> ## Documentation Index
> Fetch the complete documentation index at: https://docs.cymph.io/llms.txt
> Use this file to discover all available pages before exploring further.

# AI & data usage

> Which Cymph features use AI, what data reaches a model, and how the provider is configured.

Cymph's AI features are additive: the platform is fully usable with AI disabled, and in a self-hosted deployment it is disabled until an administrator turns it on.

## Which features use AI

| Feature                                                                     | Uses an AI provider                                     |
| --------------------------------------------------------------------------- | ------------------------------------------------------- |
| Playbook generation from a prompt                                           | Yes                                                     |
| Mindmap generation                                                          | Yes                                                     |
| Autotagging                                                                 | Yes                                                     |
| Semantic search over playbooks                                              | **No** — embeddings are generated inside the deployment |
| Everything else — editing, execution, sharing, integrations, administration | No                                                      |

Semantic search is worth calling out because it looks like an AI feature and is often assumed to be one. It runs on a local embedding service (`all-MiniLM-L6-v2`) that ships as part of the deployment, so **playbook content is never sent to an external provider in order to be searchable**. Search works whether or not AI is enabled.

## Who provides the model

**You do — in both deployment models.** Cymph does not bundle, resell or operate a model on your behalf. An administrator enters a base URL, an API key and a model name, and Cymph calls that endpoint using an OpenAI-compatible API. See [AI configuration](/deployment/settings/ai).

Because the base URL is yours to choose, you can point Cymph at a commercial API, at a provider in a specific jurisdiction, or at a model hosted inside your own network. If you need AI processing to stay in a particular region for GDPR or data-residency reasons, you satisfy that by choosing a provider that meets the requirement — it is not a property of the Cymph deployment.

This has a direct consequence for your data-processing analysis: **the relationship for AI features is between you and the provider you select.** Cymph is not a processor in that path and cannot see the requests. This holds for a [managed cloud tenant](/deployment/managed-cloud-tenant) exactly as it does for a self-hosted deployment — the AI provider is a customer-configured setting in both.

## What reaches the model

Only the content needed for the feature you invoked:

* **Playbook generation** — the prompt you write, and the context of the playbook being generated.
* **Mindmap generation** — the prompt or source material you supply.
* **Autotagging** — the playbook content being tagged.

AI features are invoked explicitly by a user action. There is no background process that sends stored playbooks, execution records, assets or integration credentials to a model, and nothing is sent from a workspace nobody has asked the AI to act on.

## Provider credentials

The AI provider's API key is stored in the deployment's database encrypted with that deployment's own key material, and decrypted only in memory when a request is made. In a self-hosted deployment it never leaves your infrastructure. See [Data protection](/security/data-protection#encryption-at-rest).

## Turning AI off

Leave the AI settings unconfigured, or clear them. Without a base URL and API key the AI features are unavailable and no outbound calls are made — in either deployment model. Semantic search continues to work.

This is the default state of a new deployment: AI is off until an administrator configures a provider, so no content reaches a model unless someone has deliberately enabled it.