How Kyndex works
Connect knowledge. Retrieve relevant evidence. Review the next action.
1. Prepare your knowledge
Documents are parsed, split into searchable passages and indexed with embeddings. Parsing and embedding providers are configured separately and may receive document content.
2. Ask in context
The signed-in user's access context and department scope shape retrieval. Relevant passages are supplied to the configured model, and source references help the user inspect the evidence.
3. Use configured capabilities
Agents can use tools and instructions made available to their department. Permitted read tools may execute directly. Write actions are proposed as pending approvals before execution.
4. Review decisions and outcomes
An authorised reviewer can approve or reject a proposed action. Approved work is executed through the action workflow, with status and audit records available for review.
Where data goes
Kyndex can run in your infrastructure while external providers handle authentication, storage, parsing, embeddings, reranking or generation. Tool calls can send data to connected systems too. A private deployment is not automatically an offline deployment.
Review each service's region, credentials, retention and network access. PII policies apply to supported processing paths and detected entities, not as a blanket guarantee over all outbound data.
What remains your responsibility
Choose appropriate source permissions, verify answers and proposed actions, configure infrastructure security, and test the deployment against your data policies. Citations and approval steps do not eliminate model or operator error.