app.py

import json
import numpy as np
from wordllama import WordLlama
import gradio as gr
from numpy.linalg import norm
import os

# Load episodes data
with open('processed_episodes.json', 'r') as f:
    episodes = json.load(f)

# Load the WordLlama model with 256 dimensions
wl = WordLlama.load(trunc_dim=256)

# Check if embeddings are already cached
if os.path.exists('summary_embeddings.npy'):
    # Load embeddings from cache
    summary_embeddings = np.load('summary_embeddings.npy')
else:
    # Compute embeddings for all summaries
    summaries = [episode['summary'] for episode in episodes]
    summary_embeddings = wl.embed(summaries)
    summary_embeddings = np.array(summary_embeddings)
    # Save embeddings to cache
    np.save('summary_embeddings.npy', summary_embeddings)

# Define the function to find matching episodes
def find_matching_episodes(query, top_k=5):
    # Compute the embedding for the query
    query_embedding = wl.embed([query])[0]  # The result is a 1D numpy array

    # Normalize embeddings
    query_norm = query_embedding / (norm(query_embedding) + 1e-10)
    summaries_norm = summary_embeddings / (norm(summary_embeddings, axis=1, keepdims=True) + 1e-10)

    # Compute cosine similarities
    similarities = summaries_norm @ query_norm

    # Get indices of the top_k most similar summaries
    top_k_indices = np.argsort(similarities)[-top_k:][::-1]

    # Retrieve the matching episodes and their similarity scores
    matching_episodes = []
    for idx in top_k_indices:
        episode = episodes[idx]
        similarity_score = similarities[idx]
        # Create a list of values instead of a dictionary
        result = [
            episode['episode_number'],
            episode['title'],
            f"{similarity_score:.4f}"
        ]
        matching_episodes.append(result)

    return matching_episodes

# Create the Gradio interface
# Create the Gradio interface
interface = gr.Interface(
    fn=find_matching_episodes,
    inputs=[
        gr.Textbox(lines=2, placeholder="Enter your query here...", label="Search Query"),
        gr.Slider(minimum=1, maximum=10, value=5, label="Number of Results")
    ],
    outputs=gr.Dataframe(
        headers=["Episode Number", "Title", "Similarity Score"],
        label="Matching Episodes"
    ),
    title="Picardle",
    description="Enter a query to find matching ST:TNG episodes based on their summaries."
)

# Launch the app
if __name__ == "__main__":
    interface.launch(share=True)