How to trim messages
This guide assumes familiarity with the following concepts:
The methods in this guide also require langchain-core>=0.2.9.
All models have finite context windows, meaning there's a limit to how many tokens they can take as input. If you have very long messages or a chain/agent that accumulates a long message is history, you'll need to manage the length of the messages you're passing in to the model.
The trim_messages util provides some basic strategies for trimming a list of messages to be of a certain token length.
Getting the last max_tokens tokens
To get the last max_tokens in the list of Messages we can set strategy="last". Notice that for our token_counter we can pass in a function (more on that below) or a language model (since language models have a message token counting method). It makes sense to pass in a model when you're trimming your messages to fit into the context window of that specific model:
# pip install -U langchain-openai
from langchain_core.messages import (
AIMessage,
HumanMessage,
SystemMessage,
trim_messages,
)
from langchain_openai import ChatOpenAI
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-4o"),
)
[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')]
If we want to always keep the initial system message we can specify include_system=True:
trim_messages(
messages,
max_tokens=45,
strategy="last",
token_counter=ChatOpenAI(model="gpt-4o"),
include_system=True,
)
[SystemMessage(content="you're a good assistant, you always respond with a joke."),
HumanMessage(content='what do you call a speechless parrot')]
If we want to allow splitting up the contents of a message we can specify allow_partial=True:
trim_messages(
messages,
max_tokens=56,
strategy="last",
token_counter=ChatOpenAI(model="gpt-4o"),
include_system=True,
allow_partial=True,
)
[SystemMessage(content="you're a good assistant, you always respond with a joke."),
AIMessage(content="\nWhy, he's probably chasing after the last cup of coffee in the office!"),
HumanMessage(content='what do you call a speechless parrot')]
If we need to make sure that our first message (excluding the system message) is always of a specific type, we can specify start_on:
trim_messages(
messages,
max_tokens=60,
strategy="last",
token_counter=ChatOpenAI(model="gpt-4o"),
include_system=True,
start_on="human",
)
[SystemMessage(content="you're a good assistant, you always respond with a joke."),
HumanMessage(content='what do you call a speechless parrot')]
Getting the first max_tokens tokens
We can perform the flipped operation of getting the first max_tokens by specifying strategy="first":
trim_messages(
messages,
max_tokens=45,
strategy="first",
token_counter=ChatOpenAI(model="gpt-4o"),
)
[SystemMessage(content="you're a good assistant, you always respond with a joke."),
HumanMessage(content="i wonder why it's called langchain")]
Writing a custom token counter
We can write a custom token counter function that takes in a list of messages and returns an int.
from typing import List
# pip install tiktoken
import tiktoken
from langchain_core.messages import BaseMessage, ToolMessage
def str_token_counter(text: str) -> int:
enc = tiktoken.get_encoding("o200k_base")
return len(enc.encode(text))
def tiktoken_counter(messages: List[BaseMessage]) -> int:
"""Approximately reproduce https://github.com/openai/openai-cookbook/blob/main/examples/How_to_count_tokens_with_tiktoken.ipynb
For simplicity only supports str Message.contents.
"""
num_tokens = 3 # every reply is primed with <|start|>assistant<|message|>
tokens_per_message = 3
tokens_per_name = 1
for msg in messages:
if isinstance(msg, HumanMessage):
role = "user"
elif isinstance(msg, AIMessage):
role = "assistant"
elif isinstance(msg, ToolMessage):
role = "tool"
elif isinstance(msg, SystemMessage):
role = "system"
else:
raise ValueError(f"Unsupported messages type {msg.__class__}")
num_tokens += (
tokens_per_message
+ str_token_counter(role)
+ str_token_counter(msg.content)
)
if msg.name:
num_tokens += tokens_per_name + str_token_counter(msg.name)
return num_tokens
trim_messages(
messages,
max_tokens=45,
strategy="last",
token_counter=tiktoken_counter,
)
[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')]
Chaining
trim_messages can be used in an imperatively (like above) or declaratively, making it easy to compose with other components in a chain
llm = ChatOpenAI(model="gpt-4o")
# Notice we don't pass in messages. This creates
# a RunnableLambda that takes messages as input
trimmer = trim_messages(
max_tokens=45,
strategy="last",
token_counter=llm,
include_system=True,
)
chain = trimmer | llm
chain.invoke(messages)
AIMessage(content='A "polygon"! Because it\'s a "poly-gone" silent!', response_metadata={'token_usage': {'completion_tokens': 14, 'prompt_tokens': 32, 'total_tokens': 46}, 'model_name': 'gpt-4o-2024-05-13', 'system_fingerprint': 'fp_319be4768e', 'finish_reason': 'stop', 'logprobs': None}, id='run-64cc4575-14d1-4f3f-b4af-97f24758f703-0', usage_metadata={'input_tokens': 32, 'output_tokens': 14, 'total_tokens': 46})
Looking at the LangSmith trace we can see that before the messages are passed to the model they are first trimmed: https://smith.langchain.com/public/65af12c4-c24d-4824-90f0-6547566e59bb/r
Looking at just the trimmer, we can see that it's a Runnable object that can be invoked like all Runnables:
trimmer.invoke(messages)
[SystemMessage(content="you're a good assistant, you always respond with a joke."),
HumanMessage(content='what do you call a speechless parrot')]
Using with ChatMessageHistory
Trimming messages is especially useful when working with chat histories, which can get arbitrarily long:
from langchain_core.chat_history import InMemoryChatMessageHistory
from langchain_core.runnables.history import RunnableWithMessageHistory
chat_history = InMemoryChatMessageHistory(messages=messages[:-1])
def dummy_get_session_history(session_id):
if session_id != "1":
raise InMemoryChatMessageHistory()
return chat_history
llm = ChatOpenAI(model="gpt-4o")
trimmer = trim_messages(
max_tokens=45,
strategy="last",
token_counter=llm,
include_system=True,
)
chain = trimmer | llm
chain_with_history = RunnableWithMessageHistory(chain, dummy_get_session_history)
chain_with_history.invoke(
[HumanMessage("what do you call a speechless parrot")],
config={"configurable": {"session_id": "1"}},
)
Parent run c87e2f1b-81ad-4fa7-bfd9-ce6edb29a482 not found for run 7892ee8f-0669-4d6b-a2ca-ef8aae81042a. Treating as a root run.
AIMessage(content="A polygon! Because it's a parrot gone quiet!", response_metadata={'token_usage': {'completion_tokens': 11, 'prompt_tokens': 32, 'total_tokens': 43}, 'model_name': 'gpt-4o-2024-05-13', 'system_fingerprint': 'fp_319be4768e', 'finish_reason': 'stop', 'logprobs': None}, id='run-72dad96e-8b58-45f4-8c08-21f9f1a6b68f-0', usage_metadata={'input_tokens': 32, 'output_tokens': 11, 'total_tokens': 43})
Looking at the LangSmith trace we can see that we retrieve all of our messages but before the messages are passed to the model they are trimmed to be just the system message and last human message: https://smith.langchain.com/public/17dd700b-9994-44ca-930c-116e00997315/r
API reference
For a complete description of all arguments head to the API reference: https://api.python.langchain.com/en/latest/messages/langchain_core.messages.utils.trim_messages.html