app.py

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

# ==========================================
# 1. Load Dataset
# ==========================================
dataset = load_dataset(
    "electricsheepafrica/nigerian_retail_and_ecommerce_product_catalog_data",
    split="train"
)
df = pd.DataFrame(dataset)

# ==========================================
# 2. Clean Data
# ==========================================
for col in ["product_name", "description", "category", "sub_category", "price"]:
    if col not in df.columns:
        df[col] = ""
    else:
        df[col] = df[col].fillna("").astype(str).str.strip()

# Merging texts with categories
df["text"] = df["product_name"] + " " + df["description"] + " " + df["category"] + " " + df["sub_category"]
df = df.drop_duplicates(subset=["text"])
df = df.head(2000)  # To reduce memory consumption on Space

# ==========================================
# 3. Create  Embeddings + FAISS Index
# ==========================================
embedding_model = SentenceTransformer("all-mpnet-base-v2")  # embeddings
embeddings = embedding_model.encode(df["text"].tolist(), convert_to_numpy=True, show_progress_bar=True)

dimension = embeddings.shape[1]
index = faiss.IndexHNSWFlat(dimension, 32)
index.add(embeddings)

# ==========================================
# 4. Load LLM
# ==========================================
MODEL_ID = "microsoft/phi-2"
tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
model = AutoModelForCausalLM.from_pretrained(
    MODEL_ID,
    device_map="auto",
    torch_dtype=torch.float32
)

# ==========================================
# 5. Retrieval Function
# ==========================================
def retrieve_products(query, top_k=3):
    query_emb = embedding_model.encode([query])
    D, I = index.search(np.array(query_emb), top_k)
    return df.iloc[I[0]]

# ==========================================
# 6. PROMPT TEMPLATE
# ==========================================
PROMPT_TEMPLATE = """You are "Jawad", a professional and friendly sales assistant for our online store.
Answer in a natural, human-like, conversational style.
Use the provided context to answer the customer's question accurately.
Guidelines:
1. Be polite and professional.
2. If you don't find the answer in the context, say: "I'm sorry, I couldn't find information about that. Could you provide more details?"
3. Do not make up any information.
4. If the customer asks in Arabic, answer in Arabic; if the customer asks in English, answer in English.
Context: {context}
Question: {question}
Jawad's Response:"""

# ==========================================
# 7. Chat Function
# ==========================================
def sales_chat(message, history):
    retrieved = retrieve_products(message)
    context = ""
    for _, row in retrieved.iterrows():
        product_name = row.get('product_name', '')
        description = row.get('description', '')
        price = row.get('price', '')
        context += f"Product: {product_name}\nDescription: {description}\nPrice: {price}\n"

    prompt = PROMPT_TEMPLATE.format(context=context, question=message)
    inputs = tokenizer(prompt, return_tensors="pt").to("cuda" if torch.cuda.is_available() else "cpu")

    with torch.no_grad():
        outputs = model.generate(
            **inputs,
            max_new_tokens=250,
            do_sample=True,
            temperature=0.8,      
            top_p=0.95,
            top_k=60,
            repetition_penalty=1.1  # Reduces repetition
        )

    response = tokenizer.decode(outputs[0], skip_special_tokens=True)
    response = response.replace(prompt, "").strip()
    response = response
    return response


# ==========================================
# 8. Gradio Interface
# ==========================================
demo = gr.ChatInterface(
    fn=sales_chat,
    title="Jawad - Sales Agent",
    description="Your smart sales advisor — delivering fast, relevant product recommendations tailored to your needs."
)

demo.launch()