app.py

import gradio as gr
import faiss
import pandas as pd
from transformers import AutoModelForCausalLM, AutoTokenizer
from sentence_transformers import SentenceTransformer
import numpy as np

# Load your FAISS index
index_path = "faiss_index/index.faiss"  # Update with your FAISS index file path
index = faiss.read_index(index_path)


# Load the metadata
df = pd.read_pickle('df_news (1).pkl')

# Load the Hugging Face model and tokenizer
model = SentenceTransformer("sentence-transformers/all-mpnet-base-v2")
hf_tokenizer = AutoTokenizer.from_pretrained('BAAI/bge-large-zh-v1.5')
hf_model = AutoModel.from_pretrained('BAAI/bge-large-zh-v1.5')

# Define the function for similarity search
def search(query, k=10):
    query_embedding = embedding_model.encode(query).astype('float32')
    D, I = index.search(np.array([query_embedding]), k)
    
    results = []
    for idx in I[0]:
        if idx < len(df):  # Ensure the index is within bounds
            doc = df.iloc[idx]
            results.append({
                'title': doc['title'],
                'author': doc['author'],
                'content': doc['full_text'],
                'source': doc['url']
            })
    return results

# Define the function to generate a response based on the retrieved documents
def generate_answer(query, max_tokens, temperature, top_p):
    # Perform similarity search
    search_results = search(query)
    context = "\n\n".join([f"Title: {doc['title']}\nContent: {doc['content']}" for doc in search_results])
    
    # Construct the prompt
    full_prompt = f"Context:\n{context}\n\nQuestion: {query}"
    
    # Tokenize the input prompt
    inputs = hf_tokenizer(full_prompt, return_tensors="pt")
    
    # Generate a response using the model
    output = hf_model.generate(
        inputs["input_ids"],
        max_length=max_tokens,
        temperature=temperature,
        top_p=top_p,
        pad_token_id=hf_tokenizer.eos_token_id
    )
    
    # Decode the response and return it
    response = hf_tokenizer.decode(output[0], skip_special_tokens=True)
    return response

# Define the Gradio interface
def respond(message, max_tokens, temperature, top_p):
    response = generate_answer(message, max_tokens, temperature, top_p)
    return response

# Set up the Gradio demo
demo = gr.Interface(
    fn=respond,
    inputs=[
        gr.Textbox(value="What is the latest news?", label="Query"),
        gr.Slider(minimum=1, maximum=2048, value=150, step=1, label="Max new tokens"),
        gr.Slider(minimum=0.1, maximum=2.0, value=1.0, step=0.1, label="Temperature"),
        gr.Slider(minimum=0.1, maximum=1.0, value=0.9, step=0.1, label="Top-p (nucleus sampling)")
    ],
    outputs=[gr.Textbox()]
)

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