Jev: Locate in image

Description

Experimental. TypeSafe’s images and box questions sit behind a “Danger-Zone” header (TypeSafe-Danger-Zone-Images): they may change without notice. The node sends that header for you.

Locates something in an image (for example “the invoice total” or “the logo”) and returns its box: coordinates from 0 to 1 relative to the image’s height and width, and also in pixels (the node reads the size from the PNG or JPEG header). With that you can crop the area, check that it exists or hand it to another node.

The image is sent to TypeSafe as an asset and the node references it with a placeholder, as TypeSafe documents. Only PNG and JPEG are supported; the format is deduced from the bytes, not the extension.

Configuration

Parameters

Parameter	Type	Required	Description
credentials_id	credentials	Yes	TypeSafe credential (API key).
model	select	Yes	Jev model. Images work best with `jev-preview` (the one TypeSafe’s documentation uses in its examples).
image	textarea	Yes	Image as base64 (the `data:image/png;base64,` prefix is accepted) or a variable, e.g. `\{\{image\}\}`. Up to 20 MB.
target	textarea	Yes	What to locate, e.g. “Locate the invoice total”. Describe it precisely; English works better.
output_key	text	No	Key in the data where the result is stored. Default `jev`.

Credentials

Requires a TypeSafe credential (credentials_id) holding the api_key from the TypeSafe console (TYPESAFE_API_KEY). The key is yours: TypeSafe bills the usage, not Floogos.

Output

The result is added to the input data. Read it as \{\{jev.box\}\}, \{\{jev.pixels\}\}…

{
  "nextModule": "next-node",
  "data": {
    "...": "input data, untouched",
    "jev": {
      "box": {
        "top": 0.1,
        "bottom": 0.5,
        "left": 0.2,
        "right": 0.6
      },
      "confidence": 0.9,
      "pixels": {
        "x": 200,
        "y": 50,
        "width": 400,
        "height": 200
      },
      "image": {
        "format": "png",
        "width": 1000,
        "height": 500,
        "bytes": 48213
      },
      "model": "jev-preview",
      "usage": {
        "input_tokens": 900,
        "output_tokens": 8
      },
      "cost_usd": 0.000038
    }
  }
}

A single output (nextModule).

Example

Check that a scanned invoice has a total

Download the image with an HTTP node, convert it to base64 and pass it in image. With target = Locate the invoice total, an If node on \{\{jev.confidence\}\} decides whether a total can be located (for example confidence above 0.7) or the invoice needs manual review.

Notes

Experimental: it depends on a TypeSafe “Danger-Zone” header. Do not use it in critical processes without a human-review path.
Privacy: the image leaves Floogos for TypeSafe, a third party. Do not send documents you cannot share (IDs, invoices with personal data…) without checking their terms.
Limits: PNG or JPEG, up to 20 MB (the API accepts a 32 MiB body and base64 weighs a third more). Coordinates are relative: top and bottom to the height, left and right to the width.
Variables: the engine substitutes \{\{image\}\} before running the node; the node does not substitute again.
Errors: missing_credentials, validation (invalid image or format, missing target), auth_error, rate_limited, overloaded, state_too_large (image too large), jev_error and connection_error. 429, 529, 5xx and network drops retry automatically.
Language: Jev performs best in English; other languages have lower accuracy.