Prompt Function
Marvin puts the engineering in prompt engineering. We expose a low-level prompt_fn decorator that lets you write prompts as functions. This lets you build fully type-hinted prompts that other engineers can introspect, version, and test.
This is the easiest way to use Azure / OpenAI's function calling API.
Basic Use¶
Type Hinting¶
Example
Marvin translates your Python code into English. We'll simply write a Python function,
tell it that we expect an integer input n input and that it'll
output list[str], or a list of strings. With Marvin, we'll use prompt_fn and decorate this function.
When we do, this function can be cast to a payload that can be send to an LLM.
from marvin.prompts import prompt_fn
@prompt_fn
def list_fruits(n: int, color: str = 'red') -> list[str]:
'''Generates a list of {{ n }} {{ color }} fruits'''
list_fruits(3, color = 'blue')
Click to see results: list_fruits(3, color = 'blue')
{
"messages": [
{
"role": "system",
"content": "Generates a list of 3 blue fruits"
}
],
"functions": [
{
"parameters": {
"type": "object",
"properties": {
"output": {
"title": "Output",
"type": "array",
"items": {
"type": "string"
}
}
},
"required": [
"output"
]
},
"name": "Output",
"description": ""
}
],
"function_call": {
"name": "Output"
}
}
Advanced Use¶
Use with Pydantic¶
Example
Marvin supports type-hinting with Pydantic, so your return annotation can be a more complex data-model.
from marvin.prompts import prompt_fn
from pydantic import BaseModel
class Fruit(BaseModel):
color: str
@prompt_fn
def list_fruits(n: int, color: str = 'red') -> list[Fruit]:
'''Generates a list of {{ n }} {{ color }} fruits'''
list_fruits(3, color = 'blue')
Click to see results: list_fruits(3, color = 'blue')
{
"messages": [
{
"role": "system",
"content": "Generates a list of 3 blue fruits"
}
],
"functions": [
{
"parameters": {
"type": "object",
"properties": {
"output": {
"title": "Output",
"type": "array",
"items": {
"$ref": "#/definitions/Fruit"
}
}
},
"required": [
"output"
],
"definitions": {
"Fruit": {
"title": "Fruit",
"type": "object",
"properties": {
"color": {
"title": "Color",
"type": "string"
}
},
"required": [
"color"
]
}
}
},
"name": "FruitList",
"description": ""
}
],
"function_call": {
"name": "FruitList"
}
}
Full Customization¶
Example
Marvin supports full customization of every element of your prompts. You can customize the
name, description, and field_names of your response_model.
Say we want to change the task to generate in Swedish.
from marvin.prompts import prompt_fn
from pydantic import BaseModel
class Fruit(BaseModel):
color: str
@prompt_fn(
response_model_name = 'Fruktlista',
response_model_description = 'A list of fruits in Swedish',
response_model_field_name = 'Frukt'
)
def list_fruits(n: int, color: str = 'red') -> list[Fruit]:
'''Generates a list of {{ n }} {{ color }} fruits'''
list_fruits(3, color = 'blue')
Click to see results: list_fruits(3, color = 'blue')
{
"messages": [
{
"role": "system",
"content": "Generates a list of 3 blue fruits"
}
],
"functions": [
{
"parameters": {
"type": "object",
"properties": {
"Frukt": {
"title": "Frukt",
"type": "array",
"items": {
"$ref": "#/definitions/Fruit"
}
}
},
"required": [
"Frukt"
],
"definitions": {
"Fruit": {
"title": "Fruit",
"type": "object",
"properties": {
"color": {
"title": "Color",
"type": "string"
}
},
"required": [
"color"
]
}
}
},
"name": "Fruktlista",
"description": "A list of fruits in Swedish"
}
],
"function_call": {
"name": "Fruktlista"
}
}
Referencing Globals¶
Example
Marvin passes the name of your response model to the prompt for you to reference as a convenience.
Say we want to change the task to generate in Swedish.
from marvin.prompts import prompt_fn
from pydantic import BaseModel
class Fruit(BaseModel):
color: str
@prompt_fn(response_model_name = 'Fruits')
def list_fruits(n: int, color: str = 'red') -> list[Fruit]:
'''Generates a list of {{ n }} {{ color }} {{ response_model.__name__.lower() }}'''
list_fruits(3, color = 'blue')
Click to see results: list_fruits(3, color = 'blue')
{
"messages": [
{
"role": "system",
"content": "Generates a list of 3 blue fruits"
}
],
"functions": [
{
"parameters": {
"type": "object",
"properties": {
"output": {
"title": "Output",
"type": "array",
"items": {
"$ref": "#/definitions/Fruit"
}
}
},
"required": [
"output"
],
"definitions": {
"Fruit": {
"title": "Fruit",
"type": "object",
"properties": {
"color": {
"title": "Color",
"type": "string"
}
},
"required": [
"color"
]
}
}
},
"name": "Fruits",
"description": ""
}
],
"function_call": {
"name": "Fruits"
}
}
Contexts¶
Example
Marvin supports full passing context dictionaries to your prompt's rendering environment.
Say we want to list 'seasonal' fruits. We'll pass the datetime.
from marvin.prompts import prompt_fn
from datetime import date
@prompt_fn(ctx = {'today': date.today()})
def list_fruits(n: int, color: str = 'red') -> list[str]:
'''
Generates a list of {{ n }} {{ color }} fruits in season.
- The date is {{ today }}
'''
This function can now be run and serialized to an Azure / OpenAI Function Calling payload.
Click to see results: list_fruits(3, color = 'blue')
{
"messages": [
{
"role": "system",
"content": "Generates a list of 3 blue fruits in season.\n- The date is 2023-09-22"
}
],
"functions": [
{
"parameters": {
"type": "object",
"properties": {
"output": {
"title": "Output",
"type": "array",
"items": {
"type": "string"
}
}
},
"required": [
"output"
]
},
"name": "Output",
"description": ""
}
],
"function_call": {
"name": "Output"
}
}
Multi-Turn Prompts¶
Example
Marvin supports multi-turn conversations. If no role is specified, the whole block is assumed to be a system prompt. To override this default behavior, simply break into Human, System, Assistant turns.
from marvin.prompts import prompt_fn
from datetime import date
@prompt_fn(ctx = {'today': date.today()})
def list_fruits(n: int, color: str = 'red') -> list[str]:
'''
System: You generate a list of {{ n }} fruits in season.
- The date is {{ today }}
User: I want {{ color }} fruits only.
'''
This function can now be run and serialized to an Azure / OpenAI Function Calling payload.
Click to see results: list_fruits(3, color = 'blue')
{
"messages": [
{
"role": "system",
"content": "You generate a list of 3 fruits in season.\n- The date is 2023-09-22"
},
{
"role": "user",
"content": "I want blue fruits."
}
],
"functions": [
{
"parameters": {
"type": "object",
"properties": {
"output": {
"title": "Output",
"type": "array",
"items": {
"type": "string"
}
}
},
"required": [
"output"
]
},
"name": "Output",
"description": ""
}
],
"function_call": {
"name": "Output"
}
}
Use Cases¶
Classification¶
Example
Marvin supports multi-turn conversations. If no role is specified, the whole block is assumed to be a system prompt. To override this default behavior, simply break into Human, System, Assistant turns.
from marvin.prompts import prompt_fn
from typing import Optional
from enum import Enum
class Food(Enum):
'''
Food classes
'''
FRUIT = 'Fruit'
VEGETABLE = 'Vegetable'
@prompt_fn
def classify_fruits(food: str) -> Food:
'''
Expertly determines the class label of {{ food }}.
'''
This function can now be run and serialized to an Azure / OpenAI Function Calling payload.
Click to see results: classify_fruits('tomato')
{
"messages": [
{
"role": "system",
"content": "Expertly determines the class label of tomato."
}
],
"functions": [
{
"parameters": {
"type": "object",
"properties": {
"output": {
"$ref": "#/definitions/Food"
}
},
"required": [
"output"
],
"definitions": {
"Food": {
"title": "Food",
"description": "Food classes",
"enum": [
"Fruit",
"Vegetable"
]
}
}
},
"name": "Output",
"description": ""
}
],
"function_call": {
"name": "Output"
}
}
Entity Extraction¶
Example
In this example, Marvin is configured to perform entity extraction on a list of fruits mentioned in a text string. The function extract_fruits identifies and extracts fruit entities based on the input text. These entities are then returned as a list of Pydantic models.
from marvin.prompts import prompt_fn
from typing import List
from pydantic import BaseModel
class FruitEntity(BaseModel):
'''
Extracted Fruit Entities
'''
name: str
color: str
@prompt_fn
def extract_fruits(text: str) -> List[FruitEntity]:
'''
Extracts fruit entities from the given text: {{ text }}.
'''
This function can now be run and serialized to an Azure / OpenAI Function Calling payload.
extract_fruits('There are red apples and yellow bananas.')
{
"messages": [
{
"role": "system",
"content": "Extracts fruit entities from the given text: There are red apples and yellow bananas.."
}
],
"functions": [
{
"parameters": {
"type": "object",
"properties": {
"output": {
"title": "Output",
"type": "array",
"items": {
"$ref": "#/definitions/FruitEntity"
}
}
},
"required": [
"output"
],
"definitions": {
"FruitEntity": {
"title": "FruitEntity",
"description": "Extracted Fruit Entities",
"type": "object",
"properties": {
"name": {
"title": "Name",
"type": "string"
},
"color": {
"title": "Color",
"type": "string"
}
},
"required": [
"name",
"color"
]
}
}
},
"name": "Output",
"description": ""
}
],
"function_call": {
"name": "Output"
}
}