How to track token usage for LLMs
Trackingtoken usage to calculate cost is an important part of putting your app in production. This guide goes over how to obtain this information from your LangChain model calls.
This guide assumes familiarity with the following concepts:
Using LangSmith
You can useLangSmith to help track token usage in your LLM application. See theLangSmith quick start guide.
Using callbacks
There are some API-specific callback context managers that allow you to track token usage across multiple calls. You'll need to check whether such an integration is available for your particular model.
If such an integration is not available for your model, you can create a custom callback manager by adapting the implementation of theOpenAI callback manager.
OpenAI
Let's first look at an extremely simple example of tracking token usage for a single Chat model call.
The callback handler does not currently support streaming token counts for legacy language models (e.g.,langchain_openai.OpenAI). For support in a streaming context, refer to the corresponding guide for chat modelshere.
Single call
from langchain_community.callbacksimport get_openai_callback
from langchain_openaiimport OpenAI
llm= OpenAI(model_name="gpt-3.5-turbo-instruct")
with get_openai_callback()as cb:
result= llm.invoke("Tell me a joke")
print(result)
print("---")
print()
print(f"Total Tokens:{cb.total_tokens}")
print(f"Prompt Tokens:{cb.prompt_tokens}")
print(f"Completion Tokens:{cb.completion_tokens}")
print(f"Total Cost (USD): ${cb.total_cost}")
Why don't scientists trust atoms?
Because they make up everything.
---
Total Tokens: 18
Prompt Tokens: 4
Completion Tokens: 14
Total Cost (USD): $3.4e-05
Multiple calls
Anything inside the context manager will get tracked. Here's an example of using it to track multiple calls in sequence to a chain. This will also work for an agent which may use multiple steps.
from langchain_community.callbacksimport get_openai_callback
from langchain_core.promptsimport PromptTemplate
from langchain_openaiimport OpenAI
llm= OpenAI(model_name="gpt-3.5-turbo-instruct")
template= PromptTemplate.from_template("Tell me a joke about {topic}")
chain= template| llm
with get_openai_callback()as cb:
response= chain.invoke({"topic":"birds"})
print(response)
response= chain.invoke({"topic":"fish"})
print("--")
print(response)
print()
print("---")
print(f"Total Tokens:{cb.total_tokens}")
print(f"Prompt Tokens:{cb.prompt_tokens}")
print(f"Completion Tokens:{cb.completion_tokens}")
print(f"Total Cost (USD): ${cb.total_cost}")
Why did the chicken go to the seance?
To talk to the other side of the road!
--
Why did the fish need a lawyer?
Because it got caught in a net!
---
Total Tokens: 50
Prompt Tokens: 12
Completion Tokens: 38
Total Cost (USD): $9.400000000000001e-05
Streaming
get_openai_callback does not currently support streaming token counts for legacy language models (e.g.,langchain_openai.OpenAI). If you want to count tokens correctly in a streaming context, there are a number of options:
- Use chat models as described inthis guide;
- Implement acustom callback handler that uses appropriate tokenizers to count the tokens;
- Use a monitoring platform such asLangSmith.
Note that when using legacy language models in a streaming context, token counts are not updated:
from langchain_community.callbacksimport get_openai_callback
from langchain_openaiimport OpenAI
llm= OpenAI(model_name="gpt-3.5-turbo-instruct")
with get_openai_callback()as cb:
for chunkin llm.stream("Tell me a joke"):
print(chunk, end="", flush=True)
print(result)
print("---")
print()
print(f"Total Tokens:{cb.total_tokens}")
print(f"Prompt Tokens:{cb.prompt_tokens}")
print(f"Completion Tokens:{cb.completion_tokens}")
print(f"Total Cost (USD): ${cb.total_cost}")
Why don't scientists trust atoms?
Because they make up everything!
Why don't scientists trust atoms?
Because they make up everything.
---
Total Tokens: 0
Prompt Tokens: 0
Completion Tokens: 0
Total Cost (USD): $0.0