llm-axe

Goal

llm-axe is meant to be a flexible toolkit that provides simple abstractions for commonly used functions related to LLMs. It's not meant to intrude in your development workflow as other larger frameworks often do.

It has functions for automatic schema generation, pre-made agents with self-tracking chat history and fully customizable agents.

Have feedback/questions? Join the Discord

Read the Development Documentation

Installation

pip install llm-axe

Example Snippets

• Easily Work With Non-Persistent Embeddings:

from llm_axe import read_pdf, find_most_relevant, split_into_chunks
text = read_pdf("./super_long_text.pdf")
sentences = split_into_chunks(text, 3)
pairs = []
for chunk in sentences:
    embeddings = client.embeddings(model='nomic-embed-text', prompt=chunk)["embedding"]
    pairs.append((chunk, embeddings))

prompt = "What do the Hobbit traditions say about second breakfast?"
prompt_embedding = client.embeddings(model='nomic-embed-text', prompt=prompt)["embedding"]
relevant_texts = find_most_relevant(pairs, prompt_embedding, top_k=4)

• Function Calling

  A function calling LLM can be created with just 3 lines of code:
  No need for premade schemas, templates, special prompts, or specialized functions.

prompt = "I have 500 coins, I just got 200 more. How many do I have?"

llm = OllamaChat(model="llama3:instruct")
fc = FunctionCaller(llm, [get_time, get_date, get_location, add, multiply])
result = fc.get_function(prompt)

• Custom Agent

llm = OllamaChat(model="llama3:instruct")
agent = Agent(llm, custom_system_prompt="Always respond with the word LLAMA, no matter what")
resp = agent.ask("What is the meaning of life?")
print(resp)

# Output
# LLAMA

• Online Agent

prompt = "Tell me a bit about this website:  https://toscrape.com/?"
llm = OllamaChat(model="llama3:instruct")
searcher = OnlineAgent(llm)
resp = searcher.search(prompt)

#output: Based on information from the internet, it appears that https://toscrape.com/ is a website dedicated to web scraping.
# It provides a sandbox environment for beginners and developers to learn and validate their web scraping technologies...

• PDF Reader

llm = OllamaChat(model="llama3:instruct")
files = ["../FileOne.pdf", "../FileTwo.pdf"]
agent = PdfReader(llm)
resp = agent.ask("Summarize these documents for me", files)

• Data Extractor

llm = OllamaChat(model="llama3:instruct")
info = read_pdf("../Example.pdf")
de = DataExtractor(llm, reply_as_json=True)
resp = de.ask(info, ["name", "email", "phone", "address"])

#output: {'Name': 'Frodo Baggins', 'Email': 'frodo@gmail.com', 'Phone': '555-555-5555', 'Address': 'Bag-End, Hobbiton, The Shire'}

• Object Detector

llm = OllamaChat(model="llava:7b")
detector = ObjectDetectorAgent(llm, llm)
resp = detector.detect(images=["../img2.jpg"], objects=["sheep", "chicken", "cat", "dog"])

#{
#  "objects": [
#    { "label": "Sheep", "location": "Field", "description": "White, black spots" },
#    { "label": "Dog", "location": "Barn", "description": "Brown, white spots" }
#  ]
#}

See more complete examples

How to setup llm-axe with your own LLM

Features

• Local LLM internet access with Online Agent
• PDF Document Reader Agent
• Premade utility Agents for common tasks
• Compatible with any LLM, local or externally hosted
• Built-in support for Ollama

Important Notes

The results you get from the agents are highly dependent on the capability of your LLM. An inadequate LLM will not be able to provide results that are usable with llm-axe

Testing in development was done using llama3 8b:instruct 4 bit quant