RITA — Radiotherapy Intelligent Treatment Assistant

RITA (SI_RITA) is an AI assistant that runs inside RayStation as a scripting tool. It provides a chat GUI backed by a LangGraph agent that can call tools — including tools that read data from the currently open RayStation patient — to help with radiotherapy treatment-planning tasks.

The agent can be driven by either OpenAI (GPT‑5) or Anthropic (Claude Sonnet) models, selectable from a dropdown in the interface.

---

Features

• 🖥️ Desktop chat GUI built with tkinter / ttkbootstrap (dark theme).
• 🤖 Single general agent built on LangGraph, with an automatic tool-calling loop.
• 🔀 Model switching at runtime — gpt-5 (OpenAI) or sonnet (Anthropic Claude Sonnet).
• 🛠️ Tools the agent can call:
  • RayStation — read patient name, date of birth, gender, and ID from the open case.
  • Math — add, subtract, multiply, divide.
  • General — count characters, get the current time.
• 📊 LangSmith tracing for observability (project SI_RITA).

---

Architecture

The agent is a small LangGraph state machine with one node and a tool loop:

START ──▶ general_agent ──┬─ (requested a tool?) ──▶ general_tool_node ──┐
                          │                                              │
                          │◀─────────────────────────────────────────────┘
                          └─ (no tool) ──▶ END

• general_agent sends the conversation (plus a system prompt) to the selected LLM.
• If the model requests a tool, general_tool_node executes it and loops the result back.
• When the model responds without a tool call, the turn ends.

---

Project structure

SI_RITA/
├── rita.py                       # GUI entry point (RITA_GUI) — run this
├── requirements.txt              # Python dependencies
├── .env                          # API keys (gitignored — never commit)
├── .gitignore
├── README.md
└── src/
    ├── rita_graph.py             # LangGraph definition, AgentState, compiled `agent`
    ├── my_langgchain_tool/
    │   ├── raystation_tool.py    # patient-data tools (RayStation only)
    │   ├── math_tool.py          # arithmetic tools
    │   ├── general_tool.py       # misc utility tools
    │   └── code_tool.py
    ├── sub_agent/general/
    │   ├── general.py            # the general_agent node + model selection
    │   └── general_sys_prompt.py # system prompt
    └── util/
        └── helpers.py            # tool/message format conversion helpers

---

Configuration

All secrets live in a .env file at the project root (loaded automatically at startup).

Create .env with:

# OpenAI — https://platform.openai.com/api-keys
OPENAI_API_KEY=sk-proj-...

# Anthropic / Claude — https://console.anthropic.com/settings/keys
ANTHROPIC_API_KEY=sk-ant-...

# LangSmith tracing — https://smith.langchain.com
LANGSMITH_API_KEY=...

⚠️ Never commit .env. It is already listed in .gitignore. Keep your keys out of source files.

---

Local development (macOS / Linux / Windows, outside RayStation)

Use this to edit and iterate. Note that the RayStation tools (raystation_tool.py) only work inside RayStation, since the raystation module ships only with RayStation's own Python.

# 1. Create and populate a virtual environment
python3 -m venv .venv
.venv/bin/python -m pip install -r requirements.txt   # Windows: .venv\Scripts\python -m pip ...

# 2. Add your keys to .env (see Configuration above)

# 3. Run
.venv/bin/python rita.py

---

Deployment inside RayStation

RayStation runs its own embedded Python interpreter into which you cannot pip install directly. Instead, the libraries are installed into a separate venv and made importable by prepending its site-packages to sys.path at the top of the script (setupPath in rita.py).

1. Find RayStation's Python version

In RayStation's scripting console:

import sys
print(sys.version)          # note the major.minor version (e.g. 3.9)
print(sys.maxsize > 2**32)  # True = 64-bit

2. Build a matching venv on Windows

The venv must be built with the same OS (Windows) and Python major.minor version as RayStation — several dependencies ship compiled wheels that are version- and platform-specific.

Install the dependencies into the site-packages folder that rita.py points at:

py -3.9 -m pip install --target "\\zkh\...\RITA\.venv\Lib\site-packages" -r requirements.txt

3. Point the script at that folder

rita.py sets sys.path before importing any third-party library:

path = "//zkh/appdata/Raystation/Research/ML/Paritt/RITA/.venv/Lib/site-packages"
setupPath(path)   # sys.path.insert(0, path)

Update path to wherever your Windows venv's site-packages lives.

4. Provide the .env

Place a .env (with your real keys) at the project root next to the code on the share. It is loaded via an absolute path, so it is found regardless of RayStation's working directory.

---

Usage

1. Launch RITA (python rita.py, or run the script from within RayStation).
2. Pick a model from the dropdown: gpt-5 or sonnet.
3. Type a prompt and send. Responses stream into the chat window; tool calls are shown inline.

Example prompts:

• "What is the patient's name and ID?" → calls the RayStation tools.
• "What is 128 divided by 4?" → calls the math tools.
• "What time is it right now?" → calls get_current_time.

---

Observability

Runs are traced to LangSmith under the project SI_RITA (configured in src/rita_graph.py). View traces at https://smith.langchain.com with the account that owns your LANGSMITH_API_KEY.

---

Security notes

• .env, .venv/, and __pycache__/ are gitignored.
• Do not hardcode API keys in source files — read them from the environment / .env.
• If a key was ever committed or shared, rotate it (regenerate it in the provider console).
• Patient data accessed via the RayStation tools stays within your RayStation session; be mindful of what patient information is sent to third‑party LLM providers.

---

Requirements

• Python matching your run target (RayStation's version for deployment; any 3.9+ for local dev).
• Dependencies: langgraph, langchain-core, langchain-openai, langchain-anthropic, langsmith, ttkbootstrap, python-dotenv (see requirements.txt).
• API keys for the providers you intend to use (OpenAI and/or Anthropic) and LangSmith.