We want to cook by loading recipe data, right? You can easily create a bot using LLM. This time, we will make cooking easier by loading recipes from ABC Cooking Studio, a shareholder of our company. Eggplant and Ground Meat Bolognese Rice Link to ABC Cooking Studio&#39;s recipe<a href="https://www.abc-cooking.co.jp/plus/recipe/detail/?id=44750" rel="noopener noreferrer" target="_blank">https://www.abc-cooking.co.jp/plus/recipe/detail/?id=44750</a> Even if you don&#39;t know what to focus on just by looking at the website, you can have a conversation with AI like below, and it will tell you only the information you want from the website. <img src="https://assets.blueqat.com/public/uploads/us-east-2:4805ff4b-c3cc-4344-b165-86544c34d0bf/2024/04/11/Gradio.png"/> The necessary technologies are: - Langchain (to create a general AI workflow)- LLM (this can be a service like ChatGPT, or it can be hosted on your own local PC)- Web crawler (this time, we will directly retrieve information from websites)- Gradio (you can easily create web interfaces) First, we have already checked how to create basic responses with LLM in another article. <a href="https://blueqat.com/yuichiro_minato2/128b8855-7689-4ce1-8a4b-f10da27b8aa2" rel="noopener noreferrer" target="_blank">https://blueqat.com/yuichiro_minato2/128b8855-7689-4ce1-8a4b-f10da27b8aa2</a> First, for this project, we loaded the model from Hugging Face. It&#39;s Mistral7B. <pre class="ql-syntax" spellcheck="false">pip install transformers accelerate langchain langchain-community sentence-transformers faiss-gpu pypdf
</pre> <pre class="ql-syntax" spellcheck="false">from transformers import AutoTokenizer, pipeline

model_id = &#34;mistralai/Mistral-7B-Instruct-v0.2&#34;
tokenizer = AutoTokenizer.from_pretrained(model_id)
pipe = pipeline(&#34;text-generation&#34;, model=model_id, tokenizer=tokenizer, device=0, max_new_tokens=300)
</pre> Next, although we used PDFLoader last time, we will change this part for the current project. <pre class="ql-syntax" spellcheck="false">from langchain_community.document_loaders import PyPDFLoader
from langchain_community.vectorstores import FAISS
from langchain_community.embeddings.huggingface import HuggingFaceEmbeddings
from langchain_text_splitters import CharacterTextSplitter

loader = PyPDFLoader(&#34;https://arxiv.org/pdf/2312.03982.pdf&#34;)
documents = loader.load()
text_splitter = CharacterTextSplitter(chunk_size=1000, chunk_overlap=0)
docs = text_splitter.split_documents(documents)

embeddings = HuggingFaceEmbeddings(
  model_name=&#34;intfloat/multilingual-e5-large&#34;
)

db = FAISS.from_documents(docs, embeddings)
</pre> Let&#39;s change it into web based loader <pre class="ql-syntax" spellcheck="false">from langchain_community.document_loaders import WebBaseLoader

loader = WebBaseLoader(&#34;https://www.abc-cooking.co.jp/plus/recipe/detail/?id=44750&#34;)
</pre> It&#39;s incredibly easy, isn&#39;t it? With this, the setup for the Loader is complete. Finally, regarding the chat part, including the prompts in the responses, <pre class="ql-syntax" spellcheck="false">from langchain_core.prompts import ChatPromptTemplate
from langchain_core.output_parsers import StrOutputParser
from langchain_core.runnables import RunnablePassthrough
from langchain.llms import HuggingFacePipeline

llm = HuggingFacePipeline(pipeline=pipe)

template = template = &#34;&#34;&#34;Answer the question in English based only on the following context. If there is Japanese language please translate it into English:

{context}

Question: {question}
&#34;&#34;&#34;

prompt = ChatPromptTemplate.from_template(template)

def format_docs(docs):
  return &#34;\n\n&#34;.join([d.page_content for d in docs])

chain = (
  {&#34;context&#34;: retriever | format_docs, &#34;question&#34;: RunnablePassthrough()}
  | prompt
  | llm
  | StrOutputParser()
)

answer = chain.invoke(query)
print(answer)
</pre> This was the basis. Next we will make web interface. To make the web interface we adopt gradio.First, install Gradio. <pre class="ql-syntax" spellcheck="false">!pip install gradio
</pre> And by writing Python as shown below, a web server will start up.It launches locally, but if you set parameters inside `.launch()`, it can also generate a web address for external access.The key point is to replace the response part in the middle with the RAG constructed with the current LLM, and that&#39;s it. <pre class="ql-syntax" spellcheck="false">import gradio as gr
import os

def add_text(history, text):
 history = history + [(text, None)]
 return history, gr.Textbox(value=&#34;&#34;, interactive=False)

def bot(history):
 query = history[-1][0]
 response = chain.invoke(query)
 history[-1][1] = &#34;&#34;
 for character in response:
 history[-1][1] += character
 yield history

with gr.Blocks() as demo:
 chatbot = gr.Chatbot([])
 with gr.Row():
 txt = gr.Textbox(
 scale=4,
 show_label = False,
 container = False
 )
 clear = gr.Button(&#34;Clear&#34;) 

 txt_msg = txt.submit(add_text, [chatbot, txt], [chatbot, txt], queue = False).then(bot, chatbot, chatbot)
 txt_msg.then(lambda: gr.Textbox(interactive = True), None, [txt], queue = False)
 clear.click(lambda: None, None, chatbot, queue=False)

demo.queue() 
demo.launch()
</pre> We quickly created a RAG with an easy-to-use interface!<img src="https://assets.blueqat.com/public/uploads/us-east-2:4805ff4b-c3cc-4344-b165-86544c34d0bf/2024/04/11/Gradio.png"/>

Easy-to-Make Cooking Recipe Robot with Langchain + Mistral7B + Gradio!

Yuichiro Minato