HK Policy RAG

A Retrieval-Augmented Generation (RAG) assistant for Hong Kong government policies, built with Streamlit and OpenAI.

The app provides a chat interface where users can ask questions about HK policies. The RAG backend is designed to retrieve relevant document chunks and inject them into the LLM prompt — retrieval sources are pluggable stubs ready to be wired up.

---

Features

• Conversational chat UI with message history
• Streaming LLM responses via OpenAI (gpt-4o-mini)
• Parallel RAG retrieval scaffold (pluggable backends)
• Automatic summarisation of older conversation history
• Suggestion pills for common questions
• Debug mode (?debug=true) to inspect full prompts

---

Requirements

• Python 3.14+
• uv (recommended) or pip
• An OpenAI API key

---

Setup

1. Clone the repo

git clone https://github.com/<your-username>/hk-policy-rag.git
cd hk-policy-rag

2. Create a .env file

cp .env.example .env   # then fill in your keys

OPENAI_API_KEY=sk-...

3. Install dependencies

With uv (recommended):

uv sync

Or with pip:

pip install -r requirements.txt

---

Running

# uv
uv run streamlit run streamlit_app.py

# pip / activated venv
streamlit run streamlit_app.py

The app will be available at http://localhost:8501.

---

Project Structure

hk-policy-rag/
├── streamlit_app.py   # Streamlit UI + LLM integration
├── main.py            # Entry point (placeholder)
├── pyproject.toml     # Project metadata & dependencies (uv)
├── requirements.txt   # Pinned direct dependencies (pip)
├── .env               # API keys (not committed)
└── .env.example       # Template for .env

---

Wiring up RAG

The two stub functions in streamlit_app.py are the integration points:

def search_relevant_docs(query: str) -> str:
    """Return relevant policy document chunks for the query."""
    ...

def search_extra_context(query: str) -> str:
    """Return supplementary context for the query."""
    ...

Both are called in parallel via ThreadPoolExecutor. Return a non-empty string to have the content injected into the LLM prompt automatically.