cellmap_flow.dashboard.routes.finetune.ai_annotate

AI-assisted annotation: a hosted model’s first guess at one plane, for the user to review.

The user hovers a point in the viewer and presses Shift+G (or clicks “Annotate at view centre”). A background job reads one 2D plane of raw EM around the point, in the plane the user is looking at (XY, XZ or YZ), asks the configured model to paint the target structure in a colour, turns the image it returns into a mask and stages it beside the session’s corrections (ai_annotate.staging). Nothing is written to the annotation volume until the user accepts the preview.

GET /api/finetune/ai-annotate/config: whether the feature is on (and if not, why and how to turn it on), the providers and models the server’s config allows with where each sends the data, the day’s calls against the limit, the organelle catalog with each one’s editable prompt, and the session’s settings and acknowledgements.
POST /api/finetune/ai-annotate/settings: the provider, model, target and prompt to use; acknowledge: true records that the user accepts that the selected provider’s destination receives this dataset’s images. Also binds Shift+G in the viewer.
POST /api/finetune/ai-annotate/run: start a job at a point (by default the view centre, in the layout’s plane).
GET /api/finetune/ai-annotate/status: the job’s stage, or its preview (input | model output | overlay) once it is ready.
POST /api/finetune/ai-annotate/resend: ask the model again about the same staged plane, with an edited prompt.
POST /api/finetune/ai-annotate/accept: write the staged mask into the annotation volume, over unannotated voxels only unless overwrite says otherwise. The existing Undo (/api/finetune/view-labels/undo) takes it back, as it takes back a seed.
POST /api/finetune/ai-annotate/reject: throw the staged mask away.

The feature is off unless the server’s config file turns it on, and the browser only ever picks a provider and a model that the file lists: it never sends an endpoint or a credential. A run is refused until the session has acknowledged the provider’s destination for the dataset, and once the day’s call limit is reached. Error answers carry an AIAnnotateError’s short user_message, never an SDK’s own text, which may echo a request; anything unexpected is “AI annotation failed; see the server log”, with the detail logged after redaction. Every request, staging, failure, resend and decision is appended to the volume’s audit log (ai_annotate.audit).

One job runs per dashboard session, and a staged result must be accepted or rejected before the next run: two previews of overlapping planes would be confusing to review and could be accepted out of order.

Attributes

`logger`
`ACTION`
`KEY`
`KEYBINDING`
`STATUS_KEY`
`GENERIC_FAILURE`
`NOT_SENT`
`STAGE_LABELS`

Functions

`ai_annotate_config`()	What the page needs to show the AI-annotate panel; see the module docstring.
`ai_annotate_settings`()	Set the provider, model, target and prompt; optionally acknowledge the provider's destination.
`ai_annotate_run`()	Start a job: at `point_nm` (z, y, x) in the plane normal to `depth_axis`.
`ai_annotate_status`()	The session's job: its status and stage, its error, or (`?preview=1`) its preview once ready.
`ai_annotate_resend`()	Ask the model again about the staged plane, with `prompt` (the editable text).
`ai_annotate_accept`()	Write the staged mask into the annotation volume: each object an id of its own, the rest background.
`ai_annotate_reject`()	Throw the staged mask away. JSON body: `annotate_id`.

Module Contents

cellmap_flow.dashboard.routes.finetune.ai_annotate.logger

cellmap_flow.dashboard.routes.finetune.ai_annotate.ACTION = 'ai-annotate'

cellmap_flow.dashboard.routes.finetune.ai_annotate.KEY = 'shift+keyg'

cellmap_flow.dashboard.routes.finetune.ai_annotate.KEYBINDING = 'Shift+G'

cellmap_flow.dashboard.routes.finetune.ai_annotate.STATUS_KEY = 'ai_annotate'

cellmap_flow.dashboard.routes.finetune.ai_annotate.GENERIC_FAILURE = 'AI annotation failed; see the server log.'

cellmap_flow.dashboard.routes.finetune.ai_annotate.NOT_SENT = ('auth', 'config')

cellmap_flow.dashboard.routes.finetune.ai_annotate.STAGE_LABELS

cellmap_flow.dashboard.routes.finetune.ai_annotate.ai_annotate_config(): What the page needs to show the AI-annotate panel; see the module docstring.

cellmap_flow.dashboard.routes.finetune.ai_annotate.ai_annotate_settings()

Set the provider, model, target and prompt; optionally acknowledge the provider’s destination.

JSON body: provider and model (ones the config allows), label_key (an organelle catalog key) and/or label_name (what to call the target, by default the catalog’s name), prompt (the editable prompt; null or the catalog’s own text uses the catalog’s) and acknowledge.

cellmap_flow.dashboard.routes.finetune.ai_annotate.ai_annotate_run()

Start a job: at point_nm (z, y, x) in the plane normal to depth_axis.

Both optional: the view centre, and the plane of the viewer’s layout (XY in a multi-panel layout, where which panel is meant is not known).

cellmap_flow.dashboard.routes.finetune.ai_annotate.ai_annotate_status()

The session’s job: its status and stage, its error, or (?preview=1) its preview once ready.

The preview is three PNGs of up to 1024 pixels a side, so it comes only when asked for: the page asks once when a result becomes ready, not on every poll while it waits for review. A result being written by Accept is reported as ready, its last state the page can show.

cellmap_flow.dashboard.routes.finetune.ai_annotate.ai_annotate_resend()

Ask the model again about the staged plane, with prompt (the editable text).

The plane is not read again: the same image goes with the new prompt, and the preview is replaced under the same annotate_id.

cellmap_flow.dashboard.routes.finetune.ai_annotate.ai_annotate_accept()

Write the staged mask into the annotation volume: each object an id of its own, the rest background.

JSON body: annotate_id and overwrite. Without overwrite only unannotated voxels are written, as a seed writes them; with it, the staged labels replace what the plane held. Either way Undo takes it back.

cellmap_flow.dashboard.routes.finetune.ai_annotate.ai_annotate_reject(): Throw the staged mask away. JSON body: annotate_id.