AI Transparency Statement

Goal

Typical LLMs have the advantage, but also the problem, of having been trained on an incredible amount of data. This means that they know a lot, but are often unable to answer very specific questions. In these cases, LLMs are very prone to hallucination, which means that they basically make things up. One way to improve the performance of LLMs for very specific questions is to use Retrieval-Augmented-Generation (RAG). Here, users provide custom documents that contain the knowledge base they want to ask questions about later. Before an LLM responds to a user’s query, the most relevant documents previously provided by the user are retrieved and provided to the LLM as additional context.

In our innovative approach, we provide a reference section at the bottom of the answer where you can find the actual part of the document our RAG associated with your question. As these references are quoted directly from the documents provided by the user, there is no possibility of hallucination.

General Functionality

This section briefly outlines the individual steps required to process a RAG request.

Ingesting Documents

In order to use RAG, a user must first provide a knowledge base, i.e. a set of documents. These documents are uploaded, converted to markup and then indexed in a database. Here they are chunked and then transformed into a vector by a special embedding model. This vector is then stored in a vector database. In this way, each document provided by a user must be processed to build the knowledge base. This knowledge base, also known as “Arcana”, is stored at the GWDG until the user explicitly deletes it!

Submitting Requests

Once the general knowledge base has been built, a user can submit queries. Before these queries are sent to the LLM, they are also transformed into a vector representation using the same embedding model that was used to create the knowledge base in the previous step. This vector is then used in a similarity search on the previously created vector database to look for parts in the previously captured documents that have a similar meaning. A configurable number of similar vectors are then returned and passed to the LLM along with the user’s original request.

Generating an Answer

The LLM uses the additional information to provide a more specific response to the user’s request. This is already much less susceptible to hallucinations, but they are still possible. In our approach, we provide explicit references to the documents containing the chunks used by the LLM to formulate the answer. These references are provided in a special reference box at the bottom, which contains the actual citations from the original documents provided by the user. Therefore, no hallucination is possible on these references.

Further Considerations

Access to the Stored Documents

The ingested knowledge base can be freely shared with other users. To do this, the Arcana ID must be provided. IMPORTANT: This gives users full access to your documents and cannot (yet) be revoked individually!

Storage of your Documents

Your indexed data will remain on GWDG systems at all times. We will not share your documents with third parties.

Processing of your Requests

The processing of your requests, including generating the embeddings, retrieving the relevant documents and performing the inference, is all done on GWDG hardware. Neither your requests nor your indexed documents are shared with third parties.

Migration Guide

With the introduction of our new RAG Manager, we have improved the UI and indexing process. Due to architectural changes, the new interface is only partially backward compatible, particularly in terms of file storage and indexing.

This guide explains:

• What still works between the old and new interface
• What does not work
• How to migrate your existing Arcanas from the old manager to the new System

What Does Work

• All existing Arcanas created by the old manager continue to work
• You can chat with a legacy Arcana via the chat interface.
• You can still modify a legacy Arcana in the old interface.
• The old interface remains fully operational.
• You can delete an Arcana created in the old interface using the new interface.

What Does Not Work

• Modifying an Arcana across interfaces is not fully supported:
  • You can only delete an Arcana or files created in the old interface using the new interface.
  • You cannot modify and see an Arcana created in the new interface using the old interface.
• Re-indexing a legacy Arcana using the new interface does not work, because the file storage architecture differs.

Migration Steps

To migrate a legacy Arcana to the new RAG Manager:

1. Identify Legacy Arcanas

In the new interface, legacy Arcanas are marked with a “Legacy Arcana” tag in the Arcana details view.

Picture 1: Arcana Details View showing “Legacy Arcana” tag

Clicking the tag will show you more information and provide a link to this migration guide (see Picture 2).

Picture 2: Pop-up after clicking the tag, with migration guidance

2. Choose Your Migration Approach

You have two options for migrating a legacy Arcana:

Option A: Re-create the Arcana

• Delete the Arcana entirely.
• Create a new Arcana in the new interface with the same name.
• Upload your files again.
• Click “Generate Index” and follow the usual steps.

Option B: Replace Files in Existing Arcana

• Delete all files in the existing legacy Arcana.
• Upload the files to the same Arcana using the new interface.
• Click “Index Generation”, then click “Delete Index” and finally click “Generate Index”.

RAG Service

RAG (Retrieval-Augmented Generation) is an advanced AI technique designed to improve the accuracy, reliability, and contextual relevance of AI-generated responses. Traditional AI models, such as large language models (LLMs), rely solely on pre-trained data to generate answers. While these models can provide insightful responses, they are limited by the information they were trained on, which may become outdated or may not cover specific topics in detail.

RAG overcomes these limitations by integrating an external retrieval process before generating a response. Instead of relying purely on static knowledge, RAG actively searches for relevant data from external sources, such as document databases, APIs, or knowledge repositories. The retrieved data is then fed into the AI model along with the original user query, ensuring that the response is fact-based, up to date, and contextually relevant.

This approach is particularly valuable for applications requiring real-time information access, such as customer support, research, healthcare, legal advice, and financial services. By leveraging external data, RAG enhances AI’s ability to provide more precise answers, reduces misinformation, and improves trust in AI-driven decision-making.

How the RAG Service Works

The RAG service follows a three-stage process: Retrieval, Augmentation, and Generation.

1. Retrieval Phase

The system maintains a structured external knowledge base, stored in Arcana-based ChromaDB. This database contains documents, articles, technical manuals, and other relevant data sources. When a user submits a query, the ChromaDB engine performs a search to find the most relevant documents. The retrieval process is powered by vector-based similarity matching, which identifies information that closely matches the meaning and context of the user’s query. This approach ensures that even if the exact wording of the query differs from stored data, the system can still locate relevant information.

2. Augmentation Phase

Once the system retrieves relevant documents, they are combined with the original user query to form an enriched input. This augmented input helps the AI model understand specialized or proprietary information that may not have been part of its original training data. The retrieved documents serve as a knowledge injection, ensuring that responses are grounded in verified, real-world data rather than relying on the model’s internal assumptions.

3. Generation Phase

The AI model processes the combined input (original query + retrieved documents). It then generates a response that integrates both its pre-trained knowledge and the newly retrieved information. This response is more accurate, relevant, and fact-based compared to responses generated by traditional AI models that lack external retrieval. The system can also provide source references, increasing transparency and allowing users to verify the information provided.

Key Benefits of the RAG Service

1. Improved Accuracy

   • Since the AI retrieves real-world data before generating responses, it significantly reduces errors and outdated information.
   • This ensures that responses are more reliable, precise, and factually correct.

2. Reduction of AI Hallucinations

   • Traditional AI models sometimes generate responses that sound plausible but are actually incorrect or misleading.
   • RAG minimizes this risk by ensuring that responses are anchored in retrieved, verifiable data, rather than being purely speculative.

3. Domain-Specific Customization

   • Organizations can integrate proprietary databases, making the AI highly specialized for their industry or use case.
   • Whether for healthcare, legal, finance, engineering, research, or customer support, RAG can be tailored to provide highly relevant responses.

4. Enhanced Explainability and Transparency

   • Unlike traditional AI models, which provide answers without explaining their reasoning, RAG can cite sources for its responses.
   • Users can trace back the information to the retrieved documents, improving trust and accountability in AI-generated content.

5. Access to Real-Time and Dynamic Knowledge

   • Unlike static AI models that rely only on pre-trained knowledge, RAG can fetch and integrate the latest available information.
   • This is especially useful for industries where information changes frequently, such as market trends, regulatory compliance, technical troubleshooting, and scientific research.

6. Better User Experience

   • By retrieving and integrating the most relevant information, RAG allows AI to provide more complete and meaningful answers to user queries.
   • This leads to better decision-making, improved efficiency, and a more user-friendly AI interaction.

Setting up an Arcana

Table of contents:

• First Login
• Creating an Arcana
• Uploading Files
• Viewing your Access Link
• Updating files in an Arcana

The process screenshots contain blank blocks. These are in the positions that will be replaced with the username.

First Login

Go to the main Arcana page and click on register. You will need to log in with your Academic Cloud account first.

Once this is done, you are now able to go to the dashboard which looks like this:

You can find an overview of your profile under the profile tab. This is also the place, where you can delete your profile data.

Creating an Arcana

Once you have the dashboard open, you can navigate to the Arcanas section by either clicking on it in the navigation bar at the top or clicking on the “My Arcanas Box”. On the Arcanas section clicking on “+ Add New Arcana” will open the Create Popup. First, you need to specify a name for this arcana.

Enabling the “Secure Arcana Name” slider will append a random string to the Arcana’s name. This ensures that only users who know the exact name of the Arcana can chat with it.

Note: All of your Arcanas are public by default! This means that anyone who knows the exact Arcana name can chat with it. However, this does not mean that they can directly access your files.

Once created, it should appear in your Arcana list.

By clicking on an Arcana, you can open the view of that Arcana content.

Uploading Files

On the Arcana page click on “Upload Files” to start adding your files to the arcana

Select one or multiple files to upload. You can upload files of various types. The following file formats are supported:

• Text (.txt)
• Markdown (.md)
• Word (.docx, .dotx, .docm, .dotm)
• Powerpoint(.pptx, .potx, .ppsx, .pptm, .potm, .ppsm)
• PDF (.pdf)
• HTML (.html, .htm, .xhtml)

Once uploaded, your files should appear.

You can check the file details by clicking on the “File Info” icon next to each file.

This is also the page, where after generating the index you can download the Docling JSON file or download the annotated Markdown file. Also, you have the option to upload a new version of the file. More importantly, you can download the Markdown file and update the markers mentioned in the Docling process. This is very useful to fine tune the splitting and marking of the file for better indexing and retrieval. Once these changes are made, uploading it will start a check process to make sure the content still matches the uploaded material.

Now that the file or files are uploaded, you can start the index generation and file conversion by clicking on “Index Generation”.

Which will change the status to a blue “pending” and finally a green “indexed”.

Viewing your Access Link

Once your Arcana is indexed, you will be able to generate the Access Link and get the ID for this Arcana. Click on “Access Link” in the Actions tab of the respective Arcanas to generate it.

You can click on the link or copy it into a new tab, which will open Chat AI with the Arcana ID preset into the advanced options.

Now you are able to share the link as well as the IDs for this Arcana.

Updating files in an Arcana

There is an option to download and customize annotations for uploaded files, or modify previous manual or generated annotations. After the file has been uploaded and processed, you can update its annotation if necessary.

The details dialog for a PDF looks like this:

For all file types, there is the option to download the annotated Text/Markdown file. In case of a PDF, this annotated Markdown file - also called Markdown Plus - is automatically generated by the Docling process. There is also the option to upload a manual annotated file (text or markdown). This is very useful to manually set the annotations or to update them. Additionally, many files have a JSON file from the Docling process that can be downloaded and viewed.

The available annotations are explained in the Docling process under the heading Markdown Plus Annotations.

How to use Arcana

This guide walks you through two primary ways to access an Arcana:

1. Quick Start: Using an Access Link (recommended for beginners)
2. Advanced: Manual Access via Arcana-ID (for power users and integration)

To use an Arcana, you first need to set one up or use a publicly available Arcana.

Quick Start: Use an Access Link (Easiest Way)

If you have been provided with an Access Link to an Arcana (e.g., for a research paper, course material, or institutional document) or if you have generated your own following the Getting Started Guide, this is the fastest and most straightforward way to begin.

Steps

1. Click the Access Link shared by the Arcana owner or yourself.
2. This opens a pre-configured Chat-AI interface, automatically configured for that Arcana.
3. Begin asking questions - the system will retrieve relevant content from the indexed material and generate responses grounded in the source.

Manual Access: Use the Arcana-ID

For greater control, you can manually connect to an Arcana using its unique Arcana-ID. The arcana ID is also needed when interacting with an Arcana via the Chat Completions API.

Steps

1. Navigate to Chat-AI.

2. In the Settings menu (right side), enable GWDG-Tools.

3. In the Arcana field, enter the Arcana-ID (e.g., user1/arcana-123456).

4. Select the recommended model: Qwen 3 30B A3B Instruct 2507

   Info

   If you see the message “Could not use tools: tools possibly not supported”, while interacting with the RAG, switch to this model - it is optimized for tool-based retrieval.

5. Adjust temp and top_p to the lowest possible.

6. Start interacting with the Arcana through the chat interface.

   Info

   While Arcanas and GWGD-Tools can work with other models, Qwen 3 30B A3B Instruct 2507 delivers the most consistent performance with the RAG functionality.

Generating an Access Link for an Arcana

To generate an Access Link for your Arcana follow the getting started guide section Viewing your Access Link.

Best Practices for Effective Use

To ensure high-quality, accurate responses, follow these guidelines:

1. Adjust ’temp’ and ’top_p'

For Arcanas we recommend to set ’temp’ to 0.0 and ’top_p’ to 0.05 for optimal results. For more information on the influence of ’temp’ and ’top_p’ see this section in the Chat-AI Documentation.

2. Use Specific Keywords from the Source

Our RAG-System processes documents by converting PDFs into structured Markdown. The LLM response is then improved by extracting relevant information from those documents for your prompt. For further information, see the Docling process.

Include terms, phrases, or concepts present in the original document in your prompt:

• Vague: “Tell me about this dataset.”
• Specific: “What evaluation metrics were used in the ImageNet benchmark?”

3. Customize the System Prompt

For important Arcanas (e.g., study guides, lab protocols or scientific papers), adjust the System Prompt to define the desired tone, format, and response style.

Example

You are a helpful assistant named Sandra working at the GWDG for many years. reply precisely to the question.
You will be provided with references at the beginning that might be helpful.
If you do not know the answer answer honestly that you do not know the details.
Please refer to context you use for your reply using the [REF NUMBER] Syntax close to its usage.
If the references provided are confusing, you may ask the user to provide the necessary details to identify which to use.

4. Review References in the Response

After each query, the system returns:

• The response, enriched with context from the indexed material.
• References below the answer, including:
  • Source file name (e.g., imagenet.pdf)
  • Page reference

References are rendered in Markdown, allowing you to trace each claim back to its original source. Originally, the information might have been in a PDF, which got converted into an annotated Markdown file as you can learn in the Docling process.

Example

• Prompt: What is ImageNet?
• Example Response: ImageNet is a large-scale ontology of images built upon the backbone of the WordNet structure. It aims to populate the majority of the 80,000 synsets of WordNet with an average of 500-1000 clean and full-resolution images, resulting in tens of millions of annotated images organized by the semantic hierarchy of WordNet.

API Usage

The Arcana Manager exposes and api endpoint for you to use. With the api endpoint you can perform all the actions that the frontend Arcana Manager is capable of.

To use the API, you’ll first need an API key, which you can request on our website. This key works across all our services. If you’re already using the LLMs via the api, you can use the same key. Otherwise you can request your key following this documentation.

Base Route

All Routes for the RAG/Arcana-API start with https://chat-ai.academiccloud.de/v1/arcanas/api/v1/

Note that every api request needs to start with https://chat-ai.academiccloud.de/v1/. Routes that start with https://chat-ai.academiccloud.de/arcanas/ are not compatible with API keys.

See below for a full list of API routes.

API Minimal Example

Here is a minimal working example of how use the API.

curl \
  --url https://chat-ai.academiccloud.de/v1/arcanas/api/v1/ \
  --header 'Accept: application/json' \
  --header 'Authorization: <api_key>'

{
    "message":"Welcome to the SAIA RAG Web API",
    "docs_url":"http://chat-ai.academiccloud.de/arcanas/api/v1/docs"
}

Examples and API Documentation

Get All Your Arcanas

curl \
  --url https://chat-ai.academiccloud.de/v1/arcanas/api/v1/arcana/ \
  --header 'Accept: application/json' \
  --header 'Authorization: <api_key>'

Get a Specific Arcana

curl \
  --url https://chat-ai.academiccloud.de/v1/arcanas/api/v1/arcana/<arcana-name> \
  --header 'Accept: application/json' \
  --header 'Authorization: <api_key>'

Add a new File to your Arcana

Every new file needs to be send in its own request.

curl -i -X POST "https://chat-ai.academiccloud.de/v1/arcanas/api/v1/arcana/<arcana-name>/files/" \
    -H "accept: application/json" \
    -H 'Authorization: <api_key>' \
    -H "Content-Type: multipart/form-data" \
    -F "file=@<filename>"

The response will be null with a status code of 200 if the upload worked

Index an Arcana

Index generation is synchronous i.e. the HTTP request only responds once the index has been created. If the request is canceled, the index generation will finish in the background.

Full List of Routes

The list of all available routes can be found under:

• In Web-Browser: https://chat-ai.academiccloud.de/arcanas/api/v1/docs
• Via API: https://chat-ai.academiccloud.de/v1/arcanas/api/v1/openapi.json

Note that the routes are relative i.e. they need to be prefixed with https://chat-ai.academiccloud.de/v1/arcanas/api/v1.

Docling process

The RAG Service provides users with an efficient way to upload and process PDF documents using Docling. The system converts uploaded PDFs into Markdown format while also automatically annotating them. This enhanced Markdown output, referred to as Markdown Plus, includes metadata and structural annotations for improved document parsing and customization.

This service is using a fork of the Docling API with many modifications, which will be published in the future.

Process Flow

1. File Upload: Users upload their PDF documents through the RAG interface.
2. Conversion Process: Once the PDF file is uploaded, the user can convert into the Markdown Plus format by clicking on Index Generation -> Generate Index. This will also convert the files.
3. Annotation & Metadata: The Markdown Plus output is automatically annotated with structural markers.
4. By clicking the “File Info” icon on the right, users can download the annotated Markdown file, adjust split markers, and re-upload it for further processing.

   

5. Validation: Upon re-upload, a validation process ensures that the paging structure remains intact.

Markdown Plus Annotations

Docling generates a structured and annotated Markdown file using the following markers:

1. Page Marker

• Format: [Page (number)]: #

• Purpose: Indicates the beginning of a new page in the original PDF document.

• Example:

  [Page 1]: #

2. Vertical Position Marker

• Format: [Y: (number)]: #

• Purpose: Represents the approximate vertical position of an item on the page (The height of the pages is scaled to 1000 lines).

• Details:

  • Each page is divided into five sections.
  • If no continuous item (such as a table or image) exists in a given section, a Y marker is assigned.

• Example:

  [Y: 300]: #

3. Split Marker

• Format: [Split]: #

• Purpose: Defines segmentation points in the document for later processing.

• Usage:

  • The split markers guide document chunking for downstream applications.
  • Users can manually adjust split markers in the Markdown Plus file before re-uploading.

• Example:

  [Split]: #

Metadata Header

Each annotated Markdown file contains a header section with metadata about the document, including:

1. Author
2. Title
3. Description
4. Filename
5. Extension
6. Number of Pages
7. Version

For example:

---
Author: jkunkel1
Title: Title of the file
Description: ''
Filename: file name
Extension: pdf
Number of Pages: 20
Version: 1.0
---

User Interaction with Annotated Markdown

1. Download Markdown Plus: Users can export the annotated Markdown file for review.
2. Modify Split Markers: If desired, users can manually edit [SPLIT]: # markers to customize segmentation.
3. Re-Upload Modified File: The system verifies that the paging structure remains undisturbed before processing the document further.

Public Arcana links

Here is a list of public links for Arcanas. These are sorted in categories and contain public material. Please refer to the How to use section of you need help working with these.

• GWDG Services
  • Contains knowledge about the GWDG Services and Documentation.
  • Example question: Welche Dienste könnte ich als Forscher bei der GWDG nutzen?
• Institute for Computer Science
  • Contain knowledge about the computer science study track.
  • Example question: Wie funktioniert der Studiengang angewandte Informatik?

In case you find a broken link, please let us know via our support contacts.

Support and FAQ

If you run into problem using this service please contact our KISSKI support

We also need your help providing publicly available Arcana links. If you have one that you think is relevant to a larger group of users, please also reach out to us so we can publish it.

FAQ

Which models support RAG (Retrieval-Augmented Generation)?

All internal models work with the GWDG-Tools and Arcanas. But some models are better at using Tools then others!

We recommend using: Qwen 3 30B A3B Instruct 2507 for the best results with tool calling and document retrieval.

We do NOT recommend to use: Meta Llama 3.1 8B Instruct as its is not optimized for Tool-Calling!

What is the maximum file size I can upload?

You can upload files up to 300 MB per file. For optimal performance, we recommend keeping files under 50 MB - larger files may take longer to process.

Which file types are supported?

The following formats are fully supported:

• Text: .txt
• Markdown: .md
• Word Documents: .docx, .dotx, .docm, .dotm
• PowerPoint Presentations: .pptx, .potx, .ppsx, .pptm, .potm, .ppsm
• PDFs: .pdf
• HTML: .html, .htm, .xhtml

All files are converted into a machine-readable (Markdown) format during indexing.

Is my data stored? Where is it stored?

Your documents and indexed data remain exclusively on GWDG systems at all times. We do not store, share, or transmit your data to third parties. Your privacy and data security are our top priorities.

Can I share my Arcana?

Yes - Arcanas can be freely shared with other users. To share, simply provide the Arcana ID or the full Arcana name.

How do I create a private Arcana?

To keep your Arcana private, enable the “Secure Arcana Name” option when creating it. This adds a random, unique security token to the Arcana name, making it accessible only to you or anyone who knows the full, exact name.

How does RAG (Retrieval-Augmented Generation) work?

RAG enhances large language models (LLMs) by combining document retrieval with generative AI.

Here’s how it works:

1. Your query is matched against a knowledge base (your uploaded documents).
2. The system retrieves the most relevant passages (5 Passages).
3. The LLM uses this context to generate accurate, informed, and contextually relevant responses.

This allows the model to draw from external knowledge beyond its training data, resulting in:

• Higher factual accuracy
• More coherent and relevant answers
• Reduced hallucinations

How are my requests processed?

Every step of your request is handled entirely on GWDG infrastructure:

• Document embedding generation
• Retrieval of relevant content
• Final inference and response generation

Your data and queries never leave GWDG systems, and are never shared with third parties.

Why is the indexing taking so long?

Indexing involves advanced machine learning algorithms to convert complex documents (like .docx, .pdf, .pptx) into a structured, machine-readable format (Markdown).

This process can take time, especially for:

• Large files (>50 MB)
• Documents with complex layouts, images, or tables

To speed up indexing, keep file sizes under 50 MB and avoid overly complex formatting.

Further Questions?

Let us know if you have any further questions!

Available Models

Chat AI provides a large assortment of state-of-the-art open-weight Large Language Models (LLMs) which are hosted on our platform with the highest standards of data protection. The data sent to these models, including the prompts and message contents, are never stored at any location on our systems. Additionally, Chat AI offers models hosted externally such as Anthropic Claude and OpenAI GPT-5.

Available models are regularly upgraded as newer, more capable ones are released. We select models to include in our services based on user demand, cost, and performance across various benchmarks, such as HumanEval, MATH, HellaSwag, MMLU, etc. Certain models are more capable at specific tasks and with specific settings, which are described below to the best of our knowledge.

List of open-weight models, hosted by GWDG

List of external models, hosted by external providers

Open-weight models, hosted by GWDG

The models listed in this section are hosted on our platform with the highest standards of data protection. The data sent to these models, including the prompts and message contents, are never stored at any location on our systems.

Apertus 70B Instruct

Apertus is a fully open language model designed to push the boundaries of transparent and compliant AI. It supports over 1,800 languages and a context window size of up to 65,536 tokens, using only fully compliant and open training data. The model achieves comparable performance to closed-source models while respecting opt-out consent of data owners. It was pretrained on 15T tokens with a staged curriculum of web, code, and math data.

DeepSeek V4 Flash 0731

DeepSeek V4 Flash 0731 is a strong Mixture-of-Experts (MoE) model with 304B parameters and a context window size of one million tokens. It outperforms DeepSeek V4 Pro (Preview) on several benchmarks despite its smaller activated parameter count, and is suitable for complex agentic workflows.

Devstral 2 123B Instruct 2512

Developed by mistralai, Devstral 2 is an agentic LLM designed for software engineering and coding tasks. It is capable of exploring codebases, working with multiple files, and powering software engineering agents.

Google Gemma 4 31B Instruct

Gemma 4 models offer frontier-level performance, well-suited for reasoning, agentic workflows, coding, and multimodal understanding.

GLM-4.7

GLM-4.7 is a coding-focused model that delivers significant improvements over its predecessor in multilingual agentic coding and terminal-based tasks. It achieves strong performance on SWE-bench, SWE-bench Multilingual, and Terminal Bench 2.0. GLM-4.7 also excels at tool use, web browsing, and mathematical reasoning, with notable gains on benchmarks like HLE and τ²-Bench.

GLM 5.3 Flash

GLM-5.3-Flash is the first natively multimodal model in the GLM-5 series. It is a hybrid-architecture MoE model with 320B total parameters and just 18B active, combining sparse and linear attention with Manifold-Constrained Hyper-Connections (mHC) to deliver precise long-context capabilities at significantly reduced serving cost. It outperforms GLM-5.2 across benchmarks and real-world workloads while approaching Claude Opus 4.8 on coding and agentic benchmarks.

Meta Llama 3.1 8B Instruct

The standard model we recommend. It is the most lightweight with the fastest performance and good results across all benchmarks. It is sufficient for general conversations and assistance.

Mistral Medium 3.5 128B

Mistral Medium 3.5 128B is a dense model with 128B parameters and a 256k context window, combining instruction-following, reasoning, and coding capabilities in a single model. Therefore, this model replaces its predecessors in Le Chat and the coding agent Vibe, and produces better results for a variety of tasks including instruct, reasoning, and coding, compared to previous models.

OpenAI GPT OSS 120B

In August 2025, OpenAI released the gpt-oss model series, consisting of two open-weight LLMs that are optimized for faster inference with state-of-the-art performance across many domains, including reasoning and tool use. According to OpenAI, the gpt-oss-120b model achieves near-parity with OpenAI o4-mini on core reasoning benchmarks.

Qwen 3 30B A3B Instruct 2507

This MoE model features 30.5B total parameters with 3.3B activated parameters for efficient inference. It delivers significant improvements in instruction following, logical reasoning, text comprehension, mathematics, science, coding, and tool usage, with better alignment for subjective and open-ended tasks. The model supports a 256K native context length and operates in non-thinking mode, achieving strong performance across knowledge, reasoning, coding, and multilingual benchmarks.

Qwen 3 Coder Next

Qwen 3 Coder Next has 80 billion total parameters and 3 billion activated parameters which makes it cost-effective, especially for agent deployment. With a context length of 256K and adaptability to various templates, it can also be integrated in CLI and IDE platforms.

Qwen 3 Omni 30B A3B Instruct

Qwen3 Omni is a natively multilingual omni-modal foundation model that processes text, images, audio, and video. It achieves state-of-the-art performance on many audio/video benchmarks with ASR, audio understanding, and voice conversation performance comparable to Gemini 2.5 Pro. The model features a novel MoE-based Thinker–Talker architecture with AuT pretraining, supports 119 text languages, 19 speech input languages, and enables low-latency interaction with flexible control via system prompts.

Qwen 3.5 122B A10B

Qwen 3.5 122B A10B is a powerful language model developed by Alibaba Cloud. With 122 billion parameters it delivers strong performance across reasoning, coding, and general tasks. The model supports vision capabilities for multimodal applications.

Qwen 3.5 397B A17B

Qwen 3.5 397B A17B is a MoE model with 397 billion total parameters and 17 billion activated parameters. It represents one of the most powerful open-weight models available, delivering exceptional performance across reasoning, coding, mathematics, and general tasks. The model supports vision capabilities, and provides state-of-the-art performance among open models.

Qwen 3.6 35B A3B

Qwen 3.6 35B A3B is an MoE model with 35 billion total parameters and 3 billion activated parameters for efficient inference. It is suitable for real-world tasks, including coding and agentic workflows.

Qwen 3.8 27B

Qwen 3.8 delivers substantial gains compared to its predecessors across research, coding, and large-scale agentic tasks. Qwen 3.8 is a native vision-language model that understands images and videos, with flexible thinking control, and is designed to complete complex tasks with greater reliability.

External models, hosted by external providers

Anthropic Claude Sonnet 5

Claude Sonnet 5 is built as an agentic model. It provides substantially improved cost efficiency at medium effort; its higher-effort performance can match Opus 4.8 on some tasks.

Anthropic Claude Sonnet 4.6

Claude Sonnet 4.6 provides an excellent balance between capability and speed. It is optimized for production workloads and delivers fast, reliable responses for coding, analysis, and conversational tasks.

OpenAI GPT-5, 5.1, 5.2, 5.3, 5.4, 5.5, and 5.6 Series

OpenAI’s GPT-5 series models achieve state-of-the-art performance across various benchmarks. The series consists of the following models along with their intended use cases:

• OpenAI GPT-5.6 Sol: The frontier model of the GPT-5.6 family, built for complex professional work.
• OpenAI GPT-5.6 Terra: The GPT-5.6 model that balances intelligence and cost.
• OpenAI GPT-5.6 Luna: The GPT-5.6 model optimized for cost-sensitive workloads.
• OpenAI GPT-5.5: Built for the most complex professional work, including coding use cases, tool-heavy agents, grounded assistants, and customer-facing workflows where execution quality and response polish are important.
• OpenAI GPT-5.4: Built for professional knowledge work, including document and spreadsheet tasks, coding, data analysis, agentic workflows, and software automation.
• OpenAI GPT-5.4 Mini: A lightweight variant of GPT-5.4 for cost-sensitive applications.
• OpenAI GPT-5.4 Nano: A highly optimized variant of GPT-5.4. Ideal for applications requiring low latency.
• OpenAI GPT-5/5.1/5.2: Designed for logic-heavy and multi-step tasks.
• OpenAI GPT-5 Mini: A lightweight variant of GPT-5 for cost-sensitive applications.
• OpenAI GPT-5 Nano: A highly optimized variant of GPT-5.

OpenAI GPT-4.1

OpenAI’s GPT-4.1-class models improve on the older GPT-4 series. These models also outperform GPT-4o and GPT-4o Mini, especially in coding and instruction following. They have a large context window size of 1M tokens, with improved long-context comprehension, and an updated knowledge cutoff of June 2024.

OpenAI GPT-4.1 Mini

This was developed as a more cost-effective and faster alternative to GPT-4.1.

OpenAI o1 and o1 Mini

OpenAI’s o1-class models were developed to perform complex reasoning tasks. These models have now been superceded by the o3-series, and are therefore no longer recommended.

OpenAI o3

Released in April 2025, OpenAI’s o3-class models were developed to perform complex reasoning tasks across the domains of coding, math, science, visual perception, and more. These models have an iterative thought process, and therefore take their time to process internally before responding to the user. The thought process for o3 models are not shown to the user.

OpenAI o3 Mini

This was developed as a more cost-effective and faster alternative to o3.

Chat AI FAQ

Data Privacy

Are my conversations or usage data used for AI training or similar purposes?

No, whether you use internal or external models, your conversations and data are not used to train any AI models.

When using internal models, are my messages and conversations stored on your servers at any stage?

No, user messages and AI responses are not stored at any stage on our servers. Once your message is sent and you receive the response, the conversation is only available in your browser.

What data does Chat AI keep when I access the service?

We do not keep any conversations or messages on our servers. We only record some usage stastistics, in order to monitor the load on our service and improve the user experience. This includes usernames, timestamps, and the models/services that were requested. Everything else the ChatBot remembers (like the history of your last conversation, etc.) is only stored locally in your browser.

When using external models, are my messages and conversations stored on Microsoft’s servers at any stage?

While we do not keep any conversations or messages on our servers, Microsoft retains the right to store messages/conversations for up to 30 days in order to prevent abuse. Since the request is sent directly from GWDG’s servers, no user information is included in the requests to Microsoft. For more information, see: learn.microsoft.com/en-us/legal/cognitive-services/openai/data-privacy

Availability

My institution is interested in using Chat AI. Can we advertise it to our users? Would you be able to handle an additional load for XXX users?

For large institutions, please contact us directly at info@kisski.de

Are all the models available with Chat-AI for free?

All models accessible to a user with an Academic Cloud account are for free, with the exception of the OpenAI GPT… (external) models. These models are only available to the Max Planck Society, members of the University of Göttingen, and via paid contract.

Usage

Why is model xxx taking longer than usual to respond?

There can be multiple reasons for this.

1. Most likely, your conversation history became larger over time and you didn’t clear it. Note that each time you send a message, the entire conversation history has to be processed by the model, which means a longer conversation history takes longer to process, and also uses more input tokens.
2. If the model responds slowly even when the conversation is empty, it could be due to high load, esp. during peak hours, or an issue with the hardware running the model on our infrastructure. You can wait a little or switch to a different model and see if the response time improves. Feel free to reach out to support if the problem persists.
3. Check your internet connection. It’s possible that this is caused by a slow or high-latency connection, esp. if you notice no difference when changing the model.

Can Chat AI process my images?

Yes, as long as the model supports it. Simply select a model that supports image input, as illustrated with the camera icon, then attach an image using the picture button in the prompt textbox. Note that some models may not support attaching more than one image at a time.

Can Chat AI process my PDF files?

Yes! Simply use the “attach text” button in the prompt textbox and select your PDF file. You will see the file in your attachments list as well as a “process” button. Note that PDF files must be processed before you can send a message to the model. Depending on the size and contents of your PDF file, this may take a while, or even fail if the file is too large. Once the file is processed, you can simply send a message to the model and its contents will be attached to your message.

OpenAI models

Can I use my own system prompts with the OpenAI (external) models?

No, sorry. The system prompt used by the OpenAI models can’t be changed by end users. Please use our internal models if you need to set custom system prompts.

Why is o1 and o1-mini slower / why can’t I get responses from o1 and o1-mini?

The o1 and o1-mini models have internal reasoning, meaning they need much more time to process a request. Furthermore, Microsoft’s API does not support streaming for these models yet, therefore Chat AI has to wait until the entire response is generated by the model before any data is received. In some cases, especially when there is a long conversation history, this can take so long that the connection times out and the request fails with a “Service Unavailable” error.

Are the OpenAI GPT… (external) models the real ChatGPT/GPT-4/… by OpenAI?

They are. We have signed a contract with Microsoft to be able to access the models running in their Azure cloud. Since the service costs money, it is only available to users from the Max Planck Society, members of the University of Göttingen (both GWDG shareholders), as well as organisations and groups with paid contract. Thank you for your understanding.

Why does GPT-4 refer to itself as GPT-3 when I ask what model it is?

This is a known issue when using GPT-4 via it’s API, see: community.openai.com/t/gpt-4-through-api-says-its-gpt-3/286881 Nevertheless, the model is in fact GPT-4, even if it states otherwise.

Features

This section collects all functionality that extends and customizes Chat AI beyond basic text generation. Features let you shape the assistant’s personality, integrate external knowledge, and configure model behavior for different use cases. The Version history is also tracked here, showing which features were added or updated in each release.

Available Features

• Memory Memories are pieces of information that the chatbot can learn from conversation to behave in a more personalized way.
• Personas Define roles and tones for the assistant (e.g., interviewer, tutor, casual style). Personas make it easy to quickly switch the model’s behavior.
• Arcana / Retrieval-Augmented Generation (RAG) Connect your own data sources (documents, notes, datasets) to ground responses in factual context.
• Tools Chat AI models are given access to a variety of tools to accomplish non text-based tasks or improve responses.
  • Web search
  • Image Generation
  • Image Modification
  • Text to speech (tts)
• MCP Support You can add custom public Model Context Protocol (MCP) servers to Chat AI.

Code Completion

Many code editors feature LLM integration these days. Code completion tools provide inline suggestions as you type, helping you write code faster by predicting the next line. They help by completing function signatures and suggesting entire code blocks based on context. They also offer a chat interface for explaining, generating, and editing code directly in your editor.

Contents

• Continue in Visual Studio Code (or Jetbrains)
• Zed
• Emacs

Continue

Setup

Visual Studio Code (or Jetbrains) is required as your IDE (Integrated Development Environment) to use CoCo AI.

Install the Continue extension from the VS Code Extension Marketplace. Continue is an open-source AI code assistant plugin that can query code snippets or even entire repositories using a selected model. Upon installation, Continue provides a short and easy-to-follow introduction to its features.

To use this service, you need to follow a few setup steps. After installing the Continue extension, click on its icon in the left sidebar of VS Code. This will open a new window. In the top-right corner of the chat box, you will see a “Local config” dropdown list. Click on it, and you should see a list of configuration files in YAML format (by default, there is only one file named “Local config”). Open the config.yaml file and paste the following configuration:

name: Chat AI
version: 1.0.0
schema: v1
models:
  - name: GLM-4.7
    provider: openai
    model: glm-4.7
    apiBase: https://chat-ai.academiccloud.de/v1
    apiKey: <api_key>
    defaultCompletionOptions:
      temperature: 0.2
      topP: 0.1
    roles:
      - autocomplete
      - chat
      - edit

Then, reload the window. After that, you should see a list of available models at the bottom bar of the chat box. By selecting a model, you can start using the code completion service.

Please note the following points when editing the config.yaml file:

• The “provider”, regardless of the actual provider, must be one that is supported by Continue (such as OpenAI or Ollama).

• The “model” name should be written entirely in lowercase letters.

• The “apiKey” should be written without any parentheses or quotation marks.

• The “roles” tells Continue what each model may be used for (chat, autocomplete, edit, etc.). At least one model must advertise autocomplete if you want the Tab-completion feature later.

For more detailed information about configuration in the Continue AI code agent, you can refer to their official documentation: docs.continue.dev

Basic configuration

Two important concepts to understand among completion options is temperature and top_P sampling.

• temperature is a slider from 0 to 2 adjusting the creativity, with closer to 0 being more predictable and closer to 2 being more creative. It does this by expanding or flattening the probabilities of the next token (response building block).
• top_p is a slider from 0 to 1 which adjusts the total population of probabilities considered for the next token. A top_p of 0.1 would only mean the top 10 percent of cumulative probabilities is considered. Variating top_p has a similar effect on predictability and creativity as temperature, with larger values considered to increase creativity.

Predictable results, such as for coding, require low values for both parameters, and creative results, such as for brainstorming, require high values. See the table in the current models section for value suggestions.

Our suggestion is to set the above completion options for each model according to the table in Chat AI and switch between the models based on your needs. You can also store the model multiple times with different completion options and different names to refer to, such as below.

- name: Creative writing model
  provider: openai
  model: gemma-3-27b-it
  apiBase: https://chat-ai.academiccloud.de/v1
  apiKey: <api_key>
  defaultCompletionOptions:
    temperature: 0.7
    topP: 0.8
  roles: [chat]

- name: Accurate code model
  provider: openai
  model: glm-4.7
  apiBase: https://chat-ai.academiccloud.de/v1
  apiKey: <api_key>
  defaultCompletionOptions:
    temperature: 0.2
    topP: 0.1
  roles: [chat, autocomplete]

- name: Exploratory code model
  provider: openai
  model: glm-4.7
  apiBase: https://chat-ai.academiccloud.de/v1
  apiKey: <api_key>
  defaultCompletionOptions:
    temperature: 0.6
    topP: 0.7
  roles: [chat, autocomplete]

Another completion option to consider setting, particularly for long responses, is max_tokens. It is a value less than the context-window that specifies how many tokens may be considered per prompt, plus generated for the response to that prompt. Each model has a different context-window size (see the table in current models for sizes). Similarly, each model has a default max_tokens length. This length is optimal for most tasks, but could be changed for longer tasks, such as “Name the capital of each country in the world and one interesting aspect about it”.

The context-window wants to look at system info, chat history, its training memory, the last prompt and the previous tokens from the current response. Therefore max_tokens limits the response generation in order not to risk degenerating the quality of the response by allocating context away from the other context sources. This is why it is recommended to split a large task into smaller tasks for the requirement of smaller response generation. It could be that this is difficult or unachievable however, in which case the max_tokens would be preferred to be increased (with the risk of degradation). See API Use Cases for an example of how to change max_tokens.

Further configuration options can be found at the Continue configuration page.

Functionality

The three main abilities of the Continue plugin is to analyse code, generate code and resolve errors. A new useful ability is a tab autocomplete key shortcut that is still in Beta.

Analyse code

For all below commands, Ctrl can be substituted with Cmd for Mac users.

Highlight a code snippet and press the command Ctrl + L . This will open the Continue side bar with the snippet as context for the question of your choice. From this side bar you can also access any file in your repository, as well as provide different types of context, such as entire package or language documentations, problems, git, terminal, or even your entire codebase in your chosen repository. This can be done either by pressing @ or clicking the + Add Context button. Typical functionality is provided, such as opening multiple sessions, retrieving previous sessions and toggling full screen. Models can be changed easily to, say, a model with a creative configuration, to which prompts without context can be sent, the same way the web interface of Chat AI works.

Generating code

Highlight a code snippet and press the command Ctrl + I . This will open a dropdown bar where your prompt of choice about this code can entered. When entered, this will generate further code based on the code snippet, or edit the code snippet itself. These edits can range from correcting faulty code, generating in-line documentation, renaming functions, etc. The generated code and potentially deleted code will show in a format reminiscent of a git merge conflict with Accept and Reject options. Bear in mind that there is no clear indication within VSCode whether cluster resources are available for code generation or whether code generation is not being performed actually for some other reason. We suggest to wait a short moment before trying again.

Before code generation with prompt

After code generation

Notice from the example that the code completion model is capable of more than just generating what seems like functional code. It also has all the benefits and knowledge that is expected from an LLM: semantics, grouping and linguistic reasoning. There are still limitations to this knowledge based on the date until which model training was performed.

Resolve errors

If errors have been encountered in your VS Code Problems, Output or Terminal, press Ctrl + Shift + R to place the errors in-context in the Continue side bar and prompt a solution for them. The results of this would explain the errors in great detail and possibly provide solution code for the identified faulty code. The same could be done manually from the Continue sidebar by providing the error as context and requesting it to be fixed.

Tab Autocomplete

Continue repetitively analyses the other code in your current file, regardless of programming language, and provides suggestions for code to fill in. To enable this function, ensure that at least one model in config.yaml includes roles: [autocomplete]. A common pattern is to duplicate the GLM4.7 entry:

- name: GWDG Code Completion
  provider: openai
  model: glm-4.7
  apiBase: https://chat-ai.academiccloud.de/v1
  apiKey: "<api_key>"
  defaultCompletionOptions:
    temperature: 0.2
    topP: 0.1
  roles: [autocomplete]

If the model selected is not a model particularly well-trained for code completion, Continue will prompt you accordingly. Now you should receive code suggestions from the selected model and be able to insert the suggested code simply by pressing Tab , much like the functionality of the default code suggestions VS Code provides when inspecting the packages loaded. Both suggestions could appear simultaneously, for which pressing Tab would prioritise the VS Code functionality over Continue. It may happen also that there is a conflict of hotkeys between tabAutocomplete and tab spacing, in which case the tab spacing hotkey needs to be disabled or remapped in your VS Code settings. In Settings, go to Keyboard Shortcuts and search ’tab’, then disable or replace the keybinding of the tab command. You can disable tabAutocomplete with the commands Ctrl + K Ctrl + A . Unfortunately there is no way to change the keybind of tabAutocomplete.

It is also possible to step through an autocompletion suggestion word-by-word by typing Ctrl + Right . Note that Ctrl + Left does NOT undo any steps. The below code example was almost entirely generated with tabAutocomplete only from initially typing def plus, beside the need for correcting some indentation.

Partial Tab Autocompletion

Full Tab Autocompletion

More information about tabAutocomplete, including further configuration options, can be found at the Continue documentation.

Zed

Zed is a popular VSCode competitor with builtin AI integration. Since the Chat AI API is OpenAI compatible, we follow Zed’s documentation on that. Your settings.json should look similar to the following:

{
  "language_models": {
    "openai": {
      "api_url": "https://chat-ai.academiccloud.de/v1",
      "available_models": [
        { "name": "glm-4.7", "max_tokens": 200000 },
        { "name": "devstral-2-123b-instruct-2512", "max_tokens": 256000 },
        { "name": "gemma-3-27b-it", "max_tokens": 128000 }
      ],
      "version": "1"
    }
  }
}

The model names are taken from here, the context sizes from here. Your API Key is configured via the UI. In the command palette open agent: open settings. Set your API Key in the dialog for “OpenAI”. For Zed’s AI editing functionality, check out their documentation.

MCP

The Model Context Protocol (MCP) is a common interface for LLM interfaces to call tools and receive additional context. Zed has builtin support for running MCP servers and letting LLMs call the exposed tools via the OpenAI API tool call requests automatically. Here is an example configuration to add a local MCP server to get you started:

{
  "context_servers": {
    "tool-server": {
      "command": {
        "path": "~/dev/tool-server/tool-server",
        "args": [
          "--transport",
          "stdio"
        ],
        "env": null
      },
      "settings": {}
    }
  }
}

Emacs

Emacs is an extensible, customizable, free/libre text editor and more.

With the help of gptel, a simple Large Language Model client, we make use of LLMs from within Emacs. gptel is available on MELPA and NonGNU-devel ELPA.

As SAIA implements the OpenAI API standard, the configuration is straight forward.

(setq gptel-model 'qwen3-30b-a3b-instruct-2507
      gptel-backend
      (gptel-make-openai "gwdg"
        :host "chat-ai.academiccloud.de"
        :endpoint "/v1/chat/completions"
        :stream t
        :key gptel-api-key
        :models '(gemma-3-27b-it
                  devstral-2-123b-instruct-2512
                  glm-4.7
                  qwen3-coder-30b-a3b-instruct
                  qwen3-30b-a3b-instruct-2507
         )
)

The SAIA API key is stored in ~/.authinfo.

machine chat-ai.academiccloud.de login apikey password <api_key>

Now you can interact with LLMs from within Emacs.

Agentic coding

Agentic coding goes beyond chat and inline suggestions. These tools leverage the model’s ability to autonomously complete multi-step tasks. It’s able to read your codebase, edit files, and run terminal commands.

Contents

• Model Selection
• Continue
• OpenCode
• Claude Code (Secure Setup)
• Cline

Model Selection

For agentic coding workloads, we recommend starting with models that have agentic capabilities. You can select from the full list of available models.

Look for models that are optimized for:

• Agentic coding: Multi-file editing, codebase exploration
• Tool use: Terminal commands, web browsing, API interactions
• Reasoning: Complex problem-solving and planning

When selecting a model, check that it mentions agentic tasks or overall performance in its advantages section or in description. For complex agentic workflows, consider models with larger context windows (200K+ tokens) to handle extensive codebases and longer prompts effectively.

Continue (Agent Mode)

Continue is primarily a code completion and chat tool, see the Code Completion section for full setup and installation. In addition to code completion, Continue also supports agentic workflows via its built-in Agent mode, where the model can autonomously read files, apply edits, and complete multi-step tasks.

To use Continue for agentic tasks:

1. Open the Continue panel in VS Code
2. Switch the mode dropdown in the continue panel from Chat to Agent
3. Send a task, for example, ask it to refactor a function or add error handling. Continue will propose the changes and ask you to accept or reject them.

OpenCode

OpenCode is an open-source, terminal-first AI coding agent that enables developers to generate code, modify files, execute commands, and automate development tasks directly from the terminal. OpenCode supports both self-hosted and commercial AI models through provider API keys, making it suitable for teams and developers who want to leverage state-of-the-art models while maintaining a terminal-centric workflow. Its design is particularly well suited for SSH sessions, remote HPC environments, CI/CD pipelines, and keyboard-driven development workflows. This section explains how to configure OpenCode with your SAIA API key. For commercial models, see the commercial models guide.

Install

# npm
npm install -g opencode-ai

# or with curl
curl -fsSL https://opencode.ai/install | bash

After installation, add OpenCode to your PATH:

echo 'export PATH="$HOME/.opencode/bin:$PATH"' >> ~/.bashrc

Open a new terminal and verify the installation:

opencode --version

Set your API key

Set your SAIA API key as an environment variable:

echo 'export SAIA_API_KEY="your-saia-api-key"' >> ~/.bashrc

Run source ~/.bashrc or open a new terminal for the changes to take effect.

Create the project config

Create opencode.json in your project directory:

{
  "$schema": "https://opencode.ai/config.json",
  "provider": {
    "saia": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "SAIA",
      "options": {
        "baseURL": "https://chat-ai.academiccloud.de/v1",
        "apiKey": "{env:SAIA_API_KEY}"
      },
      "models": {
        "qwen3-coder-30b-a3b-instruct": {
          "name": "Qwen3 Coder 30B (SAIA)"
        },
        "devstral-2-123b-instruct-2512": {
          "name": "Devstral 2 123B (SAIA)"
        }
      }
    }
  }
}

You can modify the opencode.json to choose any model of your choice and ensure the model id matches exactly with the SAIA available models.

Launch

# Interactive TUI
opencode

# One-off task example
opencode run "Write unit tests for all functions in src/utils.py"

Models

Modes

• Plan mode: Create structured plans for complex multi-step tasks, review and adjust steps before execution
• Build mode: Execute approved plans, run automation, and build workflows

Basic Commands

Adding Context

During a Task

For more information, see the official documentation.

OpenCode inside VS Code

If you prefer to stay in your editor, the OpenCode VS Code extension runs the same OpenCode agent. It uses the same opencode.json and SAIA configuration from the steps above, so there’s nothing extra to set up.

1. Open the Extensions view ( Ctrl + Shift + X on Linux, Command + Shift + X on macOS).
2. Search for OpenCode and install it.
3. Reload VS Code. OpenCode opens in a panel alongside your editor, where you can run tasks the same way as in the terminal.

Optional: SAIA model plugin

Instead of listing models by hand in opencode.json, you can install a plugin that keeps your SAIA model list up to date automatically. On every OpenCode launch, it fetches the current model list from the SAIA API in the background and merges it into your global config, so new models appear without you editing anything.

It is available here: opencode-saia-plugin.

Make sure SAIA_API_KEY is set (see Set your API key), then follow the setup steps in the repository. It will fetch the live model list from the SAIA API and generate an opencode.json with all available models, categorized (reasoning, coder, vision, agentic, etc.)

OpenCode in a sandboxed container (recommended)

For local development we recommend running OpenCode inside a container. The container mounts only your current project directory, so the agent can read and edit your project but cannot run commands that affect files anywhere else on your system. This is the recommended, secure way to use OpenCode locally.

You build an image once, then reuse it from any project. The image is generic. It is not tied to a specific project, and it won’t interfere with a Dockerfile your project may already have.

Prerequisites: Docker or Podman installed on your system.

One-time setup

You only do these four steps once.

1. Create a file named Dockerfile (capital D, no extension) in a dedicated folder of your choice (create one with mkdir opencode-sandbox). This is the standard name that docker build uses by default. Using a different name like dockerfile (lowercase) or Dockerfile.txt can cause the build to not find it, so stick with Dockerfile. Paste the following code to your Dockerfile.

FROM node:20-slim

# Install OpenCode
RUN npm install -g opencode-ai

WORKDIR /workspace
CMD ["opencode"]

If your projects need a particular language runtime (for example Python, or Python with uv), add it to this Dockerfile before building. The container only has what you install here.

2. Build the image. From inside the folder that contains the Dockerfile, run:

cd opencode-sandbox      # the folder where you saved the Dockerfile
docker build -t opencode-sandbox .

The . means “build using the Dockerfile in the current folder”, so you don’t have to type the path. (Replace docker with podman if you use Podman.) You only repeat this step if you change the Dockerfile.

Confirm the image was created:

docker images | grep opencode-sandbox

3. Add the run-opencode alias. The alias must be written into your shell config file (typing it once at the prompt only lasts for that terminal). Here’s the one-step version that removes any old copies of aliases and adds the clean alias to ~/.bashrc in a single command:

sed -i '/alias run-opencode=/d' ~/.bashrc && cat >> ~/.bashrc << 'EOF'
alias run-opencode='docker run -it --rm --user "$(id -u):$(id -g)" -e HOME=/tmp -e SAIA_API_KEY="$SAIA_API_KEY" -v "$PWD":/workspace -w /workspace opencode-sandbox'
EOF

4. Reload your shell so the alias becomes available:

source ~/.bashrc      # or: source ~/.zshrc

Verify the alias is registered:

alias run-opencode

This should print the alias definition back to you. If it does, the setup is complete.

Using it in a project

This is what you do every time you want to work on a project. Note that you run run-opencode from inside the project you want to work on not from the sandbox folder where you built the image.

1. Make sure your API key is set in the current shell (see Set your API key):

echo $SAIA_API_KEY      # should print your key, not be empty

2. Go to the project you want to work on, and make sure it contains an opencode.json (the SAIA config from Create the project config):

cd ~/path/to/your/project

Place opencode.json in the root of the project (the folder you run run-opencode from). The container mounts this folder as its working directory, so OpenCode picks the config up automatically. A global/home config on your host is not visible inside the sandbox, so each project needs its own.

3. Start the sandbox:

run-opencode

This launches the container with the current folder mounted and starts OpenCode. The agent can read and edit the files in this folder, but nothing outside it.

Troubleshooting

run-opencode: command not found : the alias isn’t loaded in your shell. Check whether it made it into your config file:

grep -n run-opencode ~/.bashrc      # or ~/.zshrc

• If it prints nothing, the alias was never added to the file (it was probably only typed at the prompt). Repeat step 3 of the one-time setup, then source the file again.
• If it prints the alias but the command still isn’t found, confirm you are editing the file your shell actually reads: run echo $SHELL, and use ~/.zshrc for zsh or ~/.bashrc for bash. If OpenCode starts but can’t reach a model: then the key isn’t reaching the container, or the project has no valid config. Confirm echo $SAIA_API_KEY prints your key, and that the project directory contains a valid opencode.json.

Claude Code (Secure Setup)

Claude Code is Anthropic’s terminal-first coding agent, also available as a VS Code extension. This setup runs it on the models in our cluster with just a SAIA API key, so your code and prompts stay on GWDG infrastructure.

Before you start

You need two things:

1. A SAIA API key. If you don’t have one yet, request one here.
2. Claude Code installed. Installation is the same as for the commercial setup, so follow the existing instructions: install the CLI or install the VS Code extension.

Only the configuration differs, and that is what the next section covers.

Configuration

Create or update ~/.claude/settings.json (Windows: %USERPROFILE%\.claude\settings.json), paste the block below, and replace <your SAIA API key> with your own key:

{
  "env": {
    "ANTHROPIC_BASE_URL": "https://chat-ai.academiccloud.de",
    "ANTHROPIC_AUTH_TOKEN": "<your SAIA API key>",
    "ANTHROPIC_MODEL": "qwen3-coder-next",
    "ANTHROPIC_DEFAULT_HAIKU_MODEL": "qwen3.6-27b",
    "ANTHROPIC_DEFAULT_OPUS_MODEL": "qwen3.5-397b-a17b",
    "ANTHROPIC_DEFAULT_SONNET_MODEL": "qwen3-coder-next",
    "CLAUDE_CODE_MAX_CONTEXT_TOKENS": "131072",
    "CLAUDE_CODE_DISABLE_ADAPTIVE_THINKING": "1",
    "CLAUDE_CODE_DISABLE_EXPERIMENTAL_BETAS": "1",
    "CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC": "1",
    "DISABLE_TELEMETRY": "1",
    "DISABLE_ERROR_REPORTING": "1",
    "DISABLE_BUG_COMMAND": "1",
    "DISABLE_AUTOUPDATER": "1"
  },
  "permissions": {
    "deny": [
      "Read(./.env)",
      "Read(./.env.*)",
      "Read(~/.ssh/**)",
      "Read(~/.aws/**)",
      "Bash(git push:*)"
    ]
  }
}

Start a session with the claude command, or open the Claude sidebar in VS Code. Because the settings file lives in your user profile, these values apply to every Claude Code session regardless of the shell it is started from. Verifying the configuration works.

If you use the VS Code extension, additionally enable the user setting Claude Code: Disable Login Prompt so the extension does not ask for an Anthropic login. You can also set it directly in your VS Code user settings.json:

{
  "claudeCode.disableLoginPrompt": true
}

What the variables do

Deliberately set a small model for background tasks. This noticeably reduces token consumption without affecting result quality. Optionally, CLAUDE_CODE_SUBAGENT_MODEL in the same env block sets the model used for subagents. Setting the reasoning level (“effort”) has no effect with the internal models.

A project-level .claude/settings.json in the repository applies in addition and is a convenient way to ship a preconfigured setup to a team. The API key does not belong in that file, since it gets committed. Put only the remaining variables there.

Switching models

There are three ways to choose a model. Way 1 is enough in most cases.

Way 1: type the model name directly (no configuration)

Typed in the running session, nothing is written to a file:

/model qwen3.5-397b-a17b

At startup, correspondingly:

claude --model qwen3.5-397b-a17b

Every model on the cluster is reachable this way, even when it does not appear in the picker list. You only need to know the model name.

Way 2: put three models into the picker list

The /model command opens a list containing three fixed slots labelled “opus”, “sonnet”, and “haiku”. These three slots can be filled with SAIA models by adding the following to the env block in ~/.claude/settings.json:

"ANTHROPIC_DEFAULT_OPUS_MODEL": "qwen3.5-397b-a17b",
"ANTHROPIC_DEFAULT_SONNET_MODEL": "qwen3-coder-next",
"ANTHROPIC_DEFAULT_HAIKU_MODEL": "qwen3.6-27b"

The list then offers three SAIA models. The labels “opus”, “sonnet”, and “haiku” stay as they are; the models behind them are the ones you entered. This way offers no more than these three slots.

Way 3: put all models into the picker list.

This is optional and not officially supported. Two different files are involved:

• Prerequisite: "CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC": "1" in the env block in ~/.claude/settings.json. Without it, Claude Code overwrites the model list on every start.
• Model list: at the top level of ~/.claude.json, a different file. Extend it with the key below while no session is running, do not replace the file:

"additionalModelOptionsCache": [
  {"value":"qwen3-coder-next","label":"Qwen3 Coder Next","description":"Coding"},
  {"value":"glm-4.7","label":"GLM 4.7","description":"Coding alternative"},
  {"value":"qwen3.5-397b-a17b","label":"Qwen3.5 397B A17B","description":"Large, vision"},
  {"value":"qwen3.5-122b-a10b","label":"Qwen3.5 122B A10B","description":"Mid-size, vision"},
  {"value":"qwen3.6-35b-a3b","label":"Qwen3.6 35B A3B","description":"Fast, vision"},
  {"value":"qwen3.6-27b","label":"Qwen3.6 27B","description":"Small, background tasks"},
  {"value":"qwen3-30b-a3b-instruct-2507","label":"Qwen3 30B A3B","description":"Small, general purpose"},
  {"value":"qwen3-omni-30b-a3b-instruct","label":"Qwen3 Omni 30B A3B","description":"Multimodal"},
  {"value":"deepseek-v4-flash-0731","label":"DeepSeek V4 Flash","description":"Fast, general purpose"},
  {"value":"openai-gpt-oss-120b","label":"GPT-OSS 120B","description":"General purpose"},
  {"value":"mistral-medium-3.5-128b","label":"Mistral Medium 3.5","description":"General purpose, vision"},
  {"value":"gemma-4-31b-it","label":"Gemma 4 31B IT","description":"General purpose, vision"},
  {"value":"apertus-70b-instruct-2509","label":"Apertus 70B","description":"General purpose"},
  {"value":"meta-llama-3.1-8b-instruct","label":"Llama 3.1 8B","description":"Very small"}
]

value is the model ID; label and description are free display text. The built-in entries (Default, Opus, Sonnet) stay visible but are not usable through SAIA. The embedding models are deliberately omitted, as is devstral-2-123b-instruct-2512, which rejects requests from Claude Code. Running /logout deletes the entry.

Permissions and sandboxing

Claude Code has file access and executes commands, so the same recommendation as for OpenCode applies: run it in a sandboxed container that mounts only your current project directory. The container is the outer security boundary; the permission system is the second layer, and it should be used rather than bypassed.

• Keep the permission prompt enabled. Avoid --dangerously-skip-permissions outside a container sandbox, and even inside one it only makes sense for tasks with no network and no access to secrets.
• Exclude access to secrets and unrelated directories with deny rules.
• Deny rules can also enforce that merges and deployments stay manual.

The deny rules go into ~/.claude/settings.json as a second block next to env, as shown in the configuration above. That block can be extended, for example to cut off outbound web access:

{
  "permissions": {
    "deny": [
      "Read(./.env)",
      "Read(./.env.*)",
      "Read(~/.ssh/**)",
      "Read(~/.aws/**)",
      "Bash(git push:*)",
      "WebFetch",
      "WebSearch"
    ]
  }
}

WebFetch and WebSearch can be omitted where your network rules already block outbound traffic; the deny rules make the restriction effective independently of the container configuration. See permission modes for the full permission model.

Verifying the configuration

Before starting, check the URL and key independently of the tool:

curl -X POST "https://chat-ai.academiccloud.de/v1/messages" \
  -H "Authorization: Bearer <your SAIA API key>" \
  -H "anthropic-version: 2023-06-01" \
  -H "content-type: application/json" \
  -d '{"model":"qwen3-coder-next","max_tokens":10,
       "messages":[{"role":"user","content":"Test"}]}'

A response containing "type":"message" confirms access.

Inside the session, run /status and check two lines:

• Anthropic base URL must show https://chat-ai.academiccloud.de. If the line is missing, the variable did not reach the session and the Anthropic API is still being used.
• Auth token must name ANTHROPIC_AUTH_TOKEN. If a Login method line with a personal account appears instead, the SAIA key is not active.

Limitations

Claude Code is designed for Anthropic models, and the vendor explicitly does not support operation with other models. On larger tasks, expect weaker results than with the native models, particularly for tool use such as reading and editing files or running commands. OpenCode is provider-agnostic by design and is the better starting point if you run into such issues.

Cline

Cline is an open-source coding agent that combines large-language-model reasoning with practical developer workflows. This section outlines Cline’s main benefits, explains its Plan → Act interface, and walks through an installation that connects Cline to SAIA models.

The Plan → Act Loop

• Plan mode: You can describe a goal, such as “add OAuth2 login”. Cline replies with a numbered plan outlining file edits and commands.

• Review: Edit the checklist or ask Cline to refine it. Nothing changes in the workspace until you approve.

• Act mode: Cline executes each step: editing files, running commands, and showing differences. You confirm or reject actions in real time.

This separation gives the agent autonomy without removing human oversight.

Installation Guide (VS Code)

Please find the installation steps below:

• Prerequisites

  • Visual Studio Code (v1.93 or newer)
  • SAIA API key
  • Node 18+ for optional CLI use

• Extension installation

  • Search Cline in VScode marketplace and install it.

• Connecting to CoCo AI

  • Open Cline (Command Palette → “Cline: Open in New Tab”).

  • Click the Setup with own API Key and choose “OpenAI Compatible”.

  • Fill the fields:

    Field	Value
    Base URL	https://chat-ai.academiccloud.de/v1
    API Key	your SAIA API key
    Model ID	glm-4.7 (add others as needed)

  • Add additional models (e.g., qwen3-30b-a3b-instruct-2507) with the same URL and key if required.

  • Assign roles (if you want a different model for plan and act): For example, set glm-4.7 for Act; set qwen3-30b-a3b-instruct-2507 for Plan.

Daily Workflow

Here is the daily workflow:

Plan → Approve plan → Act → Review differences → Iterate

Cline bridges the gap between chat-based assistants and full IDE automation. With a short setup that points to CoCo AI, it becomes a flexible co-developer for complex codebases while preserving developer’s control.

Agentic Coding

This section documents agentic coding tools that can be used with the GWDG commercial model offering.

For Codex and Claude Code, you need commercial access details from support@gwdg.de. For OpenCode, both setups are possible:

• with a SAIA API key

• with keys for commercial models

• Codex

• Claude Code

• OpenCode {class=“children children-type-tree children-sort-”}

Image-AI FAQ

Data Privacy

Are my prompts or usage data used for AI training or similar purposes?

No, your prompts and data are not used to train any AI models.

Are my prompts and generated images stored on your servers at any stage?

The user prompt is only stored on the GWDG server during the inference process itself. After the end of a session in the browser, the user’s entries are no longer available. The generated images are not stored.

What data does Image AI keep when I use the service?

A log is kept which contains the number of requests per user and the respective time stamps. The logs are stored for one year in accordance with GWDG guidelines. The collection of data for the provision of the website and the storage of the data in log files is absolutely necessary for the operation of the website. Consequently, there is no possibility for the user to object.

Availability

My institution is interested in using Image AI. Can we advertise it to our users? Would you be able to handle an additional load for XXX users?

For large institutions, please contact us directly at info@kisski.de.

Are Image AI services for free?

Image AI services that are accessible to a user with an Academic Cloud account are for free.

Protein-AI FAQ

Data Privacy

Are my protein sequences or usage data used for AI training or similar purposes?

No, your protein sequences and prediction data are not used to train any AI models.

Are my sequences and predicted structures stored on your servers at any stage?

User protein sequences and predicted structures are stored temporarily (for 30 days) on the GWDG server. During this period, only the respective user has access to their data. At no point do we access it on our servers without user permission.

What data does Protein AI keep when I use the service?

We do keep protein sequences and predicted structures on our GWDG server for 30 days. We record some usage statistics to monitor the load on our service and improve the user experience. This includes usernames, timestamps, and the services that were requested.

Availability

My institution is interested in using Protein AI. Can we advertise it to our users? Would you be able to handle an additional load for XXX users?

For large institutions, please contact us directly at info@kisski.de.

Are Protein AI services free?

Protein AI services that are accessible to a user with an Academic Cloud account are for free.

Voice-AI FAQ

How to use Voice-live tool effectively?

For Best Accuracy, Select Your input Language! While the Auto Detect feature is powerful, the transcription model achieves the highest accuracy and speed when you specify the spoken language beforehand. If you know what language will be spoken, selecting it from the dropdown is highly recommended.

How to do transcription/translation in an online meeting or generally from a system audio?

For security and privacy reasons, web browsers cannot directly “listen” to your computer’s speaker output. To transcribe a meeting (from Zoom, Teams, etc.) or any other audio playing on your computer, you need to use a Virtual Audio Cable. This free software creates a virtual “loopback” device that routes your speaker audio to a virtual microphone, which you can then select in this web app.

Setup Instructions: Windows: Download and install VB-CABLE. In your meeting app’s audio settings, set the Speaker to CABLE Input. In this browser window, select CABLE Output as your microphone. macOS: Download and install BlackHole. In your Mac’s “Audio MIDI Setup,” create a “Multi-Output Device” that includes both your normal speakers and BlackHole. Set this as your system’s default output. In your meeting app, use default output. In this browser window, select BlackHole as your microphone. Linux: You can use PulseAudio Loopback Module. Load it by running the command: pactl load-module module-loopback. Use a tool like pavucontrol to route the audio from your meeting application’s monitor to your browser’s input.

Data Privacy

Are my conversations or usage data used for AI training or similar purposes?

No, your conversations and data are not used to train any AI models.

Are my audio files and conversations stored on your servers at any stage?

No, user audio files or BBB conversation are not stored at any stage on our servers. Once your audio is sent and you receive the response, the result is available in your browser. The results of the transcription/translation are saved on our data mover node which will be erased after 30 days. This gives the users the opportunity to download their results whithin 30 days. BBB transcriptions, which are written in Etherpad and use local MySQL to save data also will be kept for 30 days before they are erased (future service).

What data does Voice AI keep when I use the service?

We do not keep any conversations or audio files on our servers. Transcrition/translation results are stored in our data mover node for 30 days. BBB transcriptions, which are stored in local MySQL also will be erased after 30 days (Future service). We also record some usage statistics to monitor the load on our service and improve the user experience. This includes usernames, timestamps, and the services that were requested.

Converting Larger Audio Files to FLAC Format

If your file is too large, you can use tools like FFmpeg or Audacity to convert it to FLAC, which is a lossless format. For example, using FFmpeg:

ffmpeg -i input.wav -vn -acodec flac output.flac

Availability

My institution is interested in using Voice AI. Can we advertise it to our users? Would you be able to handle an additional load for XXX users?

For large institutions, please contact us directly at info@kisski.de.

Are Voice AI services for free?

Voice AI services that are accessible to a user with an Academic Cloud account are for free.