Personalized Chatbot Prototype

This is a very small Flask-based chatbot prototype with:

• a single HTML/JavaScript frontend
• a Flask backend
• an OpenAI Chat Completions passthrough
• SQLite-backed memories scoped to a browser session
• OpenAI embeddings for stored memories
• DBSCAN clustering for semantically similar memories
• LLM-generated cluster descriptions plus per-memory alignment markers
• prompt-based personalization injection during chat
• prompt templates stored as editable text files

Run locally with Anaconda Prompt

If PowerShell is giving you trouble, use Anaconda Prompt instead.

1. Open Anaconda Prompt

Open the Anaconda Prompt from the Start menu.

2. Go to the project folder

cd C:\Users\markd\Documents\GitHub\llm-personalization

3. Create a Conda environment

Use Python 3.11 or newer.

conda create -n llm-personalization python=3.11 -y

4. Activate the environment

conda activate llm-personalization

5. Install dependencies

pip install -r requirements.txt

6. Set environment variables

set OPENAI_API_KEY=your_api_key_here
set OPENAI_MODEL=gpt-4o-mini

If you want to keep using the repo's default chat model, you can skip OPENAI_MODEL.

7. Run the app

python app.py

Then open http://127.0.0.1:5000.

Optional: PowerShell workflow

If you want to use PowerShell later, this repo also works with a regular virtual environment:

python -m venv .venv
.venv\Scripts\Activate.ps1
pip install -r requirements.txt
$env:OPENAI_API_KEY="your_api_key_here"
$env:OPENAI_MODEL="gpt-4o-mini"
python app.py

Current scope

Version 0.1 is intentionally simple:

• no authentication
• no streaming
• no persistent chat history
• memories are injected into the chat prompt for personalization
• memories are per browser session, not shared across users
• memories expire and are capped to keep the demo DB from growing without bounds
• memory clusters are generated from embeddings and labeled by an LLM

The SQLite database is included now so we can add personalization later without changing the project shape too much.

Configuration

Prototype tuning values live in config.txt.

Right now it includes:

• MEMORY_TTL_DAYS
• MAX_MEMORIES_PER_SESSION
• MAX_TOTAL_MEMORIES
• WEIGHT_HALF_LIFE_MINUTES
• CLUSTER_EPSILON
• CLUSTER_MIN_SAMPLES
• CHAT_MODEL
• EMBEDDING_MODEL
• EMBEDDING_DIMENSIONS

Restart Flask after changing config.txt so the updated values are loaded.

Railway deployment notes

For Railway deployment:

• the app can be started with the included Procfile
• set OPENAI_API_KEY in Railway variables
• optionally set OPENAI_MODEL
• set MEMORY_DB_PATH=/data/memory.db if you attach a Railway volume mounted at /data

Using a persistent volume is recommended if you want SQLite memories to survive redeploys and restarts.

Prompt files

Prompt templates live in the prompts/ folder:

• chat_system.txt for the main assistant behavior
• memory_extraction.txt for extracting likes/dislikes
• personalization_injection.txt for injecting memories into chat prompts
• cluster_labeling.txt for generating short labels for memory clusters

These files are read by the backend at runtime, so you can inspect and edit them directly.