app.py

from langchain_community.document_loaders import UnstructuredMarkdownLoader
from langchain.text_splitter import RecursiveCharacterTextSplitter
from langchain_huggingface import HuggingFaceEmbeddings
#from langchain.vectorstores import FAISS
from langchain_community.vectorstores import FAISS
from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig
from langchain_huggingface.llms import HuggingFacePipeline
from langchain.prompts import PromptTemplate

from langchain_core.output_parsers import StrOutputParser
from langchain_core.runnables import RunnablePassthrough
import glob
import gradio as gr
from langchain_huggingface import HuggingFaceEndpoint
from langchain_community.embeddings import HuggingFaceInferenceAPIEmbeddings

import os

# I learned a lot from https://huggingface.co/learn/cookbook/en/rag_zephyr_langchain. If you want to learn RAG, check it out.

secret_value_hf = os.getenv("hf_token")

hf_embeddings = HuggingFaceInferenceAPIEmbeddings(
    api_key=secret_value_hf,
    model_name="sentence-transformers/all-MiniLM-l6-v2"
)

md_path = glob.glob( "md_files/*.md")

docs = [UnstructuredMarkdownLoader(md).load() for md in md_path]
docs_list = [item for sublist in docs for item in sublist]

# Split documents
text_splitter = RecursiveCharacterTextSplitter.from_tiktoken_encoder(
    chunk_size=1000, chunk_overlap=200
)
doc_splits = text_splitter.split_documents(docs_list)


# Create the embeddings + retriever

db = FAISS.from_documents(doc_splits,
                          hf_embeddings)


# prompt
prompt_template = '''You are an assistant for question-answering tasks. 
        Here is the context to use to answer the question:
        {context} 
        Think carefully about the above context. 
        Now, review the user question:
        {question}
        Provide an answer to this questions using only the above context. 
        Use three sentences maximum and keep the answer concise.
        Answer:'''

prompt = PromptTemplate(
    input_variables=["context", "question"],
    template=prompt_template,
)


# gradio interface

def get_output(model_name:str,is_RAG:str,questions:str):
    if model_name=="mistralai/Mistral-7B-Instruct-v0.2":
        #repo_id = "mistralai/Mistral-7B-Instruct-v0.2"
        llm = HuggingFaceEndpoint(
            repo_id=model_name,
            max_length=4096,
            temperature=0.2,
            huggingfacehub_api_token=secret_value_hf,
        )
        llm_chain = prompt | llm | StrOutputParser()
        retriever = db.as_retriever(
            search_type="similarity",
            search_kwargs={'k': 4}
        )
        
        rag_chain = (
         {"context": retriever, "question": RunnablePassthrough()}
            | llm_chain
        )
        if is_RAG== "RAG":
            generation2=rag_chain.invoke(questions)
            return generation2
        else:
            generation1=llm_chain.invoke({"context":"", "question": questions})
            return generation1
    elif model_name=="meta-llama/Llama-3.2-3B-Instruct":
        llm = HuggingFaceEndpoint(
            repo_id=model_name,
            max_length=4096,
            temperature=0.2,
            huggingfacehub_api_token=secret_value_hf,
        )
        llm_chain = prompt | llm | StrOutputParser()
        retriever = db.as_retriever()
        
        rag_chain = (
         {"context": retriever, "question": RunnablePassthrough()}
            | llm_chain
        )
        if is_RAG== "RAG":
            generation2=rag_chain.invoke(questions)
            return generation2
        else:
            generation1=llm_chain.invoke({"context":"", "question": questions})
            return generation1
    elif model_name=="Qwen/Qwen2.5-72B-Instruct":
        llm = HuggingFaceEndpoint(
            repo_id=model_name,
            max_length=4096,
            temperature=0.2,
            huggingfacehub_api_token=secret_value_hf,
        )
        llm_chain = prompt | llm | StrOutputParser()
        retriever = db.as_retriever()
        
        rag_chain = (
         {"context": retriever, "question": RunnablePassthrough()}
            | llm_chain
        )
        if is_RAG== "RAG":
            generation2=rag_chain.invoke(questions)
            return generation2
        else:
            generation1=llm_chain.invoke({"context":"", "question": questions})
            return generation1



# Custom CSS to style the output area
custom_css = """
#output_area {
    background-color: #1e1e1e; /* Dark background */
    color: #ffffff; /* White text */
    padding: 10px;
    border-radius: 5px;
    border: 1px solid #333333; /* Dark border */
    margin-top: 10px;
}

#output_area h3 {
    color: #ffcc00; /* Yellow title color */
    margin-bottom: 10px;
}
"""

with gr.Blocks(title="Ask Questions on Chalcogenide Perovskites",theme=gr.themes.Ocean(),css=custom_css) as demo:
    gr.Markdown("""
    ## Retrieval-Augmented Generation for Chalcogenide Perovskites
    This space implements Retrieval-Augmented Generation (RAG) using large language models, based on Hui Haolei's work on chalcogenide perovskite papers. You can select different models and choose whether to use RAG to enhance the responses.
    For more details, check my [github](https://github.com/HaoleiH/AI-driven-projects/blob/main/RAG-using-Llama3.2-3b/README_RAG.md).
    """)
    
    with gr.Row():
        model_name = gr.Radio(
            choices=["mistralai/Mistral-7B-Instruct-v0.2", "meta-llama/Llama-3.2-3B-Instruct", "Qwen/Qwen2.5-72B-Instruct"],
            value="mistralai/Mistral-7B-Instruct-v0.2",
            label="Model Name",
            info="Select the model you want to use."
        )
    
    with gr.Row():
        rag = gr.Radio(
            choices=["RAG", "No RAG"],
            value="RAG",
            label="RAG or Not",
            info="Choose whether to use Retrieval-Augmented Generation."
        )
    
    with gr.Row():
        question = gr.Textbox(
            label="Input Question",
            placeholder="Enter your question about chalcogenide perovskites here...",
            lines=2  # Increase the number of lines for better input experience
        )
    
    with gr.Row():
        submit_button = gr.Button("Submit")
    
    with gr.Row():
        output = gr.Textbox(label="Response",
                            lines=10,  # Increase the number of lines for the output area
                            elem_id="output_area"  # Assign a custom ID for styling
                            )
    submit_button.click(
        fn=get_output,
        inputs=[model_name, rag, question],
        outputs=output
    )

    gr.Examples(
        examples=[
            ["mistralai/Mistral-7B-Instruct-v0.2", "RAG", "What is the advantage of BaZrS3?"],
            ["mistralai/Mistral-7B-Instruct-v0.2", "RAG", "What is the bandgap of SrHfS3?"],
            ["mistralai/Mistral-7B-Instruct-v0.2", "RAG", "Why is chalcogenide perovskite important?"]
        ],
        fn=get_output,
        inputs=[model_name, rag, question],
        outputs=output,
        cache_examples=False
    )

demo.launch()