app.py

import geopandas as gpd
import sqlite3
import pandas as pd
import torch
import faiss
import numpy as np
import os
from shapely.geometry import shape
from sentence_transformers import SentenceTransformer
from transformers import AutoTokenizer, AutoModelForCausalLM
import streamlit as st

# Set the environment variables for GPU usage in Hugging Face
os.environ["CUDA_VISIBLE_DEVICES"] = "0"  # Hugging Face uses GPU 0 by default
os.environ["TOKENIZERS_PARALLELISM"] = "false"

# Set device to GPU if available
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
st.write(f"Using device: {device}")

# Step 1: Load and Process Floodland Data
conn = sqlite3.connect('NY.db')
cursor = conn.cursor()

# Load shapefile
gdf = gpd.read_file('S_FLD_HAZ_AR.shp')

# Validate geometries
gdf['geometry'] = gdf['geometry'].apply(lambda geom: geom if geom.is_valid else None)
gdf = gdf.dropna(subset=['geometry'])

# Convert CRS to UTM Zone 18N (New York)
gdf = gdf.to_crs(epsg=32618)

# Calculate acreage (1 square meter = 0.000247105 acres)
gdf['acreage'] = gdf.geometry.area * 0.000247105

# Define flood-prone zones and calculate usable area
flood_prone_zones = ['A', 'AE', 'AH', 'AO', 'VE']
gdf['usable_area'] = gdf.apply(lambda row: row['acreage'] if row['FLD_ZONE'] not in flood_prone_zones else 0, axis=1)

# Convert geometry to WKT format
gdf['wkt_geometry'] = gdf['geometry'].apply(lambda geom: geom.wkt)

# Step 2: Load Embedding Model (Sentence-Transformer)
embedder = SentenceTransformer('all-MiniLM-L6-v2')

# Convert floodland descriptions into text
gdf['text'] = gdf.apply(
    lambda row: f"Flood Zone: {row['FLD_ZONE']}, Subtype: {row['ZONE_SUBTY']}, Acreage: {row['acreage']:.2f} acres, Usable Area: {row['usable_area']:.2f} acres",
    axis=1
)

# Generate text embeddings
embeddings = embedder.encode(gdf['text'].tolist(), show_progress_bar=True)

# Create FAISS index
d = embeddings.shape[1]
index = faiss.IndexFlatL2(d)
index.add(embeddings)

# Store embeddings in DataFrame
gdf['embedding'] = list(embeddings)

# Step 3: Load LLaMA Model for Summarization
llama_model_name = "meta-llama/Llama-2-7b-chat-hf"
tokenizer = AutoTokenizer.from_pretrained(llama_model_name)
model = AutoModelForCausalLM.from_pretrained(llama_model_name, torch_dtype=torch.float16, device_map="auto")

# Function to Generate Summary using LLaMA
def llama_summarize(text, total_acreage, usable_acreage, location_data, max_length=250):
    input_text = f"""
    **Total Land Area**: {total_acreage:.2f} acres
    **Usable Area**: {usable_acreage:.2f} acres
    **Flood-prone Zones**:
     {location_data}
    
    Summarization in sentence
    """

    inputs = tokenizer(input_text, return_tensors="pt").to(device)
    
    # Calculate max_new_tokens based on input size
    input_length = inputs['input_ids'].shape[1]
    max_new_tokens = max_length - input_length
    if max_new_tokens <= 0:
        max_new_tokens = 200  # Ensure at least a few tokens are generated
    
    with torch.no_grad():
        output_tokens = model.generate(
            **inputs, 
            max_new_tokens=max_new_tokens,  # Use max_new_tokens to control the generated length
            temperature=0.7, 
            top_k=50, 
            top_p=0.9, 
            repetition_penalty=1.2
        )
    
    summary = tokenizer.decode(output_tokens[0], skip_special_tokens=True)
    return summary

# Step 4: RAG Summarization Function
def rag_summarize(query, gdf, index, k=5):
    query = query.lower().strip()
    query_embedding = embedder.encode([query])[0]

    # Retrieve top-k relevant documents
    distances, indices = index.search(np.array([query_embedding]), k)
    retrieved_docs = gdf.iloc[indices[0]]

    # Aggregate data
    total_acreage = retrieved_docs['acreage'].sum()
    usable_acreage = retrieved_docs['usable_area'].sum()
    location_data = "\n".join([
        f"- **Flood Zone**: {row['FLD_ZONE']}, **Subtype**: {row['ZONE_SUBTY']}, "
        f"**Acreage**: {row['acreage']:.2f}, **Usable Area**: {row['usable_area']:.2f} acres"
        for _, row in retrieved_docs.iterrows()
    ])

    # Use LLaMA for summarization
    summary = llama_summarize(query, total_acreage, usable_acreage, location_data)
    
    return summary

# Streamlit Interface
st.title("🌊 Floodland Summary Bot (Powered by LLaMA-2)")

# Input for location
user_input = st.text_input("Enter a location (e.g., New York)", "")

# When the user inputs a query, display the summary
if user_input:
    query = user_input.lower().strip()
    summary = rag_summarize(query, gdf, index)
    st.write(summary)