Eric J Ma's Website

How to make Python context managers aware of their code

written by Eric J. Ma on 2023-05-02 | tags: python context manager llamabot code snippet contextvars programming til

While working on llamabot, I implemented a new feature: automagically recording prompts and LLM responses. I've wanted this for a while because it would eliminate a lot of friction associated with designing prompts, namely, having to keep track of what I tried and what the LLM responded with.

On thinking about how to implement it, I went with a PromptRecorder object that uses a context manager. The syntax I was going for was to match PyMC's pm.Model() context manager, which looks something like this:

import pymc as pm

with pm.Model() as model:
    a = pm.Normal("a")

I was trying to go for the following:

from llamabot import PromptRecorder, SimpleBot

recorder = PromptRecorder()
bot = SimpleBot()

with recorder:

# we would then access an attribute of `recorder` to get the prompt-response pairs.

As an alternative, I can imagine interactively experimenting with prompts inside a Jupyter notebook cell:

with recorder:

And we could access the prompt-response pairs from the recorder object.

But how do we make a context manager aware of its internal code?

As it turns out, that's not easily doable. However, the reverse direction is! Code inside the context manager can modify the context manager's contents if referenced properly.

How do we reference a context manager from within a code chunk? I'm going to show a snippet from llamabot to illustrate.

Firstly, within the source file that we use, we start with the following:

import contextvars
prompt_recorder_var = contextvars.ContextVar("prompt_recorder")

This gives a globally-referenceable prompt_recorder variable we can call on later.

Next, within the PromptRecorder class, we implement the __enter__ and __exit__ methods as follows:

class PromptRecorder:
    def __enter__(self):
        return self

    def __exit__(self, exc_type, exc_val, exc_tb):

Upon entering the context manager, the __enter__ method sets the prompt_recorder variable as the instantiated PromptRecorder object.

One final function gives our SimpleBot objects the ability to use the PromptRecorder within them:

def autorecord(prompt: str, response: str):
    # Log the response.
    prompt_recorder: Optional[PromptRecorder] = prompt_recorder_var.get(None)
    if prompt_recorder:
        prompt_recorder.log(prompt, response)

This function accepts a prompt and a response and automatically checks if a PromptRecorder object is stored inside the prompt_recorder_var. If yes, it will record the response. If not, nothing happens. We use this function inside the SimpleBot class when calling on it:

class SimpleBot:
    def __call__(self, human_message: str) -> AIMessage:
        response = self.model(messages)
        autorecord(human_message, response.content)
        return response

And just like that, the SimpleBot is made aware of the PromptRecorder context manager!

The fundamental trick is keeping a globally referenceable prompt_recorder variable enabled by the built-in contextvars module. With that, we can modify the instantiated PromptRecorder object's state, call on the object's methods, and do other things with the object while still hiding it from the front-end code we write.

If you're curious, you can study the code here

Cite this blog post:
    author = {Eric J. Ma},
    title = {How to make Python context managers aware of their code},
    year = {2023},
    month = {05},
    day = {02},
    howpublished = {\url{}},
    journal = {Eric J. Ma's Blog},
    url = {},

I send out a newsletter with tips and tools for data scientists. Come check it out at Substack.

If you would like to sponsor the coffee that goes into making my posts, please consider GitHub Sponsors!

Finally, I do free 30-minute GenAI strategy calls for teams that are looking to leverage GenAI for maximum impact. Consider booking a call on Calendly if you're interested!