LLM Chatbot

Overview

A small, readable chatbot built directly on Google’s google-genai SDK, with no LangChain or other framework. The aim was to understand what a “chat” with an LLM actually is under the hood. In particular: how a model that is stateless between requests can remember what you said three messages ago.

The key idea: conversation memory

An LLM API call has no memory of its own. To make a conversation feel continuous, the app has to resend the earlier turns with every new message. This project does that explicitly:

1. Store the history. Streamlit’s session_state holds a chat_history list that survives page reruns.
2. Open a chat with it. On every interaction, a Gemini chat session is created seeded with that history, using the SDK’s own Content objects with user and model roles.
3. Record each exchange. When you ask a question, the answer comes back and both your message and the reply are appended to the history.
4. Show the transcript. The full conversation is displayed below the latest answer.

So when you ask “What did I ask you earlier?”, the model can answer, because it has been given the whole conversation every time.

Design choices

• Model-agnostic. The model is a single constant (gemini-2.0-flash), so switching to another Gemini model is a one-line change, and the history-handling pattern carries over to other chat APIs.
• Minimal dependencies. Only three: streamlit for the UI, google-genai for the model and python-dotenv for the API key. The whole app is easy to read end to end.

Running it

pdm install
echo "GOOGLE_API_KEY=<your key>" > .env
pdm run streamlit run app.py

The repository includes a short demo GIF of a multi-turn conversation.