Embeddings VoyageAI Multimodal icon

Embeddings VoyageAI Multimodal

Generate multimodal embeddings from text and images using VoyageAI multimodal models

Embeddings VoyageAI Multimodal

AI-generated

Summary

Supplies a VoyageAI multimodal embeddings model (text, images, or both) to a connected vector store. It reads the incoming items, embeds them in a single VoyageAI API call, and exposes the vectors through the AI Embedding output.

Inputs

  • Model — voyage-multimodal-3.5 (default) or voyage-multimodal-3.
  • Content Type (required) — What to embed: Text Only, Image URL Only, Binary Image Only, Text + Image URL, or Text + Binary Image.
  • Text Input — Text to embed. Shown for text-based content types. It is evaluated per input item. It must be non-empty for Text Only, and in the combined modes an empty value means only the image is embedded.
  • Image URL (required) — Image URL (PNG, JPEG, WEBP, GIF). Shown for the URL content types. It is evaluated per item and must be non-empty.
  • Binary Data Key (required) — Name of the binary property holding the image on each input item (default 'data'). The MIME type must be image/*. Images over 20MB are rejected.
  • Options — Input Type (None, Query or Document) optimizes the embeddings. Truncation (default true) truncates over-length inputs. For query embeddings, Input Type falls back to 'query' when unset.

Output shape

An AI Embedding connection (a LangChain Embeddings object) for a vector store node. It is not regular item data. Each embedding is a numeric vector, one per input item. Text+image items are fused into a single vector.

Embeddings for all input items are computed up front in one multimodalEmbed request, with a maximum of 1,000 items. Later embedDocuments calls look up the cached vectors by matching document text. The cache keys are the text, '[image:URL]' or '[image:binary:key]'. If only one vector was cached, it is returned for any unmatched document. embedQuery makes a fresh text-only API call. Errors are raised for a missing text, URL or binary, a non-image MIME type, a size over 20MB, or an API failure. The credential's URL is used as the base URL, defaulting to https://api.voyageai.com/v1.

Examples

Example 1: Index product photos from URLs into a vector store alongside their descriptions.

Content Type = Text + Image URL; Text Input = {{ $json.description }}; Image URL = {{ $json.imageUrl }}; Options > Input Type = Document; connect to a Vector Store node.

Example 2: Embed images downloaded by an HTTP Request node.

Content Type = Binary Image Only; Binary Data Key = data; Model = voyage-multimodal-3.5; connect to a Vector Store node.

Example 3: Embed text snippets for multimodal search.

Content Type = Text Only; Text Input = {{ $json.text }}; Input Type = Document.

Links

Discussion