app.py

# import gradio as gr

# gr.load("models/HuggingFaceH4/zephyr-7b-alpha").launch()

import os
import numpy as np
import gradio as gr
from transformers import AutoModelForCausalLM, AutoTokenizer
import faiss

# Step 1: Load Precomputed Embeddings and Metadata
def load_embeddings(embeddings_folder='embeddings'):
    all_embeddings = []
    metadata = []
    for file in os.listdir(embeddings_folder):
        if file.endswith('.npy'):
            embedding_path = os.path.join(embeddings_folder, file)
            embedding = np.load(embedding_path)  # Shape: (27, 384)
            all_embeddings.append(embedding)
            # Metadata corresponds to each .npy file
            meta_info = file.replace('.npy', '')  # Example: 'course_1'
            metadata.extend([meta_info] * embedding.shape[0])  # Repeat metadata for each sub-embedding

    # Flatten list of embeddings to shape (n * 27, 384)
    all_embeddings = np.vstack(all_embeddings)
    return all_embeddings, metadata

embeddings, metadata = load_embeddings()

# Step 2: Set Up FAISS Index with Flattened Embeddings
dimension = embeddings.shape[1]  # Should be 384 after flattening
index = faiss.IndexFlatL2(dimension)
index.add(embeddings)


# Step 3: Load the Language Model
# model_name = "HuggingFaceH4/zephyr-7b-alpha"
# tokenizer = AutoTokenizer.from_pretrained(model_name)
# model = AutoModelForCausalLM.from_pretrained(model_name)

model_name = "TheBloke/zephyr-7B-beta-GPTQ"
tokenizer = AutoTokenizer.from_pretrained(model_name, use_fast=True)
model = AutoModelForCausalLM.from_pretrained(model_name, device_map="balanced", trust_remote_code=False)


# Step 4: Define the Retrieval Function
def retrieve_documents(query, top_k=3):
    query_embedding = np.mean([embeddings[i] for i in range(len(metadata)) if query.lower() in metadata[i].lower()], axis=0)
    distances, indices = index.search(np.array([query_embedding]), top_k)
    retrieved_docs = [metadata[idx] for idx in indices[0]]
    return retrieved_docs

# Step 5: Define the Response Generation Function
def generate_response(query):
    retrieved_docs = retrieve_documents(query)
    context = " ".join(retrieved_docs)
    input_text = f"Context: {context}\n\nQuestion: {query}\n\nAnswer:"
    inputs = tokenizer(input_text, return_tensors="pt")
    output = model.generate(**inputs, max_length=512)
    answer = tokenizer.decode(output[0], skip_special_tokens=True)
    return answer

# Step 6: Create Gradio Interface
def gradio_interface(query):
    response = generate_response(query)
    return response

iface = gr.Interface(
    fn=gradio_interface,
    inputs=gr.Textbox(lines=2, placeholder="Enter your query here..."),
    outputs="text",
    title="RAG-based Course Search",
    description="Enter a query to search for relevant courses using Retrieval Augmented Generation."
)

if _name_ == "_main_":
    iface.launch()