alana 0.0.4

Utilities for Alana, and maybe you too. Mostly geared toward LLM-heavy workflows.

alana-utilities (Utilities Library for LLM-Heavy Workflows)

Install Instructions

pip install alana

⚠️ Warning: This library is in active early development! No guarantees are made for backward compatibility. The library is NOT intended to be production-ready.

Is this for me?

This library is mostly designed for me (hence the name), but I hope they're helpful for you too!

• Do you interact with LLMs? (Currently, only Anthropic LLMs are supported).
• Do you value conciseness?
• Are you writing non-production prototype code, where 1, occassional bugs and behavioral changes are acceptable, and 2, developer ergonomics are more important than performance?
  • Note: This library strongly assumes this use-case! Many functions actually write to the console by default (you can disable this with loud=False).
• Do the features appeal to you?

Philosophy:

• Programming is too slow! This is doubly true when you're interacting with LLMs. By building nice utilities with sane defaults, I hope to speed up my (and maybe your) workflow.
• I make trade-offs to speed up the developer experience:
  • I do not try hard to anticipate future upstream API changes. I'm also ok with breaking backward compatibility to make my functions more concise and more usable.
  • Usability > Principles. While I don't relish in it, I'm ok with breaking conventions designed for large production libraries if it speeds up programmers who use alana. The priority is to make the library intuitive and fast.
  • I don't try to serve every use-case.
• Simplicity is key. This library strives to be readable and straightforward.

Motivating Examples:

(I tested these, and tried to make sure my code was idiomatic in all cases. Sorry if I messed up! Please report any issues to h i ( a t ) [pip package name for this library] dot computer.)

Continuing a list of messages using Anthropic API (adapted from Anthropic API documentation:

import anthropic
from anthropic.types import MessageParam
messages = [
  MessageParam(
    role="user",
    content="Hello, Claude!"
  ),
]
output_text = anthropic.Anthropic().messages.create(
    model="claude-3-opus-20240229",
    max_tokens=1024,
    messages=messages
).content[0].text
messages.append(
  MessageParam(
    role="assistant",
    content=output_text
  )
)
print(messages)  # [{'role': 'user', 'content': 'Hello, Claude!'}, {'role': 'assistant', 'content': "Hello! It's nice to meet you. How are you doing today?"}]

Equivalent alana code:

import alana
messages = []
alana.gen(user="Hello, Claude!", messages=messages, model="opus", max_tokens=1024)
print(messages)  # [{'role': 'user', 'content': 'Hello, Claude!'}, {'role': 'assistant', 'content': "Hello! It's nice to meet you. How are you doing today?"}]

Also, equivalent alana code thanks to defaults:

import alana
messages = []
alana.gen(user="Hello, Claude!", messages=messages)
print(messages)  # [{'role': 'user', 'content': 'Hello, Claude!'}, {'role': 'assistant', 'content': "Hello! It's nice to meet you. How are you doing today?"}]

Features:

• Easy color print: alana.red, alana.green, alana.blue, alana.yellow, alana.cyan. Try alana.green("Hello!")
• Easy pretty print with Sonnnet (or an Anthropic model of your choice): alana.pretty_print. Try alana.pretty_print(t.arange(16, device='cpu').reshape(2,2,4))
• Make it easier to use the Anthropic API:
  • alana.gen, for easy Claude generations. Try alana.gen(user="Hello, Claude!"). You can pass in a messages parameter either in place of or together with a user parameter. ⚠️ messages support is not thoroughly tested yet!
  • alana.gen_examples, alana.gen_examples_list for generating few-shot examples.
  • alana.gen_prompt, for easy prompt generation (meta-prompt).
  • alana.get_xml, for using regex to get XML tag contents from model outputs. ⚠️ Regex parsing of XML may be unreliable!
  • alana.remove_xml to strip certain XML tag-enclosed content from a string (along with the tags). This is primarily intended to get rid of "..." strings. ⚠️ Regex parsing of XML may be unreliable! This function has not been thoroughly tested yet!
• A bunch of aliases (Try: alana.few_shot, alana.n_shot, or alana.xml)

Coming Soon:

• Improve docs
• Generating alternative prompts given a prompt
• Better support for multi-turn prompting
• OpenAI model support
• Support for automatic "are you sure"/"are you confused" multi-turn prompting
• Automatic error checking (are there mistakes in this code, sanity checking of model outputs)
• Automatic model-switching on rate limit
• Support for quick-and-dirty unit-testing with Claude!
• Prompt test case generation and easy prompt testing