import pandas as pd import torch from sentence_transformers import SentenceTransformer, util from transformers import BartForConditionalGeneration, BartTokenizer import gradio as gr device = torch.device("cuda" if torch.cuda.is_available() else "cpu") sbert_model = SentenceTransformer('all-mpnet-base-v2') summarization_model = BartForConditionalGeneration.from_pretrained("facebook/bart-base") sbert_model.to(device) summarization_model.to(device) summarization_tokenizer = BartTokenizer.from_pretrained("facebook/bart-base") story_embeddings = torch.load('story_embeddings.pt', map_location=device, weights_only=True) dataset = pd.read_csv('story_data.csv') def summarize_text(text): inputs = summarization_tokenizer.encode("summarize: " + text, return_tensors="pt", max_length=512, truncation=True).to(device) summary_ids = summarization_model.generate(inputs, max_length=100, min_length=30, length_penalty=2.0, num_beams=2, early_stopping=True) return summarization_tokenizer.decode(summary_ids[0], skip_special_tokens=True) def retrieve_and_summarize_stories(query, top_n=5, similarity_threshold=0.5): query_embedding = sbert_model.encode(query, convert_to_tensor=True, device=device) similarities = util.pytorch_cos_sim(query_embedding, story_embeddings)[0] top_scores, top_indices = torch.topk(similarities, k=top_n) results = [] for score, index in zip(top_scores, top_indices): if score.item() < similarity_threshold: break story = dataset.iloc[index.item()] summary = summarize_text(story['CONTENTS']) formatted_summary = ( f"summary: {summary}\n" f"Published on: {story['SOURCE']}\n" f"Read more:{story['NEWS LINK']}\n" ) results.append(formatted_summary) return "\n\n".join(results) def interface(query, threshold): return retrieve_and_summarize_stories(query, similarity_threshold=threshold) threshold_options = [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9] threshold_explanation = ( "Select a similarity threshold to filter results:\n" "- **Lower values** (e.g., 0.1) will return more results, including those that are less similar to your query.\n" "- **Higher values** (e.g., 0.9) will return fewer results, focusing on those that are highly similar to your query." ) instructions = ( "How to use ChatJourno:\n" "1. Enter a query in the text box.\n" "2. Select a similarity threshold from the dropdown menu.\n" "3. Click the 'Submit' button to retrieve and summarize similar news stories." ) iface = gr.Interface( fn=interface, inputs=[ gr.components.Textbox(label="Query", lines=2, placeholder="Enter your query here..."), gr.components.Dropdown( choices=threshold_options, label="Similarity Threshold", value=0.5, type="value" ), ], outputs="json", title="ChatJourno", description=instructions + "\n\n" + threshold_explanation ) iface.launch()