Cognera is installed onto a machine rather than signed up for, so the setup is a real process with real decisions in it. These are the chapters of that process, in the order they happen.
None of this requires you to do it yourself — it is documented because you should know what is being done to your own equipment.
A private AI stack is only private if the box it runs on is. This comes before any service is switched on.
Services bind to the local machine rather than the open network. What reaches the outside world does so through a single reviewed tunnel, not through a port opened on your router.
Anything published externally sits behind an identity gate, so reaching it requires being on a list — not merely knowing the address.
API keys and tokens live in files readable only by the account that needs them, never in the project itself and never in a backup archive.
Each surface is brought up on its own and confirmed working before the next one is added, so a fault has only one possible cause.
The local model runtime is installed and the models you will actually use are pulled down and kept resident, so the first request of the day is not a cold start.
Your documents are collected, converted where needed, and indexed. This is the step that decides how well anything answers, so it is worth doing properly.
Website Studio, Media Studio, and the Voice Agent are enabled according to your plan, each verified against the running model before moving on.
Everything is registered to come back automatically after a reboot or a power cut. An installation that needs a person to restart it is not finished.
The one piece of maintenance that genuinely matters. Everything the agent says comes from here.
Adding material is a matter of putting files in the knowledge folder and re-running the indexer. The index is rebuilt locally and the service picks it up on restart.
The indexer reads Markdown, plain text, PDF, and Word files. Other formats — including source code and spreadsheets — need converting to one of those first, or they are silently skipped.
Plans differ in which surfaces are activated, not in how well they run. There is no throttled tier.