GRID Docs | General Robotics

Embed text, images, or one interleaved text+image record into a shared space. Exactly one of three modes is selected by which arguments are present:

text only — one row per string.
image_input only — one row per image.
both — a single interleaved row; text must be one string whose <|image|> placeholders are filled from image_input in order.

Parameters

List[str] | None

Up to 64 UTF-8 strings. Required unless image_input is given.

Union[ImageInput, Sequence[ImageInput], None]

One image or up to 64 images (path, URL, PIL Image, or numpy array). Required unless text is given; capped at 8 in interleaved mode.

str | None

Task instruction prefix applied to text only — e.g. "SearchQuery" for queries and "Document" for corpus items. Omitting it still works but reduces retrieval precision.

int | None

Matryoshka output width: 768 (default), 512, 256 or 128. Rows are re-normalized after truncation, and queries must share a dimension with the corpus they are scored against.

float | None

Optional timeout in seconds for the HTTP request.

Returns

Dict[str, Any]: Dict with features of shape (N, truncate_dim), float32, L2-normalized per row (cosine similarity == dot product); mode ("text", "image" or "interleaved"); model_revision of the backing checkpoint; and normalized=True.

Example

from grid_cortex_client import CortexClient, ModelType
client = CortexClient()
query = client.run(
    ModelType.EMBEDDINGGEMMA2,
    text=["a red cup on a table"],
    prompt_name="SearchQuery",
)["features"]
corpus = client.run(
    ModelType.EMBEDDINGGEMMA2,
    image_input=["cup.jpg", "chair.jpg"],
)["features"]
int((query @ corpus.T)[0].argmax())   # cup.jpg

Overview

Cortex Models

embeddinggemma2

Parameters

Returns

Example

​Parameters

​Returns

​Example

Parameters

Returns

Example