Custom chatbots: Advanced functions & tips

If you want to adapt your own chatbot to specific requirements or use additional options, you will find tips and extensions here: These include various output formats (such as LaTeX, Markdown, diagrams or code), customizable role models, team functions and tips for use in multiple courses.

The chatbots can output various formatted content. To ensure that this formatting is rendered correctly, the bot should receive appropriate instructions in the system prompt.

LaTeX for mathematical formulas

Mathematical formulas can be output as inline LaTeX and are displayed rendered in the chat.

Example in the system prompt:

You can output mathematical formulas with LaTeX. Use `$...$` for inline formulas. Example: The quadratic formula is $x = \frac{-b \pm \sqrt{b^2 - 4ac}}{2a}$.

Examples: - Inline: $E = mc^2$ → E = mc² - Fractions: $\frac{1}{2}$ → ½ - Roots: $\sqrt{x^2 + y^2}$ → √(x² + y²) - Sums: $\sum_{i=1}^{n} x_i$ → Σ xᵢ

Markdown for text formatting

The chatbots can use Markdown syntax to display formatted text. To do this, it is sufficient to indicate in the system prompt to use Markdown for formatted output.

Example in the system prompt:

Use Markdown to format your answers.

Available formatting:

Bold text with **double asterisks**
Italic text with *single asterisks*
Bold & italic combined
Headings with # (from # to ###### for different levels)
Bulleted lists with - or *
Numbered lists with 1., 2., etc.
Tables with | column 1 | column 2 |
Quotes with >
Code with `Backticks` for inline code - Separators with --- or ***

Program code with syntax highlighting

Programming code can be highlighted in color (syntax highlighting) in different languages.

Example in the system prompt:

When outputting program code, use code blocks with language specification:
```python
def hello():
print("Hello, World!")

console.log("Hello, World!");

**Supported languages (examples):**
- `python` - Python code
- `javascript` - JavaScript code
- `java` - Java code
- `cpp` - C++ code
- `sql` - SQL queries
- `html` - HTML markup
- `css` - CSS styles
- `bash` - shell scripts
- `json` - JSON data

**Example output:**
```python
def fibonacci(n):
if n <= 1:
return n
return fibonacci(n-1) + fibonacci(n-2)

Diagrams with mermaid syntax

The chatbots can output different types of diagrams as mermaid code, which are rendered graphically.

Example in the system prompt:

You can create diagrams with mermaid syntax. Start with ```mermaid and follow the mermaid syntax.
Example:

graph TD
A[Start] --> B{Question?}
B -->|Yes| C[Action]
B -->|No| D[End]

Available diagram types:

Diagram type	Diagram type Description	Example use case
Flowchart	Flowcharts	Processes, decision paths
Sequence diagram	Sequence diagrams	Interactions between objects
Gantt diagram	Gantt diagrams	Project schedules
Class diagram	Class diagrams	UML class structures
Git graph	Git branches	Visualize version control
Entity Relationship	ER diagrams	Database relationships
User Journey	user journeys	Customer Journeys
Quadrant Chart	Quadrant charts	Prioritization matrices
XY Chart	XY charts	Function graphs

Summary of formatting in the system prompt

Here is an example section for the system prompt that covers all formatting

<formatting>
You can use different formatting:

1. **LaTeX for math**: Use $...$ for inline formulas
2. **Markdown**: Formatting with **bold**, *italics*, #headings, -lists, |tables |
3. **Code blocks**: With language specification such as ```python or ```javascript for syntax highlighting
4. **Mermaid diagrams**: For flowcharts, sequence diagrams, Gantt charts etc.
Start with ```mermaid and follow the mermaid syntax
</formatting>

The ILIAS plugin supports various forms of document integration.

Options at a glance

Option	Option Description	Suitable for
None	Only the system prompt is used	Simple FAQ bots, ethical coaches
Context window	Document text is sent with every request	Course materials, scripts, guidelines

Supported file formats

PDF (.pdf) - Recommended for scripts, papers, guidelines
Word (.docx) - For own documents, guidelines
Text (.txt) - For simple text files, FAQs

To upload documents

Open the settings of your assistant
Select "Context window" for document support
Switch to the "Documents" tab
Click on "Select file"
Select your file and click on "Save"

⚠️ Note: This refers specifically to variant 2 (ILIAS + 1 document). Variant 1 (AI toolbox full configuration) has extended options for multiple and larger documents.

Important restrictions for variant 2

Only one document can be uploaded
The document should cover a maximum of 30% of the context window
For larger documents or if you want to use multiple documents, switch to variant 1 (AI toolbox full configuration)

What are tokens?

A token corresponds to about 3/4 of a word or a character in longer words
A DIN A4 page of text corresponds to around 500-600 tokens
Context window is the total memory for system prompt + document + conversation

Recommended document sizes for variant 3 (sorted: local models first)

Model	Context window	Max. Document size (30%)
gpt-oss:120b (local)	64,000 tokens	~19,200 tokens (≈ 32-38 pages)
Qwen3.5-397b (local)	32,768+ tokens	~9,800+ tokens (≈ 16-20+ pages)
gpt-5-mini (Cloud)	128,000 tokens	~38,400 tokens (≈ 64-77 pages)
gpt-5-nano (Cloud)	128,000 tokens	~38,400 tokens (≈ 64-77 pages)

Model recommendations for the three variants:
💡 Important note: Start with smaller, resource-saving models for simple tasks! Local models (KIT infrastructure, data-sensitive content) remain within the KIT infrastructure and are preferable from a data protection perspective.

Variant	Recommended model	Justification
Variant 1 (AI toolbox full configuration)	gpt-oss:120b or Qwen3.5-397b (local, KIT infrastructure)	Local models preferred - resource-efficient, data protection-compliant, maximum flexibility
Variant 2 (ILIAS only)	gpt-oss:120b (local, KIT infrastructure) or gpt-5-nano (cloud)	For simple bots: local models or fastest cloud variant
Variant 3 (ILIAS + 1 document)	gpt-oss:120b (local, KIT infrastructure) or gpt-5-mini (cloud)	Local models for data protection; cloud for large context windows

💡 Tip: For data-sensitive applications, we recommend local models such as gpt-oss:120b or Qwen3.5-397b. These remain completely within the KIT infrastructure. A current list of all available models with detailed descriptions can be found on the KI-Toolbox website.

For larger documents: Use variant 1

If you:

want to provide several documents
Want to use larger documents (e.g. complete scripts)
Need central administration for several courses

→ Then we recommend variant 1 (AI toolbox full configuration). This offers significantly more comprehensive setting options and supports multiple documents as a knowledge base. To the instructions for variant 1

Custom chatbots can take on different roles.

Examples of role scenarios

Role	Role description	Field of application
Socratic tutor	Asks questions, does not give complete solutions	Exercise mode, comprehension questions
Expert	Subject-specific answers at level	Advanced courses
Peer coach	Collegial, supportive	Writing center, peer tutoring
Critical reflection partner	Questions answers, encourages reflection	Ethics seminars, AI competence
Simulated person	Role play (patient, customer, citizen)	Communication training

How to configure role-specific bots

The system prompt is decisive for the role:

<role>
You are [ROLE] for [TARGET GROUP] in the context [CONTEXT].
</role>

<context>
[Description of the target group and its needs]
</context>

<instructions>
[Specific behavioral instructions for this role]
</instructions>

<constraints>
[Constraints and no-goes for this role]
</constraints>

<tone>
[Tone and language style of the role]
</tone>

Instructions: System prompts

Frequently asked questions

Which model should I choose?

Prefer local models - these remain within the KIT infrastructure and are more resource-efficient:

Variant 1 (AI toolbox): gpt-oss:120b or Qwen3.5-397b (local, data protection compliant)
Variant 2 (ILIAS only): gpt-oss:120b (local) or gpt-5-nano (cloud, fast)
Variant 3 (ILIAS + 1 document): gpt-oss:120b (local) or gpt-5-mini (cloud, large context window)

For simple bots (FAQ, greeting), start with smaller models. Always choose local models for data-sensitive applications.

Why is the bot not responding as expected?

Check:

Is the system prompt clearly formulated?
Has the correct model been selected?
Are the documents up-to-date and relevant?

How can I make the bot "better"?

Iterative approach:

Test with typical student questions
Analyze the answers
Customize the system prompt
Repeat the test

Can several teachers supervise a bot?

Yes, all course admins can edit the bot. For institutional bots, we recommend variant 1 (AI toolbox) with central administration.

Troubleshooting

The problem	Possible cause	Solution
Bot does not respond	"Online" not activated	Activate "Online" in the settings
Incorrect model responses	Wrong model selected	Change model in configuration
Documents are ignored	Token limit exceeded	Reduce documents or use less
Bot outputs solutions	System prompt too lax	Sharpen the limits in the prompt
API error	API key invalid/expired	Generate new key in the AI toolbox
Slow responses	Large documents/models	Smaller documents, select faster model

Custom chatbots: Advanced functions & tips

Output formatting

Documents

Token management

Role-specific bots

Manage conversations

Multi-course use

FAQ & Troubleshooting

LaTeX for mathematical formulas

Markdown for text formatting

Program code with syntax highlighting

Diagrams with mermaid syntax

Summary of formatting in the system prompt

Examples of role scenarios

How to configure role-specific bots

For students and teachers (as users)

Storage configurable by teachers

Variant 1: Bot per course

Variant 2: Central bot (variant 1) + shared use

How to set up multi-course use

Frequently asked questions

Troubleshooting

Surname
Name
Email
Your message to us