AnacomAI for developers

Vision API

Vision is AnacomAI's own model for finding things in photos. Send a photo and say what you're looking for. You get back where it is: a box around each match, an outline if you ask for one, and any text read on it.

Get an API key Read the docs

Quickstart

curl https://api.anacomai.com/v1/search \
  -H "Authorization: Bearer $ANACOM_API_KEY" \
  -F image=@garage.jpg \
  -F query="Find the cordless drill"

POST /v1/search

Send a photo and what to find as multipart/form-data. The answer comes back in the same request, usually in a few seconds. Photos are processed in memory, never stored, never used for training, and never sent to anyone else.

image	JPEG, PNG, WebP or HEIC, up to 25 MB. Required.
query	What to find, in plain words, up to 500 characters. Required.
outlines	true to also get an outline around each match. Off by default because it's slower.

Response

{
  "request_id": "req_8fK2xQ...",
  "model": "engine-1",
  "query": "Find the cordless drill",
  "image": { "width": 4032, "height": 3024 },
  "summary": "Found 1 likely match.",
  "matches": [
    {
      "label": "cordless drill",
      "status": "likely_match",
      "box": { "x_min": 0.412, "y_min": 0.538, "x_max": 0.597, "y_max": 0.771 },
      "outline": null,
      "text": null
    }
  ],
  "warnings": [],
  "usage": {
    "plan": "starter",
    "searches_this_period": 41,
    "included": 2500,
    "duration_ms": 2800
  },
  "duration_ms": 2800
}

box and outline use 0–1 coordinates from the top-left of your photo, after any phone rotation is applied. Multiply by image.width and image.height for pixels.
status is likely_match or possible_match. Likely matches come first. There is no made-up confidence score.
matches empty and warnings empty means what you asked for isn't in the photo.
A question a photo can't answer (“which is cheapest?”) returns no matches and a warning saying why. It never guesses a box.
outline and text are null when there's nothing to show.
model says which AnacomAI model answered: engine-1 while Vision rolls out, then vision-1, vision-2 and so on. New models always get new names, so your results never change silently.

Retries and repeats

Send an Idempotency-Key header (any unique string, up to 255 characters) to retry safely. Retries with the same key get the first answer back and aren't counted again for 24 hours.
The same photo and query again within 10 minutes is answered from memory, almost instantly. It still counts as a search.

GET /v1/usage

Your plan, the current period, searches_this_period, included, overage_usd and requests_per_minute.

Errors

Errors look like {"error": {"code": "...", "message": "..."}}. Codes don't change. Failed requests are never counted.

401	missing_api_key / invalid_api_key	No key, a wrong key, or a revoked key.
402	subscription_required / spend_cap_reached	No active monthly plan, or your spending limit for extra searches is reached.
413	file_too_large	The image is over 25 MB.
415	invalid_image	The file isn't a JPEG, PNG, WebP or HEIC image.
422	validation_error / empty_query / idempotency_key_reused	The query is missing, blank or too long, or an Idempotency-Key was reused for a different request.
429	rate_limited	Too many requests this minute for your plan. Wait the number of seconds in the Retry-After header.
504	search_timeout	The search took too long. Try a smaller photo.

Limits

Requests a minute per key: 30 on Starter, 90 on Pro, 300 on Scale.
Searches beyond your plan are billed at the end of the month. You can set a spending limit for them under Your account.
Up to 10 keys per account.
Keep keys on your server. Don't put them in apps or web pages people can download.
/v1 only ever gains fields. Nothing in it is renamed or removed.