Our first Streamlit app

Rapidly Prototyping with AI

I’ve been teaching a class through Pearson on LLMs and ChatGPT with an emphasis on empowering non-coders to learn how to prompt, build test harnesses, and rapidly prototype with LLMs. On our last day I introduced Streamlit, a super simple way to build super quick and dirty prototypes. The goal was to give my students a way to share their prototypes with people with minimal coding. I figured, why not also show the same example here!

Our wine recommending app prototype, complete with a explicit feedback mechanism

Streamlit is super simple and honestly with fewer than 100 lines of code, we can be done with our prototype. Let’s start strong and see the final app, found here on github:

A wine recommending app

# Import necessary libraries
import random

import openai
import streamlit as st
from datasets import load_dataset
from supabase import create_client

# Set API Key
openai.api_key = st.secrets["OPENAI_API_KEY"]

# Initialize DB connection once
def init_connection():
    return create_client(st.secrets["SUPABASE_URL"], st.secrets["SUPABASE_KEY"])

supabase = init_connection()

# System prompt for OpenAI API
system_prompt = '''You are a wine bot that helps clients understand what kind of wine they want. Given a list of wines and a description of the client, tell me what wines they want by giving me the names of the wines. Include a reason preceding each pick to explain to the user why they might like it. Give me the information  as a numbered list of wines with reasons why they might like it.'''

# Cache wine dataset once
def load_wines():
    wine_dataset = load_dataset("alfredodeza/wine-ratings")
    return list(wine_dataset['train'])  # only use train set for now

# Convert wine to string
def convert_wine_to_string(wine):
    return f'{wine["name"]} is from {wine["region"]} and is a {wine["variety"]}. {wine["notes"]}'

# Update reaction in DB
def react_to_row(row, reaction):
        {"reaction": reaction or None}, returning="minimal"
    ).eq("id", row['id']).execute()

# User input elements
user_description = st.text_input("Describe the client",
                                 "The client likes red wine and is looking for a wine to drink with dinner.")
n = st.number_input("How many wines to pull from the cellar?", min_value=1, max_value=10, value=3, step=1)

# Function to get recommendations
def get_recommendations(n=3, user_description=''):
    wines = random.sample(load_wines(), n)
    wines_formatted = "\n---\n".join([convert_wine_to_string(w) for w in wines])
    user_prompt = f'User Description: {user_description}\nWines to select from:\n{wines_formatted}'

    # Create chat completion with OpenAI
    chat_completion = openai.ChatCompletion.create(
        messages=[{'role': 'system', 'content': system_prompt}, {'role': 'user', 'content': user_prompt}]

    # Show the wine recommendations and store in Supabase
    st.write('Wines pulled from cellar to choose from')

    row = supabase.table("response").insert(
        [{"system_prompt": system_prompt, "user_prompt": user_prompt,
          "response": chat_completion.choices[0].message.content, "prototype": "wine"}]
    st.session_state['row'] = row

# Button to get recommendations
    "Get recommendations", on_click=get_recommendations,
    kwargs={'n': n, 'user_description': user_description}

# User reaction
reaction = st.selectbox("How do you feel about the response?", ("", "👍", "👎"))
if 'row' in st.session_state:
        "Submit reaction", on_click=react_to_row,
        kwargs={'row': st.session_state['row'], 'reaction': reaction}

Here is how a user interacts with our app:

  1. The user inputs their wine preferences and selects the number of recommendations they want to receive through the application's interface.

  2. The user clicks on the "Get recommendations" button, triggering the application to randomly select wines from its dataset and request recommendations from the AI model.

  3. The application displays personalized wine recommendations from the AI model along with detailed explanations and a table of the selected wines.

  4. The user has the option to react to the AI's recommendations via a select box, expressing their approval or disapproval.

  5. If a reaction is provided, the user clicks on "Submit reaction", and the application saves the user's feedback to Supabase, which can be used for future improvements to the system.

The goal here is to help people get their prototypes out there with minimal code. Everyone deserves to share their work!