app.py

import gradio as gr
import torch
from transformers import pipeline, AutoTokenizer, AutoModelForCausalLM

from langchain_huggingface import HuggingFaceEmbeddings, HuggingFacePipeline
from langchain_community.vectorstores import FAISS
from langchain_text_splitters import RecursiveCharacterTextSplitter
from langchain_core.prompts import ChatPromptTemplate
from langchain_core.runnables import RunnablePassthrough
from langchain_core.output_parsers import StrOutputParser


def build_chain():
    embed_model_id = "BAAI/bge-small-en-v1.5"
    embeddings = HuggingFaceEmbeddings(
        model_name=embed_model_id,
        model_kwargs={"device": "cuda" if torch.cuda.is_available() else "cpu"}
    )

    texts = [
        "Kragujevac is a city in central Serbia founded in the 15th century.",
        "The main industry in Kragujevac includes automotive manufacturing.",
        "Famous landmarks: The Šumarice Memorial Park and the Old Foundry Museum."
    ]

    text_splitter = RecursiveCharacterTextSplitter(chunk_size=500, chunk_overlap=80)
    docs = text_splitter.create_documents(texts)

    vectorstore = FAISS.from_documents(docs, embeddings)
    retriever = vectorstore.as_retriever(search_kwargs={"k": 3})

    model_id = "Qwen/Qwen2.5-1.5B-Instruct"
    tokenizer = AutoTokenizer.from_pretrained(model_id)

    # sigurnosno: ako nema pad token
    if tokenizer.pad_token_id is None:
        tokenizer.pad_token = tokenizer.eos_token

    model = AutoModelForCausalLM.from_pretrained(
        model_id,
        device_map="auto",          # menjaš u "auto" ako imaš GPU space
        torch_dtype=torch.float16
    )

    pipe = pipeline(
        "text-generation",
        model=model,
        tokenizer=tokenizer,
        max_new_tokens=200,
        temperature=0.7,
        do_sample=True,
        return_full_text=False
    )

    llm = HuggingFacePipeline(pipeline=pipe)

    template = """You are a helpful assistant. Use only the provided context to answer.
If unsure, say "I don't know."

Context: {context}

Question: {question}

Answer:"""
    prompt = ChatPromptTemplate.from_template(template)

    def format_docs(docs):
        return "\n\n".join(doc.page_content for doc in docs)

    rag_chain = (
        {"context": retriever | format_docs, "question": RunnablePassthrough()}
        | prompt
        | llm
        | StrOutputParser()
    )
    return rag_chain


rag_chain = build_chain()

def answer(question: str):
    if not question.strip():
        return ""
    return rag_chain.invoke(question)


demo = gr.Interface(
    fn=answer,
    inputs=gr.Textbox(lines=2, label="Question"),
    outputs=gr.Textbox(lines=8, label="Answer"),
    title="Mini RAG demo (Kragujevac)"
)

if __name__ == "__main__":
    demo.launch()