surface↓embedding↓image_embed
Embeddingscalar · returns json

IMAGE_EMBED

SigLIP 2 embedding for an image (L2-normalized, shared image/text space)

Per-row — runs once for each row.

embeddingembedding-modelspecialist-zooimage

Syntax

IMAGE_EMBED({{ image }})

Arguments

nametypedescription
imageVARCHAR—

About

Compute a SigLIP 2 embedding for an image column. Returns the raw 1152-dim float vector as JSON, in the SAME shared image/text space used by IMAGE_MATCHES, IMAGE_SIMILARITY, and the upcoming IMAGE_SEARCH table function. Vectors are L2-normalized at the gateway, so the dot product of two vectors from this operator (whether image–image or image–text) IS the cross-modal cosine similarity. Accepts any of: - http:// or https:// URL - Local path visible to the gateway - data:image/...;base64,... URI - Raw base64-encoded image bytes This is the low-level primitive. For most use cases you want IMAGE_MATCHES (bool) or IMAGE_SIMILARITY (float) which do the embed + compare in one gateway round trip. Use IMAGE_EMBED when you need the raw vector for storage, custom distance metrics, or feeding external vector indexes.

Nearby rabbit holes

same domain
scalar

EMBED

Generate 768-dim embedding vector from text (on-box nomic-embed-text-v1.5)

embeddingembedding-modelspecialist-zooscales-large+1
scalar

EMBED_BATCH

Batch embed rows and store in lars_embeddings

embeddingembedding-modelspecialist-zootext
scalar

EMBED_BATCH_ELASTIC

Batch embed rows and store in Elasticsearch for hybrid search

embeddingembedding-modelspecialist-zootext
scalar

EMBED_BATCH_PINECONE

Batch embed rows and store in Pinecone for vector search

embeddingembedding-modelspecialist-zooscales-large+1
scalar

EMBED_STATUS

Check embedding coverage for a table/column

embeddingembedding-modelspecialist-zootext
scalar

EMBED_WITH_STORAGE

Generate embedding and store with table/column/ID tracking

embeddingembedding-modelspecialist-zootext
Climb back to The Looking Glass