app.py

import gradio as gr
from gradio.components import Textbox
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM, T5ForConditionalGeneration
from peft import PeftModel
import torch
import datasets
from sentence_transformers import SentenceTransformer, util
import math
import re
from nltk import sent_tokenize, word_tokenize
import nltk
nltk.download('punkt')

# Load bi encoder
bi_encoder = SentenceTransformer('legacy107/multi-qa-mpnet-base-dot-v1-wikipedia-search')
bi_encoder.max_seq_length = 256
top_k = 3

# Load your fine-tuned model and tokenizer
model_name = "legacy107/flan-t5-large-ia3-wiki-merged"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = T5ForConditionalGeneration.from_pretrained(model_name)
max_length = 512
max_target_length = 200

# Load your dataset
dataset = datasets.load_dataset("legacy107/qa_wikipedia_retrieved_chunks", split="test")
dataset = dataset.shuffle()
dataset = dataset.select(range(10))

# Context chunking
def chunk_splitter(context, chunk_size=100, overlap=0.20):
    overlap_size = chunk_size * overlap
    sentences = nltk.sent_tokenize(context)

    chunks = []
    text = sentences[0]

    if len(sentences) == 1:
        chunks.append(text)

    i = 1
    while i < len(sentences):
        text += " " + sentences[i]
        i += 1
        while i < len(sentences) and len(nltk.word_tokenize(f"{text} {sentences[i]}")) <= chunk_size:
            text += " " + sentences[i]
            i += 1

        text = text.replace('\"','"').replace("\'","'").replace('\n\n\n'," ").replace('\n\n'," ").replace('\n'," ")
        chunks.append(text)

        if (i >= len(sentences)):
            break

        j = i - 1
        text = sentences[j]
        while j >= 0 and len(nltk.word_tokenize(f"{sentences[j]} {text}")) <= overlap_size:
            text = sentences[j] + " " + text
            j -= 1

    return chunks


def retrieve_context(query, contexts):
    corpus_embeddings = bi_encoder.encode(contexts, convert_to_tensor=True, show_progress_bar=False)

    question_embedding = bi_encoder.encode(query, convert_to_tensor=True, show_progress_bar=False)
    question_embedding = question_embedding
    hits = util.semantic_search(question_embedding, corpus_embeddings, top_k=top_k)
    hits = hits[0]
    
    hits = sorted(hits, key=lambda x: x['score'], reverse=True)
    return " ".join([contexts[hit['corpus_id']] for hit in hits[0:top_k]]).replace("\n", " ")


# Define your function to generate answers
def generate_answer(question, context, ground):
    contexts = chunk_splitter(context)
    context = retrieve_context(question, contexts)
    
    # Combine question and context
    input_text = f"question: {question} context: {context}"

    # Tokenize the input text
    input_ids = tokenizer(
        input_text,
        return_tensors="pt",
        padding="max_length",
        truncation=True,
        max_length=max_length,
    ).input_ids

    # Generate the answer
    with torch.no_grad():
        generated_ids = model.generate(input_ids=input_ids, max_new_tokens=max_target_length)

    # Decode and return the generated answer
    generated_answer = tokenizer.decode(generated_ids[0], skip_special_tokens=True)

    return generated_answer, context, ground


# Define a function to list examples from the dataset
def list_examples():
    examples = []
    for example in dataset:
        context = example["article"]
        question = example["question"]
        answer = example["answer"]
        examples.append([question, context, answer])
    return examples


# Create a Gradio interface
iface = gr.Interface(
    fn=generate_answer,
    inputs=[
        Textbox(label="Question"),
        Textbox(label="Context"),
        Textbox(label="Ground truth")
    ],
    outputs=[
        Textbox(label="Generated Answer"),
        Textbox(label="Retrieved Context"),
        Textbox(label="Ground Truth")
    ],
    examples=list_examples()
)

# Launch the Gradio interface
iface.launch()