app.py

import nltk
nltk.download("punkt")
import gradio as gr
from transformers import AutoTokenizer, AutoModelForCausalLM, AutoModelForSeq2SeqLM
from llama_index.core import VectorStoreIndex, SimpleDirectoryReader, ServiceContext
from llama_index.llms import HuggingFaceLLM
from llama_index.embeddings.huggingface import HuggingFaceEmbedding
import os

# Load DeepSeek-R1
deepseek_tokenizer = AutoTokenizer.from_pretrained("deepseek-ai/DeepSeek-R1", trust_remote_code=True)
deepseek_model = AutoModelForCausalLM.from_pretrained("deepseek-ai/DeepSeek-R1", trust_remote_code=True)

# Load IndicBART
indicbart_tokenizer = AutoTokenizer.from_pretrained("ai4bharat/IndicBART")
indicbart_model = AutoModelForSeq2SeqLM.from_pretrained("ai4bharat/IndicBART")

# Initialize LlamaIndex components
embed_model = HuggingFaceEmbedding(model_name="sentence-transformers/all-MiniLM-L6-v2")

llm = HuggingFaceLLM(
    model_name="deepseek-ai/DeepSeek-R1",
    tokenizer_name="deepseek-ai/DeepSeek-R1",
    max_new_tokens=512,
    context_window=4096
)

service_context = ServiceContext.from_defaults(llm=llm, embed_model=embed_model)

# Build index from documents in 'data' directory
documents = SimpleDirectoryReader("data").load_data()
index = VectorStoreIndex.from_documents(documents, service_context=service_context)

# Define functions for each task
def restore_text(input_text, task_type):
    prefix_map = {
        "Restore & Correct Tamil Text": "restore: ",
        "Summarize in Tamil": "summarize: ",
        "Translate to English": "translate Tamil to English: "
    }
    prefix = prefix_map.get(task_type, "restore: ")
    input_text = prefix + input_text
    inputs = indicbart_tokenizer([input_text], return_tensors="pt", padding=True)
    outputs = indicbart_model.generate(**inputs, max_length=256, num_beams=4, early_stopping=True)
    decoded_output = indicbart_tokenizer.batch_decode(outputs, skip_special_tokens=True)
    return decoded_output[0]

def deepseek_chat(message):
    inputs = deepseek_tokenizer.encode(message + deepseek_tokenizer.eos_token, return_tensors="pt")
    outputs = deepseek_model.generate(inputs, max_length=1024, pad_token_id=deepseek_tokenizer.eos_token_id)
    return deepseek_tokenizer.decode(outputs[0], skip_special_tokens=True)

def query_documents(query):
    query_engine = index.as_query_engine()
    response = query_engine.query(query)
    return str(response)

# Gradio Interface
with gr.Blocks() as demo:
    gr.Markdown("## 🕉️ Ancient Tamil Literature Expert AI")

    with gr.Tab("IndicBART Tasks"):
        input_text = gr.Textbox(label="Input Tamil Text", lines=8, placeholder="Enter ancient Tamil text here...")
        task_type = gr.Radio(choices=["Restore & Correct Tamil Text", "Summarize in Tamil", "Translate to English"], label="Select Task")
        output_text = gr.Textbox(label="Output")
        submit_button = gr.Button("Submit")
        submit_button.click(fn=restore_text, inputs=[input_text, task_type], outputs=output_text)

    with gr.Tab("DeepSeek-R1 Chat"):
        chat_input = gr.Textbox(label="Enter your message")
        chat_output = gr.Textbox(label="DeepSeek-R1 Response")
        chat_button = gr.Button("Send")
        chat_button.click(fn=deepseek_chat, inputs=chat_input, outputs=chat_output)

    with gr.Tab("Document Query"):
        query_input = gr.Textbox(label="Enter your query")
        query_output = gr.Textbox(label="Query Response")
        query_button = gr.Button("Search")
        query_button.click(fn=query_documents, inputs=query_input, outputs=query_output)

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