app.py

import gradio as gr
import pandas as pd
from transformers import AutoTokenizer, AutoModelForCausalLM, pipeline
from sentence_transformers import SentenceTransformer
import faiss
import torch

# -------------------------------
# Load dataset
# -------------------------------
file_path = "marketing-campaigns.csv"
df = pd.read_csv(file_path)

# Combine text for embeddings
df = df.dropna(subset=["campaign_name", "description"])
df["text"] = df["campaign_name"] + ": " + df["description"]

# -------------------------------
# Embeddings + FAISS index
# -------------------------------
embed_model = SentenceTransformer("all-MiniLM-L6-v2")

embeddings = embed_model.encode(df["text"].tolist(), convert_to_tensor=True, show_progress_bar=True)
embeddings_np = embeddings.detach().cpu().numpy()

# Build FAISS index
d = embeddings_np.shape[1]
index = faiss.IndexFlatL2(d)
index.add(embeddings_np)

# -------------------------------
# Load LLM (Phi-4-mini)
# -------------------------------
model_name = "microsoft/phi-4-mini"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype=torch.float32, device_map="auto")

# -------------------------------
# RAG functions
# -------------------------------
def retrieve_context(query, k=3):
    query_vec = embed_model.encode([query], convert_to_tensor=True).cpu().numpy()
    D, I = index.search(query_vec, k)
    results = [df.iloc[i]["text"] for i in I[0]]
    return results

def generate_with_rag(prompt):
    # Step 1: Retrieve top campaigns
    context = retrieve_context(prompt, k=3)
    context_str = "\n".join(context)

    # Step 2: Build final prompt
    rag_prompt = f"""
    You are an AI marketing assistant. 
    Here are some past campaigns for reference:\n{context_str}\n
    Based on these, generate a new campaign idea for: {prompt}
    """

    # Step 3: Generate with Phi-4
    inputs = tokenizer(rag_prompt, return_tensors="pt").to(model.device)
    outputs = model.generate(**inputs, max_length=200, temperature=0.7, top_p=0.9)
    return tokenizer.decode(outputs[0], skip_special_tokens=True)

# -------------------------------
# Gradio UI
# -------------------------------
with gr.Blocks() as demo:
    gr.Markdown("## 🤖 RAG-powered AI Marketing Campaign Generator")

    with gr.Row():
        query = gr.Textbox(label="Enter campaign idea or analysis query")
        output = gr.Textbox(label="Generated Campaign")
        btn = gr.Button("Generate with RAG")

    btn.click(generate_with_rag, inputs=query, outputs=output)

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