Quickstart Tutorial:

---

LangTorch is a Python package that accelerates development of complex language model applications by leveraging familiar PyTorch concepts.

While existing frameworks focus on connecting language models to other services, LangTorch aims to change the way you approach creating LLM applications by introducing a unified framework for working with texts, chats, templates, LLMs, API calls and more.

Powered by TextTensors — "torch Tensors but with text data entries" — offering a flexible way to structure and transform text data and embeddings with seamless parallelization.

Installation

pip install langtorch

Overview

• No useless classes:

  Instead of providing wrapper classes for users to memorize, LangTorch introduces fewer, more flexible objects that enable all kinds of text formatting, templating and LLM operations.

• Unified Approach:

  TextTensors let you structure geometrically and handle in parallel text entries that can represent:

  strings, documents, prompt templates, completion dictionaries, chat histories, markup languages, chunks, retrieval queries, tokens, embeddings and so on

• You probably already know LangTorch:

  LangTorch components subclass their numerical PyTorch counterparts, which lets users apply their existing coding skills to building novel LLM app architectures.

• Other goodies like TextModules

  a subclass of torch.nn.Module working on TextTensors and able to perform:

  template completions, prompt injections, local and API LLM inference, create embedding, performing operations on embeddings in retrieval and so on, and so on

• Honestly just go to https://langtorch.org there is much more information there!

Code Examples

The examples are introduced on the main documentation page, but even without much introduction you can see how compact some pretty complex operations can be implemented with LangTorch.

TextTensors act both as texts and embeddings

import torch  
  
tensor1 = TextTensor([["Yes"], ["No"]])  
tensor2 = TextTensor(["Yeah", "Nope", "Yup", "Non"])  
  
print(torch.cosine_similarity(tensor1, tensor2))
print("Content:\n", tensor1)

tensor([[0.6923, 0.6644, 0.6317, 0.5749],
	    [0.5457, 0.7728, 0.5387, 0.7036]])
Content:
[[Yes], 
 [No ]]

LangTorch code looks weird at first, why? Since the utility of Tensors, as used in Torch, relies on their ability to calculate simultaneously products of several weights. The corresponding, and most used, feature in LangTorch allows several prompts to be formatted on several inputs, by defining the multiplication of text entries text1*text2 similarly to text1.format(**text2)

Chains

The multiplication operation lets us build chains of TextModules with a simple torch.nn.Sequential:

chain = torch.nn.Sequential(
    TextModule("Translate this equation to natural language: {}"),
    CoT,
    OpenAI("gpt-4")
    TextModule("Calculate the described quantity: {}"),
    OpenAI("gpt-4", T=0)
)

input_tensor = TextTensor(["170*32 =", "4*20 =", "123*45/10 =", "2**10*5 ="])
output_tensor = chain(input_tensor)

Retrieval & RAG from scratch

The code below is a complete working implementation of a cosine similarity-based retriever:

class Retriever(TextModule):  
    def __init__(self, documents: TextTensor):  
        super().__init__()  
        self.documents = TextTensor(documents).view(-1)  
  
    def forward(self, query: TextTensor, k: int = 5):  
        cos_sim = torch.cosine_similarity(self.documents, query.reshape(1))  
        return self.documents[cos_sim.topk(k)]
         

retriever = Retriever(open("doc.txt", "r").readlines())
query = TextTensor("How to build a retriever?")

print(retriever(query))

We can now compose this module with a TextModule making LLM calls to get a custom Retrieval Augmented Generation pipeline:

class RAG(TextModule):  
    def __init__(self, documents: TextTensor, *args, **kwargs):  
        super().__init__(*args, **kwargs)  
        self.retriever = Retriever(documents)  
  
    def forward(self, user_message: TextTensor, k: int = 5):  
        retrieved_context = self.retriever(user_message, k) +"\n"  
        user_message = user_message + "\nCONTEXT:\n" + retrieved_context.sum()  
        return super().forward(user_message)

rag_chat = RAG(paragraphs,  
			   prompt="Use the context to answer the following user query: ",
			   activation="gpt-3.5-turbo")

assistant_response = rag_chat(user_query)

Go to https://langtorch.org to understand these RAGs-to-riches code shenanigans.

License

LangTorch is available under the MIT license.