@ell.simple¶

The core unit of prompt engineering in ell is the @ell.simple decorator. This decorator transforms a function that provides system and user prompts into a callable object. When invoked, this callable sends the provided prompts to a language model and returns the model’s response.

The development of @ell.simple is driven by several important objectives:

Improve readability and usabiltiy of prompt engineering code.
Force a functional decomposition of prompt systems into reusable components.
Enable versioning, serialization, and tracking of prompts over time

Usage¶

The @ell.simple decorator can be used in two main ways:

Using the docstring as the system prompt, and the return value as the user message:

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

Explicitly defining messages:

@ell.simple(model="gpt-4")
def hello(name: str):
    return [
        ell.system("You are a helpful assistant."),
        ell.user(f"Say hello to {name}!")
    ]

Note

Messages in ell are not the same as the dictionary messages used in the OpenAI API. ell’s Message API provides a more intuitive and flexible way to construct and manipulate messages. You can read more about ell’s Message API and type coercion in the Messages page.

Invoking an `ell.simple` LMP¶

To use the decorated function, we can call it as a normal function. However, instead of receiving the typical return value, we will receive the result of passing the system and user prompts directly to the model specified in the decorator constructor, in this case GPT-4.

>>> hello("world")
'Hello, world!'

As you can see from this example, the return type of an ell.simple LMP is a string. This is to optimize for readability and usability, as most invocations of language models revolve around passing strings around. Additional metadata is only needed occasionally.

Therefore, we have two decorators within the ell framework:

@ell.simple: Returns simple strings, as shown here.
@ell.complex: Returns message objects containing all of the typical message API metadata and additional helper functions for interacting with multimodal output data. You can read more about this in the @ell.complex page.

Variable system prompts¶

One of the challenges with specifying the system prompt in the docstring of a language model program is that if you want to use variable system prompts, Python will no longer treat the string literal at the top of the function as a docstring. For example:

def my_func(var : int):
    f"""my variable doc string for my_func. {var}"""
    pass

>>> my_func.__doc__
None

This behavior makes sense because a function’s docstring should not change during execution and should be extractable through static analysis.

To address this issue with @ell.simple, you need to use the second method of defining an ell.simple language model program by creating a function that returns a list of messages (see Messages for more details).

@ell.simple(model="gpt-4")
def my_func(name : str, var : int):
    return [
        ell.system(f"You are a helpful assistant. {var}"),
        ell.user(f"Say hello to {name}!")
    ]

With this approach, ell will ignore the docstring of my_func and instead supply the messages returned by the function to the language model API.

Passing parameters to an LLM API¶

One of the most convenient functions of the @ell.simple decorator is that you can easily pass parameters to an LLM API, both at definition time and runtime. For example, models within the OpenAI API have parameters like temperature, max_tokens, stop tokens, and logit_bias. Due to how @ell.simple works, you can simply specify these in the decorator as keyword arguments.

@ell.simple(model="gpt-4", temperature=0.5, max_tokens=100, stop=["."])
def hello(name: str):
    """You are a helpful assistant."""
    return f"Hey there {name}!"

Likewise, if you want to modify those parameters for a particular invocation of that prompt, you simply pass them in as api_params keyword arguments to the function when calling it. For example:

>>> hello("world", api_params=dict(temperature=0.7))
'Hey there world!'

Multiple outputs (n>1)¶

As is often important in prompt engineering to leverage test-time compute, many language model APIs allow you to specify a count parameter, usually ‘n’, which will generate several outputs from the language model given a particular prompt.

In the OpenAI API, for example, this is actually quite cumbersome because the API specification separates different completions into ‘choices’ objects. For example:

response = openai.Completion.create(
    model="gpt-4",
    prompt="Say hello to everyone",
    n=2
)

r1 = response.choices[0].text
r2 = response.choices[1].text

In the spirit of simplicity, we’ve designed it to automatically coerce the return type into the correct shape, similar to NumPy and PyTorch. This means that when you call an ell.simple language model program with n greater than one, instead of returning a string, it returns a list of strings.

@ell.simple(model="gpt-4", n=2)
def hello(name: str):
    """You are a helpful assistant."""
    return f"Say hello to {name}!"

>>> hello("world")
['Hey there world!', 'Hi, world.']

Similarly, this behavior applies when using runtime api_params to specify multiple outputs.

>>> hello("world", api_params=dict(n=3))
['Hey there world!', 'Hi, world.', 'Hello, world!']

Note

In the future, we may modify this interface as preserving the api_params keyword in its current form could potentially lead to conflicts with user-defined functions. However, during the beta phase, we are closely monitoring for feedback and will make adjustments based on user experiences and needs.

Multimodal inputs¶

@ell.simple supports multimodal inputs, allowing you to easily work with both text and images in your language model programs. This is particularly useful for models with vision capabilities, such as GPT-4 with vision.

Here’s an example of how to use @ell.simple with multimodal inputs:

from PIL import Image
import ell
from ell.types.message import ImageContent

@ell.simple(model="gpt-4-vision-preview")
def describe_image(image: Image.Image):
    return [
        ell.system("You are a helpful assistant that describes images."),
        ell.user(["What's in this image?", image])
        # Or ell.user(["What's in this image?", ImageContent(url=image_url, detail="low")])
    ]

# Usage with PIL Image
image = Image.open("path/to/your/image.jpg")
description = describe_image(image)
print(description)  # This will print a text description of the image

In these examples, the describe_image function takes a PIL Image object as input, while describe_image_url takes a string URL. The ell.user message combines both text and image inputs. @ell.simple automatically handles the conversion of the PIL Image object or ImageContent into the appropriate format for the language model.

This approach simplifies working with multimodal inputs, allowing you to focus on your application logic rather than the intricacies of API payloads.

Note

Not all language model providers support image URLs. For example, as of the current version, Anthropic’s models do not support image URLs. Always check the capabilities and requirements of your chosen language model provider when working with multimodal inputs.

Warning

While @ell.simple supports multimodal inputs, it is designed to return text-only outputs. For handling multimodal outputs (such as generated images or audio), you need to use @ell.complex. Please refer to the @ell.complex documentation for more information on working with multimodal outputs.

What about multiturn conversations, tools, structured outputs, and other features?¶

While @ell.simple is great for straightforward text-based interactions with language models, there are scenarios where you might need more complex functionality. For instance, you may want to work with multiturn conversations, utilize tools, generate structured outputs, or handle multimodal content beyond just text.

In such cases, you’ll need an LMP that can return rich Message objects instead of just strings. This is where @ell.complex comes into play. The @ell.complex decorator provides enhanced capabilities for more sophisticated interactions with language models.

For more information on how to use @ell.complex and its advanced features, please refer to the @ell.complex documentation.

Reference¶

ell.simple(model: str, client: Any | None = None, exempt_from_tracking=False, **api_params)¶

The fundamental unit of language model programming in ell.

This decorator simplifies the process of creating Language Model Programs (LMPs) that return text-only outputs from language models, while supporting multimodal inputs. It wraps the more complex ‘complex’ decorator, providing a streamlined interface for common use cases.

Parameters:

model (str) – The name or identifier of the language model to use.
client (Optional[openai.Client]) – An optional OpenAI client instance. If not provided, a default client will be used.
exempt_from_tracking (bool) – If True, the LMP usage won’t be tracked. Default is False.
api_params (Any) – Additional keyword arguments to pass to the underlying API call.

Usage: The decorated function can return either a single prompt or a list of ell.Message objects:

@ell.simple(model="gpt-4", temperature=0.7)
def summarize_text(text: str) -> str:
    '''You are an expert at summarizing text.''' # System prompt
    return f"Please summarize the following text:\n\n{text}" # User prompt


@ell.simple(model="gpt-4", temperature=0.7)
def describe_image(image : PIL.Image.Image) -> List[ell.Message]:
    '''Describe the contents of an image.''' # unused because we're returning a list of Messages
    return [
        # helper function for ell.Message(text="...", role="system")
        ell.system("You are an AI trained to describe images."),
        # helper function for ell.Message(content="...", role="user")
        ell.user(["Describe this image in detail.", image]),
    ]


image_description = describe_image(PIL.Image.open("https://example.com/image.jpg"))
print(image_description)
# Output will be a string text-only description of the image

summary = summarize_text("Long text to summarize...")
print(summary)
# Output will be a text-only summary

Notes:

This decorator is designed for text-only model outputs, but supports multimodal inputs.
It simplifies complex responses from language models to text-only format, regardless of the model’s capability for structured outputs, function calling, or multimodal outputs.
For preserving complex model outputs (e.g., structured data, function calls, or multimodal outputs), use the @ell.complex decorator instead. @ell.complex returns a Message object (role=’assistant’)
The decorated function can return a string or a list of ell.Message objects for more complex prompts, including multimodal inputs.
If called with n > 1 in api_params, the wrapped LMP will return a list of strings for the n parallel outputs of the model instead of just one string. Otherwise, it will return a single string.
You can pass LM API parameters either in the decorator or when calling the decorated function. Parameters passed during the function call will override those set in the decorator.

Example of passing LM API params:

@ell.simple(model="gpt-4", temperature=0.7)
def generate_story(prompt: str) -> str:
    return f"Write a short story based on this prompt: {prompt}"

# Using default parameters
story1 = generate_story("A day in the life of a time traveler")

# Overriding parameters during function call
story2 = generate_story("An AI's first day of consciousness", api_params={"temperature": 0.9, "max_tokens": 500})

@ell.simple¶

Usage¶

Invoking an ell.simple LMP¶

Variable system prompts¶

Passing parameters to an LLM API¶

Multiple outputs (n>1)¶

Multimodal inputs¶

What about multiturn conversations, tools, structured outputs, and other features?¶

Reference¶

Invoking an `ell.simple` LMP¶