Working with Large Language Models (LLMs) can be challenging, but what if I told you there are ways of making it infinitely easier? Some LLM cheat codes?

You don’t have to be a gamer to understand the concept of cheat codes.They’re what lets you unlock hidden features and change the way you play the game by removing some existing limitations. Shortcuts, extra resources, you name it!

When working with an LLM, a few cheat codes can really help make your life easier. Easier to build, easier to test, easier to deploy, and easier to use. Think about it like optimizing. Work smart, not hard.

That’s what frameworks are for. Like cheat codes, LLM frameworks are designed to make your applications more versatile, efficient, scalable, and much more. All you need to know is which ones should be at the top of your list.

Look no further! Keep reading to learn more about frameworks and the top five you should be using with LLMs.

Table of Content

• What are frameworks?
• Understanding Abstraction
• LangChain
• LlamaIndex
• HuggingFace Transformers
• TensorFlow
• PyTorch

Is Tech Right For you? Take Our 3-Minute Quiz!

You Will Learn:

☑️ If a career in tech is right for you

☑️ What tech careers fit your strengths

☑️ What skills you need to reach your goals

What are frameworks?

Frameworks are a collection of libraries, tools, and features that make it easier for developers to build applications that use large language models (LLMs). Let’s start with a quick LLM crash course.

A large language model is a type of generative AI used to generate language. LLMs use natural language processing (NLP) to understand natural human languages so computers can process, analyze, and interpret both written and spoken language. And yes, that includes coding languages like Python and JavaScript!

When a user inputs text into an application like ChatGPT, the AI uses an LLM to create text output that fits the user prompt. For example, you could ask ChatGPT to “Write a Python script to automate sending daily email reports,” and it can walk you through the steps and/or create the script for you (note: not all AI are created equal and you’ll still want to double-check their work!). LLMs like GPT-3.5 or GPT-4 have this ability because developers used a combination of machine learning and deep learning to train these LLMs on huge — and we mean massive — amounts of data.

Developing, training, and deploying LLMs and their applications aren’t the easiest projects to manage, but the right LLM framework can do a lot of the heavy lifting for you.

To really understand frameworks, you have to understand “abstraction.”

Understanding Abstraction

No matter which language you’re using, coding complex processes require a lot of code. LLM frameworks utilize abstractions to replace complex code with simplified code and syntax. Consider it a shortcut that makes your life — and the work of your LLM — much more simple.

π (pi) is an example of an abstraction that represents the ratio of a circle’s circumference to its diameter. We can say π instead of the infinite number of decimal points and we still know what it means.

A more common example of abstraction in the real world is “driving to work.” You start the car, maneuver it to drive, and use the pedals. You don’t have to know the inner workings of the motor that make the car move. The idea of abstraction is that the processes — no matter how complicated they are — are contained by a component. And these components can then be linked together to create an app.

Frameworks are a library of abstractions. They represent the steps and concepts you need to work with LLMs and make your code more flexible, readable, maintainable, and scalable.

Ready to conquer LLMs? Here are five of the best frameworks you need to check out:

1. LangChain

LangChain is an open-source framework for developing applications that use large language models. Its tools and application programming interfaces (APIs) make it easier to set up some of the most novel uses of NLP and LLMs. We’re talking about chatbots, content summarization, question-answering, and intelligent search — an AI technology that gives you more accurate and relevant online search results.

LangChain supplies your LLM with various data sources and gives it the ability to make decisions on the best way to generate output. Its standardized interface lets you work with any LLM or multiple LLMs at the same time, as well as multiple data sources.

If the app you’re building needs access to current information, access to these multiple data sources is absolutely necessary. Remember — LLMs are trained on data from a set period of time. So if you need current data, like the results of the most recent and highly-watched Copa América final, LangChain gives you the ability to integrate it with an up-to-date search tool.

With Langchain, you can receive data from multiple sources and in multiple formats, making it easy for you to swap between using different LLMs like GPT (OpenAI), Gemini (Google), Cohere, Mistral, and others.

LangChain is a rapidly evolving and extremely versatile tool, and the way you use it just depends on what you’re using it for!

2. LlamaIndex

There’s probably a good reason why companies like T-Mobile and organizations like the University of California, Berkeley use LlamaIndex, right? LlamaIndex used to be known as GPT Index, but even before the name change, it’s been one of the best LLM frameworks for building context-augmented generative AI applications with LLMs.

“Context augmentation” is an important phrase in the LlamaIndex universe. In fact, it’s the phrase. When you use an LLM, it comes pre-trained on large amounts of public information. But what if your data is private or specific to the problem you’re trying to solve? That’s where LlamaIndex comes into play. Whether your data is stuck behind APIs or in SQL databases and PDFs, context augmentation makes your data available to the LLM.

And if you haven’t heard of context augmentation, you might’ve heard of its most popular example — RAG or Retrieval-Augmented Generation. LlamaIndex uses RAG systems that combine LLMs with private data.

Here’s how LlamaIndex breaks down their process. You can use it in four stages:

• Loading: Load in 160+ data sources and formats (APIs, PDFs, documents, SQL databases, etc.) from unstructured, semi-structured, and structured data
• Indexing: Store and structure your data — as vector embeddings that capture its meaning — to make it easy for the LLM to consume the information and create a searchable database specific to the data you loaded
• Querying: Search for the most relevant data — using a RAG pipeline — for question-answering (finding the most relevant information and answering the query)
• Evaluating: Evaluate your app’s performance by measuring retrieval and LLM response quality

Anyone can take advantage of LlamaIndex and its tools. You don’t have to be a seasoned developer. Even if you’re a beginner working on your first app, harnessing LlamaIndex can make your life — and the quality of your LLM application — a lot better.

Is Tech Right For you? Take Our 3-Minute Quiz!

You Will Learn:

☑️ If a career in tech is right for you

☑️ What tech careers fit your strengths

☑️ What skills you need to reach your goals

3. HuggingFace Transformers

HuggingFace Transformers is an open-source library where you can access thousands — over 25,000 to be more exact — of pre-trained transformer models for NLP, computer vision, and audio and speech processing.

We touched on transformer models in our deep dive into LLMs and how they work, but here’s the Sparknotes version. The TL;DR: Transformer models are artificial neural networks used in modern LLMs that make decisions by trying to copy the complex way our brains process information. They deal with data that comes in order — like a sentence — and process entire sequences simultaneously. Transformers can track relationships in a sequence and learn their context and meaning, allowing LLMs to respond to prompts and generate text faster and more efficiently than ever before.

If you had all the time in the world to filter through their transformer models, you’d find models for tons of different tasks. But since you probably don’t (we get it), we’ve got a list!

The Huggingface Transformers library has models for:

• Text generation: Generate new text with LLMs
• Text classification: Classify text, for example, spam or not spam
• Question-answering: Answer questions using context
• Content summarization: Create summaries of large text
• Object detection: Detect objects from an image
• Speech recognition: Transcribe audio to text

Huggingface Transformers also makes it easier to use their models with other machine learning — specifically deep learning — libraries like TensorFlow and Pytorch.

4. TensorFlow

Speaking of TensorFlow…TensorFlow is Google’s open-source framework that AI engineers use to build machine learning applications and develop learning models and algorithms (artificial neural networks). While you could say this is the main objective of TensorFlow, that’s not all it does. The framework also makes it easier to minimize the complexity of implementing computations on large data sets.

Deep learning models require tons (and tons) of computations so they can be as accurate as possible. TensorFlow uses “tensors” — algorithms that are expressed as multidimensional arrays (lines) — as the data structure for computations.

There’s also a feature you can use to create structures for machine learning models. These structures are data flow graphs that denote the functionalities you want to use. Data flow graphs are a set of nodes (in a clearly defined order) where you can specify how the model will compute the information.

The data that gets fed into the model needs to be a multidimensional array. And these arrays are extremely helpful when you have large amounts of data. When looking at a graph, the nodes (circles) represent a mathematical operation while each edge (or connection) between the nodes is the tensor that represents the data flowing between them.

Here’s how you’d use TensorFlow:

• Model building: Use the TensorFlow API to create a computational graph where you can define your model architecture and create your model.
• Data processing: Prepare your data — remove duplicate values, standardize your data, etc. — to feed it to the model.
• Model training: Test algorithms to adjust model parameters based on training data.
• Model evaluation: Evaluate the model’s performance and fine-tune parameters.
• Model deployment: Deploy the model.

To be honest, TensorFlow comes with a pretty steep learning curve, but its value is unbeatable when working with LLMs. It’s a lot easier to use TensorFlow if you have a basic understanding of machine learning, so make sure you check out our beginner’s guide to machine learning.

5. PyTorch

Facebook’s AI Research Lab developed PyTorch, an open-source machine learning library. There’s nothing wrong with a little competition, and in a conversation on the best frameworks for deep learning, TensorFlow and PyTorch will often come up in the same conversation. PyTorch is typically thought to be more beginner-friendly, while TensorFlow might be better suited for programmers looking to deploy scalable, highly optimized, and high-performance models.

PyTorch is based on Python — one of the top programming languages in AI — and Torch library, an open-source library that was developed for numerical computations and machine learning tasks. The PyTorch framework was designed to speed up the process between experimentation, prototyping, and deployment.

To understand exactly how PyTorch works, it’s easier to break it down into its features and components.

• Tensor computation: Tensors — algorithms that are expressed as multidimensional arrays (lines) — are used for computations on large data sets and can be manipulated with APIs.
• Dynamic computational graphs: These graphs are constructed as operations are performed, making it easier to debug and change the model architecture as it runs. It also lets developers run and test portions of the code without waiting for the entire program to be written.
• Accelerated computation: PyTorch supports graphics processing unit (GPU) acceleration, making it faster to implement computations on large models and datasets.

There are a lot of reasons why the PyTorch community is growing and thriving, and it has everything to do with the quality of their offerings. Among them is the PyTorch Hub which features an archive of pre-trained models. And because PyTorch is written in Python and integrates with popular Python libraries like NumPy, it’s a lot easier to learn if you already have experience working with Python.

Is Tech Right For you? Take Our 3-Minute Quiz!

You Will Learn:

☑️ If a career in tech is right for you

☑️ What tech careers fit your strengths

☑️ What skills you need to reach your goals