banks

Banks is the linguist professor who will help you generate meaningful LLM prompts using a template language that makes sense. If you're still using f-strings for the job, keep reading.

Docs are available here.

Table of Contents

• banks
  • Installation
  • Features
  • Cookbooks
  • Examples
    • :point_right: Create a blog writing prompt
    • :point_right: Create a summarizer prompt
    • :point_right: Lemmatize text while processing a template
    • :point_right: Use a LLM to generate a text while rendering a prompt
    • :point_right: Go meta: create a prompt and generate its response
    • :point_right: Go meta(meta): process a LLM response
  • Reuse templates from registries
  • Async support
  • License

Installation

pip install banks

Features

Prompts are instrumental for the success of any LLM application, and Banks focuses around specific areas of their lifecycle:

• :orange_book: Templating: Banks provides tools and functions to build prompts text and chat messages from generic blueprints.
• :tickets: Versioning and metadata: Banks supports attaching metadata to prompts to ease their management, and versioning is first-class citizen.
• :file_cabinet: Management: Banks provides ways to store prompts on disk along with their metadata.

Cookbooks

• :blue_book: Prompt caching with Anthropic
• :blue_book: Prompt versioning

Examples

:point_right: Create a blog writing prompt

Given a generic template to instruct an LLM to generate a blog article, we use Banks to generate the actual prompt on our topic of choice, "retrogame computing":

from banks import Prompt


p = Prompt("Write a 500-word blog post on {{ topic }}.\n\nBlog post:")
topic = "retrogame computing"
print(p.text({"topic": topic}))

This will print the following text, that can be pasted directly into Chat-GPT:

Write a 500-word blog post on retrogame computing.

Blog post:

The same prompt can be written in form of chat messages:

prompt_text = """{% chat role="system" %}
I want you to act as a title generator for written pieces.
{% endchat %}

{% chat role="user" %}
Write a 500-word blog post on {{ topic }}.

Blog post:
{% endchat %}"""

p = Prompt(prompt_text)
print(p.chat_messages({"topic":"prompt engineering"}))

This will output the following:

[
  ChatMessage(role='system', content='I want you to act as a title generator for written pieces.\n'),
  ChatMessage(role='user', content='Write a 500-word blog post on .\n\nBlog post:\n')
]

:point_right: Create a summarizer prompt

Instead of hardcoding the content to summarize in the prompt itself, we can inject it starting from a generic one:

from banks import Prompt


prompt_template = """
Summarize the following documents:
{% for document in documents %}
{{ document }}
{% endfor %}
Summary:
"""

# In a real-world scenario, these would be loaded as external resources from files or network
documents = [
    "A first paragraph talking about AI",
    "A second paragraph talking about climate change",
    "A third paragraph talking about retrogaming"
]

p = Prompt(prompt_template)
print(p.text({"documents": documents}))

The resulting prompt:

Summarize the following documents:

A first paragraph talking about AI

A second paragraph talking about climate change

A third paragraph talking about retrogaming

Summary:

:point_right: Lemmatize text while processing a template

Banks comes with predefined filters you can use to process data before generating the prompt. Say you want to use a lemmatizer on a document before summarizing it, first you need to install simplemma:

pip install simplemma

then you can use the lemmatize filter in your templates like this:

from banks import Prompt


prompt_template = """
Summarize the following document:
{{ document | lemmatize }}
Summary:
"""

p = Prompt(prompt_template)
print(p.text({"document": "The cats are running"}))

the output would be:

Summarize the following document:
the cat be run
Summary:

:point_right: Use a LLM to generate a text while rendering a prompt

Sometimes it might be useful to ask another LLM to generate examples for you in a few-shot prompt. Provided you have a valid OpenAI API key stored in an env var called OPENAI_API_KEY you can ask Banks to do something like this (note we can annotate the prompt using comments - anything within {# ... #} will be removed from the final prompt):

from banks import Prompt


prompt_template = """
Generate a tweet about the topic {{ topic }} with a positive sentiment.

{#
    This is for illustration purposes only, there are better and cheaper ways
    to generate examples for a few-shots prompt.
#}
Examples:
{% for number in range(3) %}
- {% generate "write a tweet with positive sentiment" "gpt-3.5-turbo" %}
{% endfor %}
"""

p = Prompt(prompt_template)
print(p.text({"topic": "climate change"}))

The output would be something similar to the following:

Generate a tweet about the topic climate change with a positive sentiment.


Examples:

- "Feeling grateful for the amazing capabilities of #GPT3.5Turbo! It's making my work so much easier and efficient. Thank you, technology!" #positivity #innovation

- "Feeling grateful for all the opportunities that come my way! With #GPT3.5Turbo, I am able to accomplish tasks faster and more efficiently. #positivity #productivity"

- "Feeling grateful for all the wonderful opportunities and experiences that life has to offer! #positivity #gratitude #blessed #gpt3.5turbo"

If you paste Banks' output into ChatGPT you would get something like this:

Climate change is a pressing global issue, but together we can create positive change! Let's embrace renewable energy, protect our planet, and build a sustainable future for generations to come. 🌍💚 #ClimateAction #PositiveFuture

[!IMPORTANT] The generate extension uses LiteLLM under the hood, and provided you have the proper environment variables set, you can use any model from the supported model providers.

[!NOTE] Banks uses a cache to avoid generating text again for the same template with the same context. By default the cache is in-memory but it can be customized.

:point_right: Go meta: create a prompt and generate its response

We can leverage Jinja's macro system to generate a prompt, send the result to OpenAI and get a response. Let's bring back the blog writing example:

from banks import Prompt

prompt_template = """
{% from "banks_macros.jinja" import run_prompt with context %}

{%- call run_prompt() -%}
Write a 500-word blog post on {{ topic }}

Blog post:
{%- endcall -%}
"""

p = Prompt(prompt_template)
print(p.text({"topic": "climate change"}))

The snippet above won't print the prompt, instead will generate the prompt text

Write a 500-word blog post on climate change

Blog post:

and will send it to OpenAI using the generate extension, eventually returning its response:

Climate change is a phenomenon that has been gaining attention in recent years...
...

:point_right: Go meta(meta): process a LLM response

When generating a response from a prompt template, we can take a step further and post-process the LLM response by assinging it to a variable and applying filters to it:

from banks import Prompt

prompt_template = """
{% from "banks_macros.jinja" import run_prompt with context %}

{%- set prompt_result %}
{%- call run_prompt() -%}
Write a 500-word blog post on {{ topic }}

Blog post:
{%- endcall -%}
{%- endset %}

{# nothing is returned at this point: the variable 'prompt_result' contains the result #}

{# let's use the prompt_result variable now #}
{{ prompt_result | upper }}
"""

p = Prompt(prompt_template)
print(p.text({"topic": "climate change"}))

The final answer from the LLM will be printed, this time all in uppercase.

Reuse templates from registries

We can get the same result as the previous example loading the prompt template from a registry instead of hardcoding it into the Python code. For convenience, Banks comes with a few registry types you can use to store your templates. For example, the DirectoryTemplateRegistry can load templates from a directory in the file system. Suppose you have a folder called templates in the current path, and the folder contains a file called blog.jinja. You can load the prompt template like this:

from banks import Prompt
from banks.registries import DirectoryTemplateRegistry

registry = DirectoryTemplateRegistry(populated_dir)
prompt = registry.get(name="blog")

print(prompt.text({"topic": "retrogame computing"}))

Async support

To run banks within an asyncio loop you have to do two things:

1. set the environment variable BANKS_ASYNC_ENABLED=true.
2. use the AsyncPrompt class that has an awaitable run method.

Example:

from banks import AsyncPrompt

async def main():
    p = AsyncPrompt("Write a blog article about the topic {{ topic }}")
    result = await p.text({"topic": "AI frameworks"})
    print(result)

asyncio.run(main())

License

banks is distributed under the terms of the MIT license.