Multimodal Embeddings - Content API

The Content API provides a unified interface for embedding text, images, and other modalities. It introduces Content and Part as the canonical request types alongside portable Intent constants and per-request options like Dimension.

Quick Start

Embed a single piece of text using EmbedContent:

{% codetabs group="lang" %} {% codetab label="Go" %}

import (
    "context"

    "github.com/amikos-tech/chroma-go/pkg/embeddings"
)

content := embeddings.Content{
    Parts: []embeddings.Part{
        embeddings.NewTextPart("What is Chroma?"),
    },
}
embedding, err := ef.EmbedContent(ctx, content)

{% /codetab %}

The ef variable is any embeddings.ContentEmbeddingFunction — for example, a Roboflow embedding function (see the mixed-part section below for construction).

Mixed-Part Requests

Use EmbedContents to embed a batch of content items with different modalities. Because Roboflow processes one item per request, pass text and image as separate Content items:

{% codetabs group="lang" %} {% codetab label="Go" %}

import (
    "context"

    "github.com/amikos-tech/chroma-go/pkg/embeddings"
    "github.com/amikos-tech/chroma-go/pkg/embeddings/roboflow"
)

ef, err := roboflow.NewRoboflowEmbeddingFunction(roboflow.WithEnvAPIKey())

contents := []embeddings.Content{
    {Parts: []embeddings.Part{embeddings.NewTextPart("A dog running on a beach")}},
    {Parts: []embeddings.Part{
        embeddings.NewPartFromSource(
            embeddings.ModalityImage,
            embeddings.NewBinarySourceFromURL("https://example.com/dog.jpg"),
        ),
    }},
}
results, err := ef.EmbedContents(ctx, contents)

{% /codetab %}

Other binary source constructors are available for different input types:

• embeddings.NewBinarySourceFromFile("/path/to/image.png")
• embeddings.NewBinarySourceFromBase64("base64data==")
• embeddings.NewBinarySourceFromBytes(rawBytes)

Portable Intents

Set an Intent on a Content to communicate the purpose of the request to providers that support it:

{% codetabs group="lang" %} {% codetab label="Go" %}

content := embeddings.Content{
    Parts:  []embeddings.Part{embeddings.NewTextPart("retrieval query text")},
    Intent: embeddings.IntentRetrievalQuery,
}

{% /codetab %}

The five neutral intent constants map to a provider-independent vocabulary:

Constant	Value
IntentRetrievalQuery	retrieval_query
IntentRetrievalDocument	retrieval_document
IntentClassification	classification
IntentClustering	clustering
IntentSemanticSimilarity	semantic_similarity

Request Options

Use the Dimension field to request a specific output vector size from providers that support truncated embeddings:

{% codetabs group="lang" %} {% codetab label="Go" %}

dim := 256
content := embeddings.Content{
    Parts:     []embeddings.Part{embeddings.NewTextPart("document text")},
    Dimension: &dim,
}

{% /codetab %}

Advanced

You can pass raw intent strings and provider-specific hints via the ProviderHints field. See godoc for details.

Compatibility with Legacy API

Both the Content API and the legacy EmbedDocuments / EmbedQuery API coexist and neither is deprecated. Use whichever fits your use case:

Use Case	Recommended API
Text-only embeddings	EmbedDocuments / EmbedQuery
Mixed media (text + images)	Content API (EmbedContent / EmbedContents)
Portable intents or dimensions	Content API (EmbedContent / EmbedContents)

The legacy API continues to work exactly as before:

{% codetabs group="lang" %} {% codetab label="Go" %}

embeddings, err := ef.EmbedDocuments(ctx, []string{"text1", "text2"})
queryEmb, err := ef.EmbedQuery(ctx, "query text")

{% /codetab %}

Existing providers automatically work with the Content API when retrieved through the registry (BuildContent). The registry wraps them with built-in adapters.