LLM-based ChatBot in Streamlit
This tutorial shows how to build a production-ready LLM chatbot in Streamlit using st.chat_message and st.chat_input, with optional audio input/output, session logging, and a simple admin panel. The code below is a cleaned-up, working template you can adapt for your own app.
1. What This App Does
- Chat UI with streaming responses
- Azure OpenAI backend
- Optional audio input (Whisper) + audio output (gTTS)
- Session logging with TinyDB
- Simple admin prompt editor and analytics
2. Install Dependencies
pip install streamlit openai python-dotenv tinydb pytz gTTS langdetect librosa whisper pandas
If you plan to use Whisper locally, make sure you can run it on your machine (it can be heavy).
3. Set Environment Variables
Create a .env file with:
AZURE_OPENAI_API_KEY=...
OPENAI_API_VERSION=...
AZURE_OPENAI_ENDPOINT=...
PASS_TOKEN=...
4. Full Example
import os
import uuid
from io import BytesIO
from datetime import datetime
import pandas as pd
import pytz
import streamlit as st
from dotenv import load_dotenv
from tinydb import TinyDB
from langdetect import detect
from gtts import gTTS
import librosa
import whisper
from openai import AzureOpenAI
# ---------------------------------
# Config
# ---------------------------------
load_dotenv(override=True)
APP_ID = "sel"
st.set_page_config(layout="wide")
# ---------------------------------
# Models & Clients
# ---------------------------------
@st.cache_resource
def load_whisper_model():
return whisper.load_model("small.en")
@st.cache_resource
def get_azure_openai_client():
return AzureOpenAI(
api_key=os.getenv("AZURE_OPENAI_API_KEY"),
api_version=os.getenv("OPENAI_API_VERSION"),
azure_endpoint=os.getenv("AZURE_OPENAI_ENDPOINT"),
)
whisper_model = load_whisper_model()
client = get_azure_openai_client()
# ---------------------------------
# Helpers
# ---------------------------------
EASTERN_TZ = pytz.timezone("US/Eastern")
def get_eastern_time():
return datetime.now(EASTERN_TZ).strftime("%Y-%m-%d %H:%M:%S")
def play_audio(text, autoplay=False):
try:
text_language = detect(text)
except Exception:
text_language = "en"
tts = gTTS(text, lang=text_language)
mp3_fp = BytesIO()
tts.write_to_fp(mp3_fp)
mp3_fp.seek(0)
st.audio(mp3_fp, format="audio/mp3", autoplay=autoplay)
def input_audio_to_text():
audio_data = st.audio_input(
":red[Click the microphone button to talk, then stop to submit]"
)
if audio_data:
audio_array, _ = librosa.load(audio_data, sr=16000)
text_msg = whisper_model.transcribe(audio_array)["text"]
return text_msg
return None
# ---------------------------------
# DB
# ---------------------------------
os.makedirs("user_logs", exist_ok=True)
db = TinyDB("user_logs/user_log.json")
db_table = db.table("sel_listening")
# ---------------------------------
# UI: Header
# ---------------------------------
st.markdown(
'<p style="text-align:center; font-size:42px;">Your App Name</p>',
unsafe_allow_html=True,
)
st.markdown("***")
st.markdown("### About")
st.markdown(
"This is an example of creating a Streamlit chatbot using an LLM. "
"For questions and support, contact: <jianganghao@gmail.com>"
)
# ---------------------------------
# Login
# ---------------------------------
st.markdown("### Login")
col_login = st.columns([2, 1, 2, 2])
userid = col_login[0].text_input("Please enter your username (nickname):")
access_token = col_login[2].text_input(
"Please enter your access token:", type="password"
)
if access_token != os.getenv("PASS_TOKEN"):
st.markdown(
"#### :blue[Sorry, invalid login token. Please try again or contact the administrator.]"
)
st.stop()
# ---------------------------------
# Session State
# ---------------------------------
if "app_page" not in st.session_state:
st.session_state.app_page = APP_ID
if "user_id" not in st.session_state:
st.session_state.user_id = userid
if st.session_state.app_page != APP_ID or st.session_state.user_id != userid:
st.session_state.clear()
st.session_state.app_page = APP_ID
if "system_instruction" not in st.session_state:
st.session_state.system_instruction = (
"You are an experienced teacher on ***. "
"Ask the user questions to check understanding. "
"Keep responses concise and age-appropriate."
)
if "session_id" not in st.session_state:
st.session_state.session_id = str(uuid.uuid4())[:8]
if "messages" not in st.session_state:
st.session_state.messages = []
intro = {
"role": "assistant",
"content": "Hi, I am your AI learning assistant. Let's start!",
"datetime": get_eastern_time(),
"user_id": userid,
"session_id": st.session_state.session_id,
"app_id": APP_ID,
"event_name": "chat",
}
st.session_state.messages.append(intro)
db_table.insert(intro)
# ---------------------------------
# Admin Panel
# ---------------------------------
if userid == "admin":
with st.expander("Admin: edit system instruction"):
st.session_state.system_instruction = st.text_area(
"", value=st.session_state.system_instruction, height=300
)
st.download_button(
label="Download instruction",
data=st.session_state.system_instruction.encode("utf-8"),
file_name="prompt.txt",
mime="text/plain",
)
st.markdown("***")
# ---------------------------------
# Chat UI
# ---------------------------------
show_audio = st.checkbox("Enable Audio Chat")
container = st.container(height=550)
for message in st.session_state.messages:
if message["event_name"] == "chat":
with container.chat_message(message["role"]):
st.markdown(message["content"])
prompt = input_audio_to_text() if show_audio else st.chat_input("Enter to chat")
if prompt:
user_msg = {
"role": "user",
"content": prompt,
"datetime": get_eastern_time(),
"user_id": userid,
"session_id": st.session_state.session_id,
"app_id": APP_ID,
"event_name": "chat",
}
st.session_state.messages.append(user_msg)
db_table.insert(user_msg)
with container.chat_message("user"):
st.markdown(prompt)
with container.chat_message("assistant"):
try:
stream = client.chat.completions.create(
model="gpt-4o",
messages=[
{"role": "system", "content": st.session_state.system_instruction}
]
+ [
{"role": m["role"], "content": m["content"]}
for m in st.session_state.messages
],
stream=True,
)
response = st.write_stream(stream)
except Exception:
response = "Please avoid inappropriate content in classroom settings."
st.write(response)
assistant_msg = {
"role": "assistant",
"content": response,
"datetime": get_eastern_time(),
"user_id": userid,
"session_id": st.session_state.session_id,
"app_id": APP_ID,
"event_name": "chat",
}
st.session_state.messages.append(assistant_msg)
db_table.insert(assistant_msg)
if show_audio:
with container:
play_audio(response, autoplay=True)
# ---------------------------------
# Analytics (Admin)
# ---------------------------------
if userid == "admin":
st.markdown("### Analytics")
with st.expander("Click to see details"):
df = pd.DataFrame(st.session_state.messages)
df["datetime"] = pd.to_datetime(df["datetime"])
df["session_time"] = df.datetime - df.datetime.iloc[0]
df["session_time"] = df.session_time.apply(lambda x: x.total_seconds())
st.markdown(
f"#### Total turns: {df.query('role==\"user\"').content.count() * 2 + 1}"
)
st.markdown(
f"#### Total words by user: {df.query('role==\"user\"').content.apply(lambda x: len(x.split())).sum()}"
)
st.markdown(
f"#### Total words by agent: {df.query('role==\"assistant\"').content.apply(lambda x: len(x.split())).sum()}"
)
st.dataframe(df)
5. Notes and Improvements
- Security: Use environment variables and rotate keys regularly.
- Scaling: Add rate limiting and store logs in a database (Postgres, SQLite, etc.).
- UX: Add a “Clear chat” button and default welcome messages per session.
- Audio: Whisper + gTTS are optional; disable for lighter deployments.
That’s it. You now have a solid Streamlit chatbot template with LLM streaming, logging, and optional audio support.