Prompt Functions

Defining chat interactions as python functions is a great abstraction on top of chat pipelines, allowing you to leverage parsing logic and type hints to define a code-less function that will use a generator/pipline underneath to produce an output.

A Prompt is typically created using one of the following decorators:

• @rg.prompt - Optionally takes a generator id, generator, or pipeline.
• @generator.prompt - Use this generator when executing the prompt.
• @pipeline.prompt - Use this pipeline when executing the prompt.

import rigging as rg

@rg.prompt(generator_id="mistral/mistral-medium-latest")
def summarize(text: str) -> str:
    """Summarize this text."""

Prompts are optionally bound to a pipeline/generator underneath, hence the generator and pipeline decorator variants, but they don’t have to be. We refer to bound prompts as “standalone”, because they can be executed directly as functions. Otherwise, you are required first “bind” the prompt to a specific generator id, generator, or pipeline to make it callable. Do this with Prompt.bind or related methods.

import rigging as rg

@rg.prompt
def summarize(text: str) -> str:
    """Summarize this text."""

generator = rg.get_generator("gpt-4o-mini")

await summarize.bind(generator)("...")

​

Templates and Docstrings

Underneath, the function signature will be analyzed for inputs and outputs, and the docstring will be used to create a final jinja2 template which will be used as the prompt text. You can always access this .template attribute to inspect how the prompt will be formatted.

Here are the general processing docstring rules:

• Any docstring content will always be at the top of the prompt.
• If no docstring is provided, a generic one will be substituted.
• Any inputs not explicitly defined in the docstring will be appended after the docstring.

import rigging as rg

@rg.prompt
def summarize(text: str) -> str:
    ...

print(summarize.template)

# Convert the following inputs to outputs (summarize).
#
# {{ text }}
#
# Produce the following output:
#
# <str></str>

In other words, you can define how inputs will be included in the prompt by passing them inside the docstring in jinja2 formats, or let Rigging handle this for you by ommiting them.

​

Outputs and Context

In the example above, you’ll notice that defining our function to output a str results in the following text be appended to the prompt template:

# Produce the following output:
#
# <str></str>

This is pretty light on context, and we can improve this by updating our signature with a Ctx annotation:

from typing import Annotated
import rigging as rg

Summary = Annotated[str, rg.Ctx(tag="summary", example="[2-3 sentences]")]

@rg.prompt
def summarize(text: Annotated[str, rg.Ctx(tag="long-text")]) -> Summary:
    """Summarize this text."""

print(summarize.template)

# Summarize this text.
#
# {{ long_text }}
#
# Produce the following output:
#
# <summary>[2-3 sentences]</summary>

We can apply Ctx annotations to any of the inputs and outputs of a prompt. We can override the xml tag, provide an example, and add prefix text.

Output processing is optional, and can be omitted by returning a Chat object from the wrapped function. This allows you to do with the generated output as you please.

import rigging as rg

@rg.prompt
def summarize(text: str) -> rg.Chat:
    """Summarize this text."""

print(summarize.template)

# Summarize this text.
#
# {{ text }}

​

Complex Outputs

You can also define more complex outputs by using a rigging model, list, tuple, or dataclass. Not every construction will be supported, and we attempt to pre-validate the output structure to ensure it can be processed correctly.

import rigging as rg

class User(rg.Model):
    name: str
    email: str
    age: int

@rg.prompt
def generate_user() -> User:
    """Generate a fake test user."""

print(generate_user.template)

# Generate a fake test user.
#
# Produce the following output:
#
# <user>
#   <name/>
#   <email/>
#   <age/>
# </user>

You can also embed a Chat object inside other objects, which will be excluded from any prompt guidance, but supplied the value when the prompt is executed. This is great for gathering both structured data and the original chat.

import rigging as rg

joke = Annotated[str, rg.Ctx(tag="joke")]

@rg.prompt
def tell_joke() -> tuple[joke, rg.Chat]:
    """Tell a joke."""

print(tell_joke.template)

# Tell a joke.
#
# Produce the following output:
#
# <joke></joke>

​

Rendering Prompts

In addition to templates, you can use .render with valid inputs to view the exact prompt as it will be sent to a generator. You can also use this to pass your prompt into pipelines at your discretion.

from typing import Annotated
import rigging as rg

email = Annotated[str, rg.Ctx(tag="email")]

@rg.prompt
def convert_to_email(name: str, top: int = 5) -> list[email]:
    """Convert this name into the best {{ top }} email addresses."""

print(convert_to_email.render("John Doe"))

# Convert this name into the best 5 email addresses.
#
# <name>John Doe</name>
#
# Produce the following output for each item:
#
# <email></email>

​

Running Prompts

Prompt objects expose the following methods for execution:

• Prompt.run() (Aliased with __call__)
• Prompt.run_many()
• Prompt.run_over()

(Available if the prompt was supplied/bonded to a pipeline or generator)

You can also bind a prompt at runtime with any of the following:

• Prompt.bind()
• Prompt.bind_many()
• Prompt.bind_over()

import rigging as rg

@rg.prompt(generator_id="claude-3-sonnet-20240229")
def write_code(description: str, language: str = "python") -> str:
    """Write a single function."""

code = await write_code("Calculate the factorial of a number.")