ChatAI the GWDG ChatBot
ChatAI is GWDG’s AI chat service, built for researchers. The service runs on on-premises hardware and guarantees data-protection-compliant use of multiple language models.
Access: https://chat-ai.academiccloud.de (Academic Cloud login required)
Available Models and Selection
GWDG offers both internal models (hosted on its own hardware) and external models (via third-party providers).
Adjustable Parameters (Quick Overview)
System Prompt
Defines the role of the AI (e.g., “You are a climate-science expert”). Clear role assignment greatly improves answer quality.
ChatAI includes a library of ready-made system prompts—for example, “You are a proofreader,” “You are a data-analysis assistant,” or “You are a scientific translator.” Open the Templates menu (right of the prompt box) and pick the role that best matches your task; the chosen prompt is inserted automatically, saving time and ensuring consistent role setup.
Temperature (range: 0–2)
Controls answer creativity:
- 0–0.5: Precise, factual answers (ideal for data analysis, code)
- 0.5–1: Balanced creativity
- 1–2: Highly creative answers (best for brainstorming)
Top-p (range: 0–1)
Controls answer diversity. Not every model lets you change top-p. Some models—including those that use deterministic decoding (beam search) or that have a server-side safety profile—ignore the setting entirely or apply an internal fixed value. If the UI shows the control as disabled or adds a “fixed” tag next to the model name, any value you enter will be ignored. Always check the model description or tooltip to see whether top-p is honored for that specific model.
- 0.1–0.3: Focused, precise answers
- 0.8–1: More varied phrasing
- Default: 0.9 for most tasks; several models provide their own recommended values.
Toolbar Actions and Advanced Features
Arcana (RAG System)
RAG – Retrieval-augmented generation combines on-the-fly document retrieval with an LLM so it can cite and reason over up-to-date, user-supplied sources.
- Upload Your Own Documents (PDF, text, Markdown)
- AI answers based on your papers
- Share With Colleagues via ID and key
- Transmitted over HTTPS/TLS; documents are not stored after preprocessing
- Requires an Arcana account and selection of the model “Meta Llama 3.1 8B RAG.”
Other Toolbar Actions
- Attach (PDF Upload): Add documents directly in chat
- Listen (Speech Recognition): Ask questions by microphone (supported mainly in Chrome- or Edge-based browsers)
- Import / Export: Save and reuse conversations
- API Access: Request an OpenAI-compatible API key via the KISSKI booking form (SAIA service)
Note: A traditional web search is not currently available. Models work with their training data up to their respective cut-off dates.
ChatAI and Data-Protection Compliance
Internal Models (recommended for research):
- ✅ No conversation content stored server-side
- ✅ Data exists only locally during the session
- ✅ Minimal usage statistics (username, timestamp, model) logged for service monitoring
- ✅ GDPR-compliant
External Models (OpenAI):
- ⚠️ Data is forwarded to Microsoft Azure (hosts GPT-4, GPT-4o, etc.)
- ⚠️ Prompts and responses retained for up to 30 days; early deletion is not yet configurable
AI Inside the Minerva Messenger (MPG)
The Max Planck Digital Library (MPDL) bundles key digital services for the Max Planck Society (MPG), including access to advanced AI tools like ChatGPT and Llama within the Minerva Messenger for all MPG members.
1. Access Requirements
-
These services are available to MPG members.
-
Access is provided via the central Single Sign-On (SSO) system.
-
You will need your MPG credentials.
2. Logging In
-
Go to the Minerva Messenger login page:
https://minervamessenger.mpdl.mpg.de/login -
Log in using your SSO account.
3. Using AI Tools
-
Once logged in, you can chat directly with ChatGPT or Llama in the Minerva Messenger.
-
Start a new direct message or add the desired bot to a private channel.
-
Begin your conversation as you would with any other user—the AI will respond directly in the chat.
Note:
The MPDL ensures that all services are provided securely and centrally for the MPG. For more information, visit the MPDL website.
Via MPDL you also have acces to various other helpful tools, including iThenticate, for plagiarism detection as well as several tools to share raw data and results.
Support:
If you have any questions or encounter issues, I’d recommend to contact the MPDL support team.
Important Links:
