iKB combines access rules, privacy checks and approved connections. Decision Records keep the sources, policy checks and model use behind answers for human review.
Choose who can access each topic, which actions can run and which models and external services are allowed.
Check the sources, policy events, model choices and recorded results behind an answer.
Set export permissions and retention periods to support your organisation's reviews and audits.
A Decision Record brings together the sources, policy checks and model use behind an answer. Review records from native chat (iKB's built-in interface), the web widget, Slack, Microsoft Teams, Chatwoot and API conversations. It does not expose a model's hidden reasoning.
Export Decision Records as JSON or PDF packages. Integrity checks help detect changes. Set access and retention rules. Conversation evidence packages support signed links that expire and audit records for security monitoring systems (SIEM).
Review each connection and its individual read actions, assign permitted tools to topics and retain execution records. Discovering a connector does not grant it permission to run.
Check document text and table cells for personal data. Add word lists and pattern checks to existing categories. Controls include personal information (PII) redaction in text and images, data loss prevention (DLP), data subject access requests (DSAR) and two-stage retention.
Check outgoing research queries against private document text. Block queries that fail the privacy checks. Topic settings control external sources and shared search budgets. External research needs network access and approved services.
Monitor evidence checks without changing answers, or enable filtering of clearly irrelevant material while keeping protected evidence. A low evidence score does not itself add an automatic refusal rule. Optional native-chat quality scores and checks after an answer inform administrators without rewriting the delivered answer.
Keep prompt versions and apply access rules, moderation and prompt-injection controls. Check and remove restricted content from outputs. Require permission for source-based inference, external tools and destinations. Linked audit events are append-only and hash-chained to help detect changes. They record configuration, sources, decisions and usage.
Keep feedback and correction learning within the relevant conversation. Topic controls and correction limits still apply.
Sign support-feedback links for each topic and set an expiry time. External support services can verify these links. Cloudflare Turnstile can protect configured sign-in, public-chat and embedded-widget forms.
Features depend on the installed version, topic permissions, platform settings, supported models and deployment. Test the workflow with your own sources before rollout.
Bring sample sources and clear requirements. Review the answers and controls before rollout.
