app.py

import gradio as gr
from sentence_transformers import SentenceTransformer
import pandas as pd
from datasets import load_dataset
from annoy import AnnoyIndex
import numpy as np

# Load the dataset
dataset = load_dataset("nickprock/AIRC_FAQ")
df = pd.DataFrame(dataset["train"])

# Extract questions and answers
questions = df["question"].tolist()
answers = df["answer"].tolist()

# Sentence-transformers models to test
model_names = [
    "nickprock/multi-sentence-BERTino",
    "nickprock/sentence-bert-base-italian-uncased",
    "nickprock/sentence-bert-base-italian-xxl-uncased",
    "nickprock/mmarco-bert-base-italian-uncased",
]

models = {name: SentenceTransformer(name) for name in model_names}
annoy_indexes = {}  # Store Annoy indexes for each model

def build_annoy_index(model_name):
    """Builds an Annoy index for a given model."""
    model = models[model_name]
    embeddings = model.encode(answers)
    embedding_dim = embeddings.shape[1]
    annoy_index = AnnoyIndex(embedding_dim, "angular")  # Use angular distance for cosine similarity
    for i, embedding in enumerate(embeddings):
        annoy_index.add_item(i, embedding)
    annoy_index.build(10)  # Build with 10 trees
    return annoy_index

# Build Annoy indexes for each model
for model_name in model_names:
    annoy_indexes[model_name] = build_annoy_index(model_name)

def find_similar_answer_annoy(question, model_name):
    """Finds the most similar answer using Annoy."""
    model = models[model_name]
    annoy_index = annoy_indexes[model_name]
    question_embedding = model.encode(question)
    nearest_neighbors = annoy_index.get_nns_by_vector(question_embedding, 1)  # Get the nearest neighbor
    best_answer_index = nearest_neighbors[0]
    return answers[best_answer_index]

def compare_models_annoy(question, model1_name, model2_name, model3_name, model4_name):
    """Compares the results of different models using Annoy."""
    answer1 = find_similar_answer_annoy(question, model1_name)
    answer2 = find_similar_answer_annoy(question, model2_name)
    answer3 = find_similar_answer_annoy(question, model3_name)
    answer4 = find_similar_answer_annoy(question, model4_name)
    return answer1, answer2, answer3, answer4

iface = gr.Interface(
    fn=compare_models_annoy,
    inputs=[
        gr.Textbox(lines=2, placeholder="Enter your question here..."),
        gr.Dropdown(model_names, value=model_names[0], label="Model 1"),
        gr.Dropdown(model_names, value=model_names[1], label="Model 2"),
        gr.Dropdown(model_names, value=model_names[2], label="Model 3"),
        gr.Dropdown(model_names, value=model_names[3], label="Model 4"),
    ],
    outputs=[
        gr.Textbox(label=model_names[0]),
        gr.Textbox(label=model_names[1]),
        gr.Textbox(label=model_names[2]),
        gr.Textbox(label=model_names[3]),
    ],
    title="Sentence Transformer Model Comparison (Annoy)",
    description="Enter a question and compare the answers generated by different sentence-transformer models (using Annoy for faster search).",
)

iface.launch()