cellmap_flow.ai_annotate.backends.vertex_gemini

Gemini image models on Google Cloud Vertex AI, asked to paint the structure.

The plane is sent as a PNG at its own size (padded to a square when it is not one, see _square) with the recolor prompt; the model answers with an image in which the structure is painted the prompt’s colour on black, and the mask is the pixels of that colour. The plane is not upsampled first: gemini-3-pro-image charges the same ~560 tokens for an image whatever its size, so more pixels buy nothing, and the resolution the prompt states must be the one sent.

Authentication is Application Default Credentials (gcloud auth application-default login, or a service account on a cloud host), found by the SDK, or the credentials file the config names (credentials_file, read by secrets.load_google_credentials), so this backend holds no key. The google-genai SDK is imported only when the first call is made, so the dashboard runs without it when the feature is off.

Every call has a timeout. Only “try again later” answers (429, 503) are retried, at most MAX_ATTEMPTS times in all with bounded waits; anything else fails at once. Failures become AIAnnotateError with a message written here: the SDK’s text goes to the server log, not the browser.

Attributes

`logger`
`MAX_ATTEMPTS`
`RETRY_WAITS_S`
`RETRY_CODES`
`RETRY_STATUSES`
`MAX_REPLY_CHARS`

Classes

VertexGeminiBackend

Recolor-and-threshold segmentation through Vertex AI.

Module Contents

cellmap_flow.ai_annotate.backends.vertex_gemini.logger

cellmap_flow.ai_annotate.backends.vertex_gemini.MAX_ATTEMPTS = 3

cellmap_flow.ai_annotate.backends.vertex_gemini.RETRY_WAITS_S = (10.0, 30.0)

cellmap_flow.ai_annotate.backends.vertex_gemini.RETRY_CODES = (429, 503)

cellmap_flow.ai_annotate.backends.vertex_gemini.RETRY_STATUSES = ('UNAVAILABLE', 'RESOURCE_EXHAUSTED')

cellmap_flow.ai_annotate.backends.vertex_gemini.MAX_REPLY_CHARS = 200

class cellmap_flow.ai_annotate.backends.vertex_gemini.VertexGeminiBackend(project, location, timeout_s, sleep=time.sleep, client_factory=None, credentials_file=None)

Recolor-and-threshold segmentation through Vertex AI.

project may be None, in which case the SDK takes it from GOOGLE_CLOUD_PROJECT or the credentials. credentials_file may be None, in which case the SDK finds Application Default Credentials itself. sleep is how retries wait and client_factory (no arguments, returns an object with .models.generate_content) replaces the SDK client; both are for tests.

project

credentials_file = None

location

timeout_s

segment(request: cellmap_flow.ai_annotate.backends.base.SegmentRequest, model: str) → cellmap_flow.ai_annotate.backends.base.SegmentResult