Prompt Journey
Where can your AI request go? Follow it down to the model and back to your screen.
Choose a journey
Start with the sample. No private information needed.The round trip
Select any stop to inspect it.The map follows playback. You can also scroll it in any direction.
Read or jump through every step
What can remain after the answer?
Turn on “Show possible stored copies and later use” to add the dashed branches. A branch shows something to investigate, not proof that a copy exists.
- Conversation history: the app may save your messages, attachments and answers. Saved memory may add information to a later request.
- Logs and monitoring: the app, router, host and tools can each keep their own records. Those records might contain usage totals, content, or both. Backups and support access need their own checks.
- Caches: a saved answer can avoid a model call. A prompt-processing cache reuses earlier computation while a model still generates an answer. Neither is shown as a cache hit in this walkthrough.
- Training or evaluation: using information to improve or test a model is a separate policy decision. It is not an automatic step after every answer. “No training” and “no storage” promise different things.
This diagram does not assign storage periods or legal guarantees. Ask each operator what is stored, who can access it, where it is processed and when it is deleted.
Terms, scope and sources
Prompt: the instructions and context sent to a model. Your typed message may become one part of a larger package.
API: a way for one program to ask another program to do something. Inference: running a trained model to produce an output. Tokens: small pieces of text represented as numbers. Images and audio have other preparation steps.
Router: a service that chooses an eligible model host. Retrieval: finding information to include in a request. Agent: an application that can use model outputs to choose and carry out further steps, within the permissions it has been given.
The model developer created the model. The model host operates the computers running it. They may be the same company. A model name alone does not identify the host or processing region.
These are common logical components, not an inventory of every network device. One organization may operate several boxes. Real systems can combine or reorder checks, use several models, stream partial answers, or repeat the tool loop. Document indexing happens before retrieval; it is not drawn as part of every request. A tool that fetches a URL adds that destination to the route.
Encrypted internet links normally hide message content from routers carrying the traffic. A service that ends that encryption to process the message can read the content it receives. Company inspection proxies and device extensions can add access paths not pictured here.
- Microsoft: retrieval and document indexing.
- Anthropic: application-run and provider-run tools.
- OpenRouter: provider selection and fallback.
- OpenRouter: retention, training and tool scope.
- OpenAI: API data controls and application state.
These sources explain particular implementations. Their policies are not promises made by every AI service. Content reviewed November 21, 2026.