Data and state

AI applications create conversation history, uploaded documents, generated artifacts, vector indexes, evaluation results, and prompt logs. Some of these stores begin as local files and enter production without a formal provisioning request.

Data stores

A primary database holds ordinary application data and should enter the provisioning process.

Object storage holds uploads and generated files. Hand-created buckets can inherit broad access when a builder needs a working URL quickly.

The vector index contains chunked copies of source documents and may store original text beside the embeddings. Include it in the data catalogue, classification process, and data access controls.

Session and conversation state may reside in a cache and contain full prompt and response history, including user-supplied data.

The prompt log is another data store with retention obligations. Apply the prompt-log access controls during collection.

Each production store needs an owner, classification, backup policy, and retention period. Provisioning should collect those decisions when the store is created.

Platform-provisioned stores

A store created manually can miss tags and backup policy in its reproducible configuration. A platform template can apply those defaults consistently.

Self-service provisioning gives teams a fast route to a store with the required defaults. The template applies the backup standard without relying on each requester to remember it.

Restore testing

Verify backups through scheduled restore tests.

Vector indexes may be derived artifacts, but rebuild time and embedding cost can make recovery slow. Conversation history may also be the record of what a system told a user, so its recovery requirements can differ from ordinary cache data.

Schedule restore drills with a named owner and record the result.

Deletion has to propagate

When a customer asks to be deleted, or a document is withdrawn, the record has to leave every copy: primary database, object storage, vector index, conversation history, prompt log, and any backup you keep long enough for it to matter.

Test vector indexes and prompt logs explicitly when they sit outside the normal deletion workflow. Document propagation time before a deletion request arrives.