Generative-AI-Using-Langchain-and-Huggingface

I have created a Text Generation Chat bot using the Huggingface pretrained models and combine them with Langchain to create a workflow.

For API Calls I Installed the required libraries from HuggingFace

1. huggingface_hub
2. transformers
3. accelerate
4. bitsandbytes
5. Langchain

I have acced the pretrained model using different methods like:

1. Through Hugging Face Endpoints
2. By creating a LLMChain
3. By creating a LLMpipeLine

HuggingFaceEndpoint

Used Hugging face Endpoints to access the pre trained models present in Hugging Face

 from langchain_huggingface import HuggingFaceEndpoint
 
 llm = HuggingFaceEndpoint(
 repo_id="mistralai/Mistral-7B-Instruct-v0.3",
 task="text-generation",
 max_new_tokens=100,
 do_sample=False, ) 

Model Used - mistralai/Mistral-7B-Instruct-v0.3

MistralAI has around 7 Billion parameters so it is not possible to host or download it locally as it is a very large Model hence we have used the API Method to call and access these models. For Text Generation.

LLMChain:

 from langchain import PromptTemplate, LLMChain 

LLMChain will Wraps an LLM and a prompt template. It will help in building logic, memory, and multi-step reasoning. from prompt template we have used LLMChain to create a working flow.

LLM Pipe Line:

 from langchain_huggingface import HuggingFacePipeline 

Also created a LLM pipeline by downloading the simple model gpt-2

AutoModelForCausallLM will download the model weights and configuration needed for text generation tasks

 from transformers import AutoModelForCausalLM 

Tokenizer - Tokenizer converts input text into tokens (numbers) that models can understand, and vice versa.

 from transformers import AutoTokenizer 

This will create a ready to use pipeline object and ready to use generic pipeline utility from transformers

 from transformers import pipeline 

Models can be loaded directly with the from_model_id method

The tokenizer is necessary to encode input text into tokens and decode output tokens back to text.

 tokenizer = AutoTokenizer.from_pretrained(model_id) 

As we have to download it locally so we use a lighter model like GPT2

The below pipeline will handle encoding inputs, running the model, and decoding outputs automatically

 pipe = pipeline(
    "text-generation",
    model=model,
    tokenizer=tokenizer,
    max_new_tokens=100
)