app.py

import os
from pathlib import Path
import gradio as gr

from langchain_community.vectorstores import FAISS
from langchain_huggingface import HuggingFaceEmbeddings
from langchain_core.prompts import PromptTemplate
from langchain.chains import RetrievalQA
from langchain.llms import HuggingFacePipeline
from transformers import pipeline

# Constants
DB_FAISS_PATH = "vectorstore/db_faiss"  # Pre-generated FAISS directory
EMBEDDING_MODEL_NAME = "sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2"
MODEL_NAME = "MBZUAI/LaMini-Flan-T5-783M"  # Lightweight CPU-friendly model

# Step 1: Load FAISS vectorstore (already created offline)
def load_vector_store():
    embeddings = HuggingFaceEmbeddings(model_name=EMBEDDING_MODEL_NAME)
    return FAISS.load_local(DB_FAISS_PATH, embeddings, allow_dangerous_deserialization=True)

# Step 2: Load lightweight HuggingFace model (no token needed)
def load_llm():
    pipe = pipeline("text2text-generation", model=MODEL_NAME)
    return HuggingFacePipeline(pipeline=pipe)

# Step 3: Setup QA chain
def setup_chain():
    prompt_template = """
    Use the following context to answer the question.
    If the answer is not in the context, just say you don't know.

    Context: {context}
    Question: {question}
    """
    prompt = PromptTemplate(template=prompt_template, input_variables=["context", "question"])
    retriever = load_vector_store().as_retriever(search_kwargs={"k": 3})
    llm = load_llm()
    return RetrievalQA.from_chain_type(
        llm=llm,
        chain_type="stuff",
        retriever=retriever,
        return_source_documents=True,
        chain_type_kwargs={"prompt": prompt}
    )

qa_chain = setup_chain()

# Step 4: Gradio Interface
def rag_bot(query):
    result = qa_chain.invoke({"query": query})
    return result["result"]

# Step 5: Launch Interface
demo = gr.Interface(
    fn=rag_bot,
    inputs="text",
    outputs="text",
    title="TextileVision: AI Chatbot",
    description="Ask queries about loom speed, yarn mixing, knitting prediction, and textile operations."
)
demo.launch()