Photo checks¶
Photo checks tell you whether a piece of work was done right, from the photos your users take of it. Your app sends each photo as it is taken. Sight checks it against a template, the list of things that must be true for that kind of work, and answers with a verdict, every check's outcome, and what is still missing, in words you can show to the user.
graph LR
A[User takes a photo] --> B[Your app sends it to Sight];
B --> C[Model finds things, rules judge them];
C --> D[Verdict + what is still missing];
D --> E[Your app shows it, or Sight sends it to your app];
The words used¶
| Word | Meaning |
|---|---|
| Template | The checks for one kind of work, e.g. site-completion. Set up for your account; each version is frozen once written, so a past verdict can always be explained. |
| Job | One piece of work: a task, an order, a visit. You name it with your own id (jobId), and it is opened the first time you send a photo for it. |
| Photo | One photo in a job, named with your own id (photoId) so a resend is harmless and you can remove it later. |
| Slot | Which photo this is, as the template names it, e.g. front_view or meter_reading. |
| Check | One thing that must be true, e.g. "the meter shows a reading in range". Each has an outcome. |
| Verdict | The job as a whole: pass, fail, needs_review or incomplete. |
Outcomes¶
| Outcome | Meaning |
|---|---|
pass |
Seen and right. |
fail |
Seen and wrong. The check's ask says what the user should do about it. |
needs_review |
Could not be decided with enough confidence: a person should look. Its retake_helps says whether a new photo could help (then it has an ask) or only a person can settle it. |
not_assessable |
The photo was there but could not be used for this check, e.g. too small or too dark. |
missing |
No photo yet for this check. Becomes fail once the job is completed. |
pending |
Still being read. Not a call for a person: a later answer replaces it. |
not_applicable |
The check does not apply to this job: it uses job data your app did not send, and its template says to leave it out then. It does not count for the verdict, and the user is not asked anything. |
Each check carries its evidence: which photo (by your own photoId), which
box on it, and what was read, with the reading's source (your device or Sight).
How fast¶
A photo can be answered in three ways (see Sending photos):
- At once (
async, the default): the photo is taken and checked moments later. Your app gets the result by asking, or by having it sent to your app. - Fast (
fast): the rules and the model answer in the same call (about a second); checks waiting for a reading saypendinguntil the final answer. - Final (
sync): the call waits for everything, readings included.
Next¶
- Quick start: send your first photo with the SDK.
- Sending photos: modes, readings your device made, finding just the things you name.
- Results back to your app: webhooks, signatures, the photo with its boxes drawn on.
- HTTP API: every route, for any language.
- For AI agents: how an agent should use all of this.