Organization Settings

Global settings are managed from Admin → Organization Settings. Settings are organized into tabs. Some tabs are only visible to super admins or when specific features are enabled.

System Prompt

The system prompt defines the chatbot's personality, behavior, and response style. It is prepended to every conversation.

Configuring the System Prompt

Writing an Effective System Prompt

A good system prompt should cover these areas:

• Knowledge boundaries — Define what the chatbot should and should not answer. For example, restrict it to only use information from the knowledge base.
• Greeting behavior — How the chatbot should start a conversation.
• Tone and style — Friendly, professional, formal, casual, etc.
• Response length — Short and concise, or detailed and thorough.
• Response format — Whether to use bullet points, paragraphs, numbered lists, etc.
• Fallback behavior — What to do when the chatbot cannot answer a question (e.g., refer users to a support email).
• Language — Whether to respond in a specific language or match the user's language.

Example System Prompt

**Personality traits:**

- Important - You are not allowed to respond with Your general LLM knowledge. You are allowed to only utilize information within knowledge base.
- Greet in the first message in the chat – "Hello!”. If you indicate that you need time to think, you do not need to greet once again when continuing to answer the question.
- You clearly explain what ... does and how it works, using only information available in the knowledge base.
- You stay friendly, encouraging, and professional at all times.
- When informing user of tool usage, do not respond with technical jargon. E.g., for semantic tool usage, simply advise that You are looking through the knowledge base.
- Keep responses warm and professional, but concise.
- Respond in one paragraph, no longer than five lines. Use simple, clear language that is easy to read and understand. Keep your tone friendly and approachable, making complex information digestible without losing accuracy.
- Bullet-proof formatting - use bullet points. Make scanning effortless.
- When presenting information that can be listed, format it as a bullet point list. Only add citations after the entire list, not after each bullet point.
- When answering a question You must start with the correct answer first and only then follow up with related information if applicable
- Do not include messages about asking the user to wait while searching a knowledge base—just provide the answer directly.
- Response length - short and concise.
- If you are unable to answer a related question, advise the user to contact ...@...

Theme

Customize the chatbot's appearance with a custom CSS theme based on Tailwind CSS and ShadCN UI.

Providers

Configure which AI models power your chatbot and how they behave.

API Keys

Provide API keys for the providers you want to use. Keys configured here override any environment variable defaults.

Models

Configure chat, embedding, and transcription models. Here you can see which model types each provider offers, to decide which API keys to configure and which defaults to pick in Chat Models.

Chat Models

These models power chat conversations and the default selections used across the app:

Each chat model has its own Reasoning Effort and Thinking Budget settings.

Processing Models

These models handle background document processing tasks:

Each processing model has its own Reasoning Effort and Thinking Budget settings.

Retry Settings

Observability

Monitor LLM interactions by connecting to Langfuse for trace collection and analysis. Observability settings are configured per organization under the Provider tab.

If organization-level keys are not configured, the platform falls back to the global environment variable defaults (LANGFUSE_PUBLIC_KEY, LANGFUSE_SECRET_KEY, LANGFUSE_HOST).

Traces are exported via OpenTelemetry, giving you detailed visibility into LLM calls, tool executions, and token usage.

Rate Limiting

Limit how many messages users can send within a time window to control API costs and prevent abuse.

File Attachments

By default, uploaded file attachments require authentication to access. You can make all file attachments publicly accessible by enabling the Make all file attachments publicly accessible option in the General tab.

Conversation Compaction

When a conversation grows long enough to approach the model's context window limit, the platform can automatically summarize older messages and replace them with a compact summary. This keeps the conversation going without losing the thread.

Compaction is configured per organization under the General tab:

Individual workspaces can override the global compaction settings. See Workspaces for details.

Data Retention

When the data retention feature is enabled, chat history is automatically deleted after the configured retention period.

Sandbox

When the sandbox is enabled for your organization, a Sandbox tab appears under Admin → Organization Settings:

Artifact Publishing

Publishing an artifact creates a link that anyone can open without signing in. Turn Artifact publishing off to remove that option from the whole organization: the publish action disappears everywhere, while links that are already public stay reachable and can still be revoked by their publisher or by an organization owner or admin.

See Sandboxes — Publishing an Artifact for what users see.

Network Policy

Enter a Network policy and click Save to restrict which destinations the sandbox may reach.

See Sandboxes — Network Policy for the policy format and examples.