless parrot"),
]
trim_messages(
messages,
max_tokens=45,
strategy="last",
token_counter=ChatOpenAI(model="gpt-5.5"),
# Most chat models expect that chat history starts with either:
# (1) a HumanMessage or
# (2) a SystemMessage followed by a HumanMessage
start_on="human",
# Usually, we want to keep the SystemMessage
# if it's present in the original history.
# The SystemMessage has special instructions for the model.
include_system=True,
allow_partial=False,
)
```
```python
[
SystemMessage(
content="you're a good assistant, you always respond with a joke."
),
HumanMessage(content="what do you call a speechless parrot"),
]
```
Trim chat history using approximate token counting with `'approximate'`:
```python
trim_messages(
messages,
max_tokens=45,
strategy="last",
# Using the "approximate" shortcut for fast token counting
token_counter="approximate",
start_on="human",
include_system=True,
)
# This is equivalent to using `count_tokens_approximately` directly
from langchain_core.messages.utils import count_tokens_approximately
trim_messages(
messages,
max_tokens=45,
strategy="last",
token_counter=count_tokens_approximately,
start_on="human",
include_system=True,
)
```
Trim chat history based on the message count, keeping the `SystemMessage` if
present, and ensuring that the chat history starts with a HumanMessage (
or a `SystemMessage` followed by a `HumanMessage`).
trim_messages(
messages,
# When `len` is passed in as the token counter function,
# max_tokens will count the number of messages in the chat history.
max_tokens=4,
strategy="last",
# Passing in `len` as a token counter function will
# count the number of messages in the chat history.
token_counter=len,
# Most chat models expect that chat history starts with either:
# (1) a HumanMessage or
# (2) a SystemMessage followed by a HumanMessage
start_on="human",
# Usually, we want to keep the SystemMessage
# if it's present in the original history.
# The SystemMessage has special instructions for the model.
include_system=True,
allow_partial=False,
)
```python
[
SystemMessage(
content="you're a good assistant, you always respond with a joke."
),
HumanMessage(content="and who is harrison chasing anyways"),
AIMessage(
content="Hmmm let me think.\n\nWhy, he's probably chasing after "
"the last cup of coffee in the office!"
),
HumanMessage(content="what do you call a speechless parrot"),
]
```
Trim chat history using a custom token counter function that counts the
number of tokens in each message.
```python
messages = [
SystemMessage("This is a 4 token text. The full message is 10 tokens."),
HumanMessage(
"This is a 4 token text. The full message is 10 tokens.", id="first"
),
AIMessage(
[
{"type": "text", "text": "This is the FIRST 4 token block."},
{"type": "text", "text": "This is the SECOND 4 token block."},
],
id="second",
),
HumanMessage(
"This is a 4 token text. The full message is 10 tokens.", id="third"
),
AIMessage(
"This is a 4 token text. The full message is 10 tokens.",
id="fourth",
),
]
def dummy_token_counter(messages: list[BaseMessage]) -> int:
# treat each message like it adds 3 default tokens at the beginning
# of the message and at the end of the message. 3 + 4 + 3 = 10 tokens
# per message.
default_content_len = 4
default_msg_prefix_len = 3
default_msg_suffix_len = 3
count = 0
for msg in messages:
if isinstance(msg.content, str):
count += (
default_msg_prefix_len
+ default_content_len
+ default_msg_suffix_len
)
if isinstance(msg.content, list):
count += (
default_msg_prefix_len
+ len(msg.content) * default_content_len
+ default_msg_suffix_len
)
return count
```
First 30 tokens, allowing partial messages:
```python
trim_messages(
messages,
max_tokens=30,
token_counter=dummy_token_counter,
strategy="first",
allow_partial=True,
)
```
```python
[
SystemMessage("This is a 4 token text. The full message is 10 tokens."),
HumanMessage(
"This is a 4 token text. The full message is 10 tokens.",
id="first",
),
AIMessage(
[{"type": "text", "text": "This is the FIRST 4 token block."}],
id="second",
),
]
```
"""
# Validate arguments
if start_on and strategy == "first":
msg = "start_on parameter is only valid with strategy='last'"
raise ValueError(msg)
if include_system and strategy == "first":
msg = "include_system parameter is only valid with strategy='last'"
raise ValueError(msg)
messages = convert_to_messages(messages)
# Handle string shortcuts for token counter
if isinstance(token_counter, str):
if token_counter in _TOKEN_COUNTER_SHORTCUTS:
actual_token_counter = _TOKEN_COUNTER_SHORTCUTS[token_counter]
else:
available_shortcuts = ", ".join(
f"'{key}'" for key in _TOKEN_COUNTER_SHORTCUTS
)
msg = (
f"Invalid token_counter shortcut '{token_counter}'. "
f"Available shortcuts: {available_shortcuts}."
)
raise ValueError(msg)
else:
# Type narrowing: at this point token_counter is not a str
actual_token_counter = token_counter # type: ignore[assignment]
if hasattr(actual_token_counter, "get_num_tokens_from_messages"):
list_token_counter = actual_token_counter.get_num_tokens_from_messages
elif callable(actual_token_counter):
if (
next(
iter(inspect.signature(actual_token_counter).parameters.values())
).annotation
is BaseMessage
):
def list_token_counter(messages: Sequence[BaseMessage]) -> int:
return sum(actual_token_counter(msg) for msg in messages) # type: ignore[arg-type, misc]
else:
list_token_counter = actual_token_counter
else:
msg = ( # type: ignore[unreachable]
f"'token_counter' expected to be a model that implements "
f"'get_num_tokens_from_messages()' or a function. Received object of type "
f"{type(actual_token_counter)}."
)
raise ValueError(msg)
text_splitter_fn: Callable[[str], list[str]]
if _HAS_LANGCHAIN_TEXT_SPLITTERS and isinstance(text_splitter, TextSplitter):
text_splitter_fn = text_splitter.split_text
elif text_splitter:
text_splitter_fn = cast("Callable[[str], list[str]]", text_splitter)
else:
text_splitter_fn = _default_text_splitter
if strategy == "first":
return _first_max_tokens(
messages,
max_tokens=max_tokens,
token_counter=list_token_counter,
text_splitter=text_splitter_fn,
partial_strategy="first" if allow_partial else None,
end_on=end_on,
)
if strategy == "last":
return _last_max_tokens(
messages,
max_tokens=max_tokens,
token_counter=list_token_counter,
allow_partial=allow_partial,
include_system=include_system,
start_on=start_on,
end_on=end_on,
text_splitter=text_splitter_fn,
)
msg = f"Unrecognized {strategy=}. Supported strategies are 'last' and 'first'." # type: ignore[unreachable]
raise ValueError(msg)
_SingleMessage = BaseMessage | str | dict[str, Any]
_T = TypeVar("_T", bound=_SingleMessage)
# A sequence of _SingleMessage that is NOT a bare str
_MultipleMessages = Sequence[_T]
@overload
def convert_to_openai_messages(
messages: _SingleMessage,
*,
text_format: Literal["string", "block"] = "string",
include_id: bool = False,
pass_through_unknown_blocks: bool = True,
) -> dict[str, Any]: ...
@overload
def convert_to_openai_messages(
messages: _MultipleMessages[Any],
*,
text_format: Literal["string", "block"] = "string",
include_id: bool = False,
pass_through_unknown_blocks: bool = True,
) -> list[dict[str, Any]]: ...
def convert_to_openai_messages(
messages: MessageLikeRepresentation | Sequence[MessageLikeRepresentation],
*,
text_format: Literal["string", "block"] = "string",
include_id: bool = False,
pass_through_unknown_blocks: bool = True,
) -> dict[str, Any] | list[dict[str, Any]]:
"""Convert LangChain messages into OpenAI message dicts.
Args:
messages: Message-like object or iterable of objects whose contents are
in OpenAI, Anthropic, Bedrock Converse, or VertexAI formats.
text_format: How to format string or text block contents:
- `'string'`:
If a message has a string content, this is left as a string. If
a message has content blocks that are all of type `'text'`, these
are joined with a newline to make a single string. If a message has
content blocks and at least one isn't of type `'text'`, then
all blocks are left as dicts.
- `'block'`:
If a message has a string content, this is turned into a list
with a single content block of type `'text'`. If a message has
content blocks these are left as is.
include_id: Whether to include message IDs in the openai messages, if they
are present in the source messages.
pass_through_unknown_blocks: Whether to include content blocks with unknown
formats in the output. If `False`, an error is raised if an unknown
content block is encountered.
Raises:
ValueError: if an unrecognized `text_format` is specified, or if a message
content block is missing expected keys.
Returns:
The return type depends on the input type:
- dict:
If a single message-like object is passed in, a single OpenAI message
dict is returned.
- list[dict]:
If a sequence of message-like objects are passed in, a list of OpenAI
message dicts is returned.
Example:
```python
from langchain_core.messages import (
convert_to_openai_messages,
AIMessage,
SystemMessage,
ToolMessage,
)
messages = [
SystemMessage([{"type": "text", "text": "foo"}]),
{
"role": "user",
"content": [
{"type": "text", "text": "what's in this"},
{
"type": "image_url",
"image_url": {"url": "data:image/png;base64,'/9j/4AAQSk'"},
},
],
},
AIMessage(
"",
tool_calls=[
{
"name": "analyze",
"args": {"baz": "buz"},
"id": "1",
"type": "tool_call",
}
],
),
ToolMessage("foobar", tool_call_id="1", name="bar"),
{"role": "assistant", "content": "that's nice"},
]
oai_messages = convert_to_openai_messages(messages)
# -> [
# {'role': 'system', 'content': 'foo'},
# {'role': 'user', 'content': [{'type': 'text', 'text': 'what's in this'}, {'type': 'image_url', 'image_url': {'url': "data:image/png;base64,'/9j/4AAQSk'"}}]},
# {'role': 'assistant', 'tool_calls': [{'type': 'function', 'id': '1','function': {'name': 'analyze', 'arguments': '{"baz": "buz"}'}}], 'content': ''},
# {'role': 'tool', 'name': 'bar', 'content': 'foobar'},
# {'role': 'assistant', 'content': 'that's nice'}
# ]
```
!!! version-added "Added in `langchain-core` 0.3.11"
""" # noqa: E501
if text_format not in {"string", "block"}:
err = f"Unrecognized {text_format=}, expected one of 'string' or 'block'."
raise ValueError(err)
oai_messages: list[dict[str, Any]] = []
messages_: Sequence[MessageLikeRepresentation]
if is_single := isinstance(messages, (BaseMessage, dict, str)):
messages_ = [messages]
else:
messages_ = cast("Sequence[MessageLikeRepresentation]", messages)
for i, message in enumerate(convert_to_messages(messages_)):
oai_msg: dict[str, Any] = {"role": _get_message_openai_role(message)}
tool_messages: list[dict[str, Any]] = []
content: str | list[dict[str, Any]]
if message.name:
oai_msg["name"] = message.name
if isinstance(message, AIMessage) and message.tool_calls:
oai_msg["tool_calls"] = _convert_to_openai_tool_calls(message.tool_calls)
if message.additional_kwargs.get("refusal"):
oai_msg["refusal"] = message.additional_kwargs["refusal"]
if isinstance(message, ToolMessage):
oai_msg["tool_call_id"] = message.tool_call_id
if include_id and message.id:
oai_msg["id"] = message.id
if not message.content:
content = "" if text_format == "string" else []
elif isinstance(message.content, str):
if text_format == "string":
content = message.content
else:
content = [{"type": "text", "text": message.content}]
elif text_format == "string" and all(
isinstance(block, str) or block.get("type") == "text"
for block in message.content
):
content = "\n".join(
block if isinstance(block, str) else block["text"]
for block in message.content
)
else:
content = []
for j, block in enumerate(message.content):
# OpenAI format
if isinstance(block, str):
content.append({"type": "text", "text": block})
elif block.get("type") == "text":
if missing := [k for k in ("text",) if k not in block]:
err = (
f"Unrecognized content block at "
f"messages[{i}].content[{j}] has 'type': 'text' "
f"but is missing expected key(s) "
f"{missing}. Full content block:\n\n{block}"
)
raise ValueError(err)
content.append({"type": block["type"], "text": block["text"]})
elif block.get("type") == "image_url":
if missing := [k for k in ("image_url",) if k not in block]:
err = (
f"Unrecognized content block at "
f"messages[{i}].content[{j}] has 'type': 'image_url' "
f"but is missing expected key(s) "
f"{missing}. Full content block:\n\n{block}"
)
raise ValueError(err)
content.append(
{
"type": "image_url",
"image_url": block["image_url"],
}
)
# Standard multi-modal content block
elif is_data_content_block(block):
formatted_block = convert_to_openai_data_block(block)
if (
formatted_block.get("type") == "file"
and "file" in formatted_block
and "filename" not in formatted_block["file"]
):
logger.info("Generating a fallback filename.")
formatted_block["file"]["filename"] = "LC_AUTOGENERATED"
content.append(formatted_block)
# Anthropic and Bedrock converse format
elif (block.get("type") == "image") or "image" in block:
# Anthropic
if source := block.get("source"):
if missing := [
k for k in ("media_type", "type", "data") if k not in source
]:
err = (
f"Unrecognized content block at "
f"messages[{i}].content[{j}] has 'type': 'image' "
f"but 'source' is missing expected key(s) "
f"{missing}. Full content block:\n\n{block}"
)
raise ValueError(err)
content.append(
{
"type": "image_url",
"image_url": {
"url": (
f"data:{source['media_type']};"
f"{source['type']},{source['data']}"
)
},
}
)
# Bedrock converse
elif image := block.get("image"):
if missing := [
k for k in ("source", "format") if k not in image
]:
err = (
f"Unrecognized content block at "
f"messages[{i}].content[{j}] has key 'image', "
f"but 'image' is missing expected key(s) "
f"{missing}. Full content block:\n\n{block}"
)
raise ValueError(err)
b64_image = _bytes_to_b64_str(image["source"]["bytes"])
content.append(
{
"type": "image_url",
"image_url": {
"url": (
f"data:image/{image['format']};base64,{b64_image}"
)
},
}
)
else:
err = (
f"Unrecognized content block at "
f"messages[{i}].content[{j}] has 'type': 'image' "
f"but does not have a 'source' or 'image' key. Full "
f"content block:\n\n{block}"
)
raise ValueError(err)
# OpenAI file format
elif (
block.get("type") == "file"
and isinstance(block.get("file"), dict)
and isinstance(block.get("file", {}).get("file_data"), str)
):
if block.get("file", {}).get("filename") is None:
logger.info("Generating a fallback filename.")
formatted_block = {
**block,
"file": {**block["file"], "filename": "LC_AUTOGENERATED"},
}
content.append(formatted_block)
else:
content.append(block)
# OpenAI audio format
elif (
block.get("type") == "input_audio"
and isinstance(block.get("input_audio"), dict)
and isinstance(block.get("input_audio", {}).get("data"), str)
and isinstance(block.get("input_audio", {}).get("format"), str)
):
content.append(block)
elif block.get("type") == "tool_use":
if missing := [
k for k in ("id", "name", "input") if k not in block
]:
err = (
f"Unrecognized content block at "
f"messages[{i}].content[{j}] has 'type': "
f"'tool_use', but is missing expected key(s) "
f"{missing}. Full content block:\n\n{block}"
)
raise ValueError(err)
if not any(
tool_call["id"] == block["id"]
for tool_call in cast("AIMessage", message).tool_calls
):
oai_msg["tool_calls"] = oai_msg.get("tool_calls", [])
oai_msg["tool_calls"].append(
{
"type": "function",
"id": block["id"],
"function": {
"name": block["name"],
"arguments": json.dumps(
block["input"], ensure_ascii=False
),
},
}
)
elif block.get("type") == "function_call": # OpenAI Responses
if not any(
tool_call["id"] == block.get("call_id")
for tool_call in cast("AIMessage", message).tool_calls
):
if missing := [
k
for k in ("call_id", "name", "arguments")
if k not in block
]:
err = (
f"Unrecognized content block at "
f"messages[{i}].content[{j}] has 'type': "
f"'tool_use', but is missing expected key(s) "
f"{missing}. Full content block:\n\n{block}"
)
raise ValueError(err)
oai_msg["tool_calls"] = oai_msg.get("tool_calls", [])
oai_msg["tool_calls"].append(
{
"type": "function",
"id": block.get("call_id"),
"function": {
"name": block.get("name"),
"arguments": block.get("arguments"),
},
}
)
if pass_through_unknown_blocks:
content.append(block)
elif block.get("type") == "tool_result":
if missing := [
k for k in ("content", "tool_use_id") if k not in block
]:
msg = (
f"Unrecognized content block at "
f"messages[{i}].content[{j}] has 'type': "
f"'tool_result', but is missing expected key(s) "
f"{missing}. Full content block:\n\n{block}"
)
raise ValueError(msg)
tool_message = ToolMessage(
block["content"],
tool_call_id=block["tool_use_id"],
status="error" if block.get("is_error") else "success",
)
# Recurse to make sure tool message contents are OpenAI format.
tool_messages.extend(
convert_to_openai_messages(
[tool_message], text_format=text_format
)
)
elif (block.get("type") == "json") or "json" in block:
if "json" not in block:
msg = (
f"Unrecognized content block at "
f"messages[{i}].content[{j}] has 'type': 'json' "
f"but does not have a 'json' key. Full "
f"content block:\n\n{block}"
)
raise ValueError(msg)
content.append(
{
"type": "text",
"text": json.dumps(block["json"]),
}
)
elif (block.get("type") == "guard_content") or "guard_content" in block:
if (
"guard_content" not in block
or "text" not in block["guard_content"]
):
msg = (
f"Unrecognized content block at "
f"messages[{i}].content[{j}] has 'type': "
f"'guard_content' but does not have a "
f"messages[{i}].content[{j}]['guard_content']['text'] "
f"key. Full content block:\n\n{block}"
)
raise ValueError(msg)
text = block["guard_content"]["text"]
if isinstance(text, dict):
text = text["text"]
content.append({"type": "text", "text": text})
# VertexAI format
elif block.get("type") == "media":
if missing := [k for k in ("mime_type", "data") if k not in block]:
err = (
f"Unrecognized content block at "
f"messages[{i}].content[{j}] has 'type': "
f"'media' but does not have key(s) {missing}. Full "
f"content block:\n\n{block}"
)
raise ValueError(err)
if "image" not in block["mime_type"]:
err = (
f"OpenAI messages can only support text and image data."
f" Received content block with media of type:"
f" {block['mime_type']}"
)
raise ValueError(err)
b64_image = _bytes_to_b64_str(block["data"])
content.append(
{
"type": "image_url",
"image_url": {
"url": (f"data:{block['mime_type']};base64,{b64_image}")
},
}
)
elif (
block.get("type") in {"thinking", "reasoning"}
or pass_through_unknown_blocks
):
content.append(block)
else:
err = (
f"Unrecognized content block at "
f"messages[{i}].content[{j}] does not match OpenAI, "
f"Anthropic, Bedrock Converse, or VertexAI format. Full "
f"content block:\n\n{block}"
)
raise ValueError(err)
if text_format == "string" and not any(
block["type"] != "text" for block in content
):
content = "\n".join(block["text"] for block in content)
oai_msg["content"] = content
if message.content and not oai_msg["content"] and tool_messages:
oai_messages.extend(tool_messages)
else:
oai_messages.extend([oai_msg, *tool_messages])
if is_single:
return oai_messages[0]
return oai_messages
def _first_max_tokens(
messages: Sequence[BaseMessage],
*,
max_tokens: int,
token_counter: Callable[[list[BaseMessage]], int],
text_splitter: Callable[[str], list[str]],
partial_strategy: Literal["first", "last"] | None = None,
end_on: str | type[BaseMessage] | Sequence[str | type[BaseMessage]] | None = None,
) -> list[BaseMessage]:
messages = list(messages)
if not messages:
return messages
# Check if all messages already fit within token limit
if token_counter(messages) <= max_tokens:
# When all messages fit, only apply end_on filtering if needed
if end_on:
for _ in range(len(messages)):
if not _is_message_type(messages[-1], end_on):
messages.pop()
else:
break
return messages
# Use binary search to find the maximum number of messages within token limit
left, right = 0, len(messages)
max_iterations = len(messages).bit_length()
for _ in range(max_iterations):
if left >= right:
break
mid = (left + right + 1) // 2
if token_counter(messages[:mid]) <= max_tokens:
left = mid
idx = mid
else:
right = mid - 1
# idx now contains the maximum number of complete messages we can include
idx = left
if partial_strategy and idx < len(messages):
included_partial = False
copied = False
if isinstance(messages[idx].content, list):
excluded = messages[idx].model_copy(deep=True)
copied = True
num_block = len(excluded.content)
if partial_strategy == "last":
excluded.content = list(reversed(excluded.content))
for _ in range(1, num_block):
excluded.content = excluded.content[:-1]
if token_counter([*messages[:idx], excluded]) <= max_tokens:
messages = [*messages[:idx], excluded]
idx += 1
included_partial = True
break
if included_partial and partial_strategy == "last":
excluded.content = list(reversed(excluded.content))
if not included_partial:
if not copied:
excluded = messages[idx].model_copy(deep=True)
copied = True
# Extract text content efficiently
text = None
if isinstance(excluded.content, str):
text = excluded.content
elif isinstance(excluded.content, list) and excluded.content:
for block in excluded.content:
if isinstance(block, str):
text = block
break
if isinstance(block, dict) and block.get("type") == "text":
text = block.get("text")
break
if text:
if not copied:
excluded = excluded.model_copy(deep=True)
split_texts = text_splitter(text)
base_message_count = token_counter(messages[:idx])
if partial_strategy == "last":
split_texts = list(reversed(split_texts))
# Binary search for the maximum number of splits we can include
left, right = 0, len(split_texts)
max_iterations = len(split_texts).bit_length()
for _ in range(max_iterations):
if left >= right:
break
mid = (left + right + 1) // 2
excluded.content = "".join(split_texts[:mid])
if base_message_count + token_counter([excluded]) <= max_tokens:
left = mid
else:
right = mid - 1
if left > 0:
content_splits = split_texts[:left]
if partial_strategy == "last":
content_splits = list(reversed(content_splits))
excluded.content = "".join(content_splits)
messages = [*messages[:idx], excluded]
idx += 1
if end_on:
for _ in range(idx):
if idx > 0 and not _is_message_type(messages[idx - 1], end_on):
idx -= 1
else:
break
return messages[:idx]
def _last_max_tokens(
messages: Sequence[BaseMessage],
*,
max_tokens: int,
token_counter: Callable[[list[BaseMessage]], int],
text_splitter: Callable[[str], list[str]],
allow_partial: bool = False,
include_system: bool = False,
start_on: str | type[BaseMessage] | Sequence[str | type[BaseMessage]] | None = None,
end_on: str | type[BaseMessage] | Sequence[str | type[BaseMessage]] | None = None,
) -> list[BaseMessage]:
messages = list(messages)
if len(messages) == 0:
return []
# Filter out messages after end_on type
if end_on:
for _ in range(len(messages)):
if not _is_message_type(messages[-1], end_on):
messages.pop()
else:
break
# Handle system message preservation
system_message = None
if include_system and len(messages) > 0 and isinstance(messages[0], SystemMessage):
system_message = messages[0]
messages = messages[1:]
# Reverse messages to use _first_max_tokens with reversed logic
reversed_messages = messages[::-1]
# Calculate remaining tokens after accounting for system message if present
remaining_tokens = max_tokens
if system_message:
system_tokens = token_counter([system_message])
remaining_tokens = max(0, max_tokens - system_tokens)
reversed_result = _first_max_tokens(
reversed_messages,
max_tokens=remaining_tokens,
token_counter=token_counter,
text_splitter=text_splitter,
partial_strategy="last" if allow_partial else None,
end_on=start_on,
)
# Re-reverse the messages and add back the system message if needed
result = reversed_result[::-1]
if system_message:
result = [system_message, *result]
return result
_MSG_CHUNK_MAP: dict[type[BaseMessage], type[BaseMessageChunk]] = {
HumanMessage: HumanMessageChunk,
AIMessage: AIMessageChunk,
SystemMessage: SystemMessageChunk,
ToolMessage: ToolMessageChunk,
FunctionMessage: FunctionMessageChunk,
ChatMessage: ChatMessageChunk,
}
_CHUNK_MSG_MAP = {v: k for k, v in _MSG_CHUNK_MAP.items()}
def _msg_to_chunk(message: BaseMessage) -> BaseMessageChunk:
if message.__class__ in _MSG_CHUNK_MAP:
return _MSG_CHUNK_MAP[message.__class__](**message.model_dump(exclude={"type"}))
for msg_cls, chunk_cls in _MSG_CHUNK_MAP.items():
if isinstance(message, msg_cls):
return chunk_cls(**message.model_dump(exclude={"type"}))
msg = (
f"Unrecognized message class {message.__class__}. Supported classes are "
f"{list(_MSG_CHUNK_MAP.keys())}"
)
msg = create_message(message=msg, error_code=ErrorCode.MESSAGE_COERCION_FAILURE)
raise ValueError(msg)
def _chunk_to_msg(chunk: BaseMessageChunk) -> BaseMessage:
if chunk.__class__ in _CHUNK_MSG_MAP:
return _CHUNK_MSG_MAP[chunk.__class__](
**chunk.model_dump(exclude={"type", "tool_call_chunks", "chunk_position"})
)
for chunk_cls, msg_cls in _CHUNK_MSG_MAP.items():
if isinstance(chunk, chunk_cls):
return msg_cls(
**chunk.model_dump(
exclude={"type", "tool_call_chunks", "chunk_position"}
)
)
msg = (
f"Unrecognized message chunk class {chunk.__class__}. Supported classes are "
f"{list(_CHUNK_MSG_MAP.keys())}"
)
msg = create_message(message=msg, error_code=ErrorCode.MESSAGE_COERCION_FAILURE)
raise ValueError(msg)
def _default_text_splitter(text: str) -> list[str]:
splits = text.split("\n")
return [s + "\n" for s in splits[:-1]] + splits[-1:]
def _is_message_type(
message: BaseMessage,
type_: str | type[BaseMessage] | Sequence[str | type[BaseMessage]],
) -> bool:
types = [type_] if isinstance(type_, (str, type)) else type_
types_str = [t for t in types if isinstance(t, str)]
types_types = tuple(t for t in types if isinstance(t, type))
return message.type in types_str or isinstance(message, types_types)
def _bytes_to_b64_str(bytes_: bytes) -> str:
return base64.b64encode(bytes_).decode("utf-8")
def _get_message_openai_role(message: BaseMessage) -> str:
if isinstance(message, AIMessage):
return "assistant"
if isinstance(message, HumanMessage):
return "user"
if isinstance(message, ToolMessage):
return "tool"
if isinstance(message, SystemMessage):
role = message.additional_kwargs.get("__openai_role__", "system")
if not isinstance(role, str):
msg = f"Expected '__openai_role__' to be a str, got {type(role).__name__}"
raise TypeError(msg)
return role
if isinstance(message, FunctionMessage):
return "function"
if isinstance(message, ChatMessage):
return message.role
msg = f"Unknown BaseMessage type {message.__class__}."
raise ValueError(msg)
def _convert_to_openai_tool_calls(tool_calls: list[ToolCall]) -> list[dict[str, Any]]:
return [
{
"type": "function",
"id": tool_call["id"],
"function": {
"name": tool_call["name"],
"arguments": json.dumps(tool_call["args"], ensure_ascii=False),
},
}
for tool_call in tool_calls
]
def count_tokens_approximately(
messages: Iterable[MessageLikeRepresentation],
*,
chars_per_token: float = 4.0,
extra_tokens_per_message: float = 3.0,
count_name: bool = True,
tokens_per_image: int = 85,
use_usage_metadata_scaling: bool = False,
tools: list[BaseTool | dict[str, Any]] | None = None,
) -> int:
"""Approximate the total number of tokens in messages.
The token count includes stringified message content, role, and (optionally) name.
- For AI messages, the token count also includes stringified tool calls.
- For tool messages, the token count also includes the tool call ID.
- For multimodal messages with images, applies a fixed token penalty per image
instead of counting base64-encoded characters.
- If tools are provided, the token count also includes stringified tool schemas.
Args:
messages: List of messages to count tokens for.
chars_per_token: Number of characters per token to use for the approximation.
One token corresponds to ~4 chars for common English text.
You can also specify `float` values for more fine-grained control.
[See more here](https://platform.openai.com/tokenizer).
extra_tokens_per_message: Number of extra tokens to add per message, e.g.
special tokens, including beginning/end of message.
You can also specify `float` values for more fine-grained control.
[See more here](https://github.com/openai/openai-cookbook/blob/main/examples/How_to_count_tokens_with_tiktoken.ipynb).
count_name: Whether to include message names in the count.
tokens_per_image: Fixed token cost per image (default: 85, aligned with
OpenAI's low-resolution image token cost).
use_usage_metadata_scaling: If True, and all AI messages have consistent
`response_metadata['model_provider']`, scale the approximate token count
using the **most recent** AI message that has
`usage_metadata['total_tokens']`. The scaling factor is:
`AI_total_tokens / approx_tokens_up_to_that_AI_message`
tools: List of tools to include in the token count. Each tool can be a
`BaseTool` instance, a dict representing a tool schema, or a plain
callable. `BaseTool` instances and callables are converted to
OpenAI tool format before counting.
Returns:
Approximate number of tokens in the messages (and tools, if provided).
Note:
This is a simple approximation that may not match the exact token count used by
specific models. For accurate counts, use model-specific tokenizers.
For multimodal messages containing images, a fixed token penalty is applied
per image instead of counting base64-encoded characters, which provides a
more realistic approximation.
!!! version-added "Added in `langchain-core` 0.3.46"
"""
converted_messages = convert_to_messages(messages)
token_count = 0.0
ai_model_provider: str | None = None
invalid_model_provider = False
last_ai_total_tokens: int | None = None
approx_at_last_ai: float | None = None
# Count tokens for tools if provided
if tools:
tools_chars = 0
for tool in tools:
if isinstance(tool, dict):
tool_dict = tool
else:
# tool_call_schema is memoized per instance
schema = tool.tool_call_schema
if isinstance(schema, dict):
parameters = dict(schema)
else:
parameters = dict(get_model_json_schema(schema))
# Drop the schema's own `title`/`description` to avoid double-counting:
# they're duplicated at the top level.
parameters.pop("title", None)
parameters.pop("description", None)
tool_dict = {
"name": tool.name,
"description": tool.description,
"parameters": parameters,
}
tools_chars += len(json.dumps(tool_dict))
token_count += math.ceil(tools_chars / chars_per_token)
for message in converted_messages:
message_chars = 0
if isinstance(message.content, str):
message_chars += len(message.content)
# Handle multimodal content (list of content blocks)
elif isinstance(message.content, list):
for block in message.content:
if isinstance(block, str):
# String block
message_chars += len(block)
elif isinstance(block, dict):
block_type = block.get("type", "")
# Apply fixed penalty for image blocks
if block_type in {"image", "image_url"}:
token_count += tokens_per_image
# Count text blocks normally
elif block_type == "text":
text = block.get("text", "")
message_chars += len(text)
# Conservative estimate for unknown block types
else:
message_chars += len(repr(block))
else:
# Fallback for unexpected block types
message_chars += len(repr(block)) # type: ignore[unreachable]
else:
# Fallback for other content types
content = repr(message.content) # type: ignore[unreachable]
message_chars += len(content)
if (
isinstance(message, AIMessage)
# exclude Anthropic format as tool calls are already included in the content
and not isinstance(message.content, list)
and message.tool_calls
):
tool_calls_content = repr(message.tool_calls)
message_chars += len(tool_calls_content)
if isinstance(message, ToolMessage):
message_chars += len(message.tool_call_id)
role = _get_message_openai_role(message)
message_chars += len(role)
if message.name and count_name:
message_chars += len(message.name)
# NOTE: we're rounding up per message to ensure that
# individual message token counts add up to the total count
# for a list of messages
token_count += math.ceil(message_chars / chars_per_token)
# add extra tokens per message
token_count += extra_tokens_per_message
if use_usage_metadata_scaling and isinstance(message, AIMessage):
model_provider = message.response_metadata.get("model_provider")
if ai_model_provider is None:
ai_model_provider = model_provider
elif model_provider != ai_model_provider:
invalid_model_provider = True
if message.usage_metadata and isinstance(
(total_tokens := message.usage_metadata.get("total_tokens")), int
):
last_ai_total_tokens = total_tokens
approx_at_last_ai = token_count
if (
use_usage_metadata_scaling
and len(converted_messages) > 1
and not invalid_model_provider
and ai_model_provider is not None
and last_ai_total_tokens is not None
and approx_at_last_ai
and approx_at_last_ai > 0
):
scale_factor = last_ai_total_tokens / approx_at_last_ai
token_count *= min(1.25, max(1.0, scale_factor))
# round up once more time in case extra_tokens_per_message is a float
return math.ceil(token_count)
# Mapping from string shortcuts to token counter functions
def _approximate_token_counter(messages: Sequence[BaseMessage]) -> int:
"""Wrapper for `count_tokens_approximately` that matches expected signature."""
return count_tokens_approximately(messages)
_TOKEN_COUNTER_SHORTCUTS = {
"approximate": _approximate_token_counter,
}