src/app.py

from typing import Optional
import streamlit as st
import pandas as pd
import plotly.express as px
import os
import time
import requests
import torch
from transformers import pipeline

@st.cache_resource
def model_loader():
    model_name = "OrangeBottle/place_description_model_16_bit"
    #model_name = "unsloth/Llama-3.2-3B-Instruct" # Non Fine-Tuned Base Model

    st.write(f"##### Using {model_name}")

    pipe = pipeline(task="text-generation", model=model_name, dtype=torch.bfloat16, device_map="cpu")

    return pipe

@st.cache_data
def text_description_to_locations(description: str) -> list[str]:
    # # Use LLM to convert text description to list of locations
    # # For simplicity, we will just return a hardcoded list here
    # return ["Paris, France", "Rome, Italy", "Barcelona, Spain", "Amsterdam, Netherlands"]
    messages = [
        {"role": "user", "content": f"List 4 popular vacation locations that match the following description: {description}. Return the locations as a comma-separated list. Only return the location names without any additional text. The locations should be geo-searchable cities."},
    ]

    text = llm_request(messages)

    text = text.strip()

    locations = [loc.strip() for loc in text.split(",")] 

    return locations


@st.cache_data
def search_location(location_name) -> Optional[dict[str, float]]:

    url = "https://us1.locationiq.com/v1/search"

    headers = {"accept": "application/json"}

    params = {
        "key": os.environ.get("pk_key"),
        "q": location_name,
        "format": "json"
    }
    response = requests.get(url, params=params, headers=headers)

    response_json = response.json()
    location = response_json[0]
    if not location:
        return None
    latitude = location['lat']
    longitude = location['lon']

    time.sleep(1)  # To respect rate limiting

    return {
        'lat': float(latitude),
        'lon': float(longitude),
    }

@st.cache_data
def describe_location_with_llm(location_name: str) -> str:
    # # Use LLM to generate a description of the location
    # # For simplicity, we will just return a hardcoded description here
    # return f"{location_name} is a beautiful city with rich history and culture."
    messages = [
        {"role": "user", "content": f"Describe the location: {location_name} in detail."},
    ]

    text = llm_request(messages)

    text = text.strip()

    return text

@st.cache_data
def llm_request(messages) -> str:

    pipe = model_loader()

    max_new_tokens=128
    use_cache = True
    temperature = 1.5
    min_p = 0.1

    outputs = pipe(messages, max_new_tokens=max_new_tokens)
    print(outputs)
    text = outputs[0]['generated_text'][-1]["content"]

    return text

async def main():
    st.write("""
    # Vacation Location Explorer with LLM and Maps
    """)
    # Text box for what to search for
    search_query = st.text_input("Describe a vacation experience you would like to have:")

    # Use LLM to generate a couple of locations based on the search query
    # For simplicity, we will just hardcode some locations here
    if search_query:
        st.write(f"Based on your input: '{search_query}', here are some recommended locations:")
        recommended_locations = text_description_to_locations(search_query)
        for location in recommended_locations:
            st.write(f"- {location}")
    else:
        st.write("Please enter a vacation experience to get recommendations.")
        return
    
    locations = []

    for location in recommended_locations:
        result = search_location(location)
        if result:
            lat = result['lat']
            lon = result['lon']
            locations.append((location, lat, lon))
        else:
            st.write(f"Location {location} not found.")
    
    if not locations:
        st.write("No locations found to display on the map.")
        return
    
    df_locations = pd.DataFrame(locations, columns=["Location", "Latitude", "Longitude"])
    df_locations['size'] = 3  # Fixed size for all points, can be modified as needed

    fig = px.scatter_map(df_locations, lat="Latitude", lon="Longitude", hover_name="Location", zoom=3, size="size", color="Location")
    fig.update_layout(mapbox_style="open-street-map")
    fig.update_layout(margin={"r":0,"t":0,"l":0,"b":0})
    event = st.plotly_chart(fig, on_select = "rerun")
    
    if len(event['selection']["points"]) == 0:
        st.write("Please select a location on the map to get more information.")
        return
        
    description = describe_location_with_llm(event['selection']["points"][0]["hovertext"])
    st.write(description)

import asyncio
asyncio.run(main())