The short version: We do not store your prompts or completions, and we do not train on them. Nothing you send is written to disk. We keep request metadata — token counts, timing, status — because we have to bill and keep the service up. Your request is processed in the region you called and does not leave it.
Scope
This policy covers InventDB Sparkle, an inference API that serves open-weight language models on our own engine and hardware. It applies whether you call Sparkle directly or reach it through a marketplace such as OpenRouter. It does not cover InventDB SOAR or InventDB Serverless, which are governed by our product Privacy Policy.
InventDB Software LLP is the controller for the metadata described below. Where you send us personal data inside a prompt, we act as a processor on your instructions, and those instructions are the ones set out here: run the inference, return the result, keep nothing.
What we receive
To serve a request we receive the prompt you send (text, and where you use the vision input, the images or video frames attached to it), the sampling parameters, and the API key presented. We receive nothing else. We do not receive, and cannot see, your end users' identities unless you place them in a prompt yourself.
What we do not keep
The engine runs with zero retention of request content enabled. In practice that means:
- Prompts and completions are never written to disk. They exist in the worker's memory for the life of the request and are gone when it ends.
- They are not written to logs. Content logging is compiled behind a switch that the zero-retention setting overrides; our production workers run with it on.
- They are not used for training, fine-tuning or evaluation, by us or by anyone else. We do not sell, share or licence them.
- No human reads them. We have no support workflow, no review queue and no debugging path that exposes request content.
The prompt cache, stated plainly
To make repeated calls fast, a worker may keep the computed state of a prompt prefix in GPU memory so an immediately following request that begins with the same tokens does not have to recompute it. This is worth understanding precisely:
- It lives in volatile GPU memory on one worker. It is never written to disk and never leaves that worker.
- It is destroyed when the worker stops, which happens automatically after a short idle period.
- A cache entry can only be matched by a request that supplies the identical tokens. It cannot be read, listed or extracted, and it cannot be used to discover content you did not already send.
- It stores computed numeric state, not your text.
What we log
We record request metadata, and only metadata: timestamp, model requested, count of input tokens, count of those served from cache, count of output tokens, latency, HTTP status, and the identifier of the API key used. We use it to bill accurately, to size capacity, and to detect abuse. We retain it for up to 12 months and then delete it.
Where your request runs
Sparkle runs on GPU workers rented from RunPod. We operate endpoints in two regions and you choose which you call:
- European Union / EEA — Iceland.
- United States — California.
A request is processed in the region whose endpoint you called and is not moved to another region. RunPod is our only sub-processor for inference and receives prompt content solely as the operator of the machine the model runs on; no other third party receives it. Model weights are held on storage attached to the region and never carry request content.
Security
Traffic is encrypted in transit with TLS. Requests are authenticated; an unauthenticated request is rejected at the edge before reaching the engine. Because no request content is persisted, there is no stored corpus of prompts to breach.
Your rights
Access, correction, deletion and portability rights attach to data we hold. For prompts and completions we hold nothing, so there is nothing to return or erase. For request metadata, write to us and we will act on the request. If you are an EU or UK data subject you may also complain to your supervisory authority.
Children
Sparkle is a developer service and is not directed at children. Do not send us personal data of children through it.
Changes
If we change this policy we will update the version and date above. A change that materially reduces the protections stated here will be announced before it takes effect.
Contact
Questions about this policy, or a data request: contact@inventdb.com.