AI Configuration
The AI Configuration section is where you set up and manage the AI assistants that power your portal's chat experience. This includes connecting to the AI service, training your assistants on your product content, and defining guardrails that shape how the assistant communicates.
AI Provider
Before the AI assistant can work, you need to choose on which AI provider it should be based. You can choose between OpenAI (American) and MistralAI (French).
Assistants
One Assistant Per Access Level
DistiForge creates a separate AI assistant for each confidentiality level you use in your portal. This ensures each assistant only has knowledge of content that users at that level are permitted to see.
For example, if your portal uses Public, Internal, and NDA levels:
The Public assistant knows only your public product content
The Internal assistant knows your internal content plus everything in Public
The NDA assistant knows everything - NDA content, internal content, and public content
When a user chats with the portal, they are automatically connected to the assistant matching their access level. This all happens behind the scenes. The experience is seamless for the user.
Training an Assistant
Before users can chat, each assistant needs to be trained on your published product content. You will need to create and train an assistant for each level you use.
To train an assistant:
Go to AI Configuration > Assistants.
Find the access level you want to set up (for example, "Public").
Click Create Assistant.
Training takes a few minutes. Progress is shown on screen. Once complete, the assistant is ready to answer questions.
Updating an Assistant
When you publish new products, update guardrails, or change your content, your assistant becomes outdated. You will see a warning indicator when an assistant is stale and needs to be updated.
To bring an assistant up to date:
Go to AI Configuration > Assistants.
Click Update Assistant.
This retrains the assistant on your latest content and guardrail settings.
Guardrails
Guardrails define how the AI assistant behaves: its personality, what it will and will not discuss, how it handles sensitive topics, and what disclosures it must always include. You do not need to touch code or write prompts; the guardrail settings in the admin UI are translated into instructions for the assistant automatically.
Guardrails are organized into four areas:
These settings control what the assistant can draw on when answering questions.
Web Search
Turn on to allow the assistant to look up live information from the web; turn off to keep it limited to your portal's content
Public Knowledge
Allow or prevent the assistant from drawing on broader knowledge beyond your documents
Competitor Mentions
How the assistant handles questions about competing products, for example, decline to compare, or redirect to your own offerings
Unreleased Products
Whether the assistant can discuss products not yet publicly available
Pricing Policy
Rules for handling pricing questions, such as "Do not quote prices; direct users to contact sales"
Scope Instructions
A free-text field for any additional topic limits specific to your business
These settings shape the assistant's communication style and identity.
Personality
The assistant's character, such as 'technical', 'professional'
Tone
The emotional register - formal, serious, playful, and so on
Brand Voice
The communication style, such as "professional and concise" or "friendly and approachable"
Language / Formality
The preferred language and formality level, such as "English, formal technical language"
Citation Policy
To what extent your assistant should include source citations in responses or not
These settings ensure the assistant always includes required disclosures.
Confidentiality Notice
Text the assistant appends when referencing restricted or NDA-level content
Legal Disclaimer
A standard disclaimer shown when the assistant discusses safety, regulatory, or liability-sensitive topics
Support Redirect
Instructions for how the assistant directs users to technical support when it cannot fully answer a question
Sesitive Content Policy
Guidelines for handling sensitive topics
Refusal Rules
Refuses certain requests to answer, ex.: Request to customer-specific data.
Fine-grained controls for what the assistant must never do.
Forbidden Phrases
Specific words or phrases the assistant must never use
Forbidden Topics
Topics the assistant must refuse to engage with entirely
Custom Rules
A free-text field for any additional rules not covered elsewhere
Saving Guardrails and Updating Assistants
Guardrail changes save immediately when you click Save in each section. After saving, your assistants will show a stale indicator, this is expected. The new guardrails have been saved but the assistants need to be updated to use them.
Go to AI Configuration > Assistants and click Update Assistant for each level to push the new settings. Until you do this, the assistants continue to use the previous guardrail configuration.
You do not need to retrain assistants from scratch when you change guardrails - updating is faster and reuses your existing content.
Last updated