Cache completions

[abrichr]

Is there some mechanism to avoid hitting the API if the prompt hasn't changed at all?

For example:

import ell

@ell.simple(model="gpt-4o")
def hello(name: str):
    """You are a helpful assistant.""" # System prompt
    return f"Say hello to {name}!" # User prompt

greeting = hello("Sam Altman")
print(greeting)

If we run this script twice, there is no need for the API to be called on the second time if we simply persist the result of the function call to disk.

Normally we can accomplish this with joblib.memory:

from joblib import Memory
import ell

memory = Memory("./cache")

@memory.cache()
@ell.simple(model="gpt-4o")
def hello(name: str):
    """You are a helpful assistant.""" # System prompt
    return f"Say hello to {name}!" # User prompt

greeting = hello("Sam Altman")
print(greeting)

Now if we run this script twice, the API will not be hit on the second call.

This behaves as we expect if we modify the parameters to the function, e.g. if we call hello("Sam"), the API will be hit, since the arguments changed.

However, if we change the prompt literal inside the function, unfortunately joblib is not able to pick up on it, and the stale result is returned.

Any suggestions for avoiding unnecessary API calls would be appreciated!

[abrichr]

Workaround: incorporate a hash of the function’s source code (including the prompt) into the cache key.

import hashlib
import inspect
from joblib import Memory
import ell

memory = Memory("./cache")

def hash_source_code(func):
    """Hash the entire function's source code, including its docstring."""
    source = inspect.getsource(func)
    return hashlib.sha256(source.encode("utf-8")).hexdigest()

def cache_source(func):
    """Decorator to cache a function, including its source code in the hash."""

    def wrapper(*args, **kwargs):
        # Compute the hash of the source code
        func_hash = hash_source_code(func)
        cache_key = (func_hash, args, frozenset(kwargs.items()))

        # Define the cacheable function
        @memory.cache()
        def cached_call(func_hash, func_args, func_kwargs):
            return func(*func_args, **func_kwargs)

        return cached_call(func_hash, args, kwargs)

    return wrapper

# Example usage
@cache_source
@ell.simple(model="gpt-4o")
def hello(name: str):
    """You are a helpful assistant."""
    return f"Say hello to {name}!"

greeting = hello("Sam Altman")
print(greeting)

Now, modifying the prompt will avoid re-using the stale API result.

[gwpl]

cross-linking: #200

[cthorrez]

I wonder if something has changed in ell since this, when I use the above method I get an error.
CacheWarning: Unable to cache to disk. Possibly a race condition in the creation of the directory. Exception: 'NoneType' object is not iterable.