jina-ai-provider weekly downloads statistics

AI SDK - Jina AI Provider

Introduction

The Jina AI Provider is a provider for the AI SDK. It provides a simple interface to the Jina AI API for both text and multimodal embeddings.

Installation

Configuration

The Jina AI Provider requires an API key to be configured. You can obtain an API key by signing up at Jina.

Add the following to your .env file:

Provider Instance

You can use the default provider instance or create your own configured instance.

You can use the following optional settings to customize the Jina provider instance:

• baseURL string • The base URL of the Jina API. Defaults to https://api.jina.ai/v1. • apiKey string • API key sent via the Authorization header. Defaults to the JINA_API_KEY environment variable. • headers Record • Custom headers to include with every request. • fetch (input: RequestInfo, init?: RequestInit) => Promise • Custom fetch implementation or middleware (for interception, testing, etc.).

Usage

Text Embeddings

Multimodal Embeddings (Text + Images)

[!TIP] Use MultimodalEmbeddingInput type to ensure type safety when using multimodal embeddings.

Provider options

Pass Jina embedding options via providerOptions.jina. See supported fields below.

Supported provider options via providerOptions.jina:

• inputType 'text-matching' | 'retrieval.query' | 'retrieval.passage' | 'separation' | 'classification' • Intended downstream application to help the model produce better embeddings. Defaults to 'retrieval.passage'. • 'retrieval.query': input is a search query. • 'retrieval.passage': input is a document/passage. • 'text-matching': for semantic textual similarity tasks. • 'classification': for classification tasks. • 'separation': for clustering tasks. • outputDimension number • Number of dimensions for the output embeddings. See model docs for ranges. • jina-embeddings-v3: min 32, max 1024. • jina-clip-v2: min 64, max 1024. • jina-clip-v1: fixed 768. • embeddingType 'float' | 'binary' | 'ubinary' | 'base64' • Data type for the returned embeddings. Defaults to 'float'. • normalized boolean • Whether to L2-normalize embeddings. Defaults to true. • truncate boolean • Whether to truncate inputs beyond the model context limit instead of erroring. Defaults to false. • lateChunking boolean • Split long inputs into 1024-token chunks automatically. Defaults to false. Only for text embedding models.

Max Embeddings Per Call

The Jina AI Provider supports up to 2048 embeddings per call.

Jina embedding models:

| Model | Context Length (tokens) | Embedding Dimension | Modalities |

| jina-embeddings-v3 | 8,192 | 1024 | Text | | jina-clip-v2 | 8,192 | 1024 | Text + Images | | jina-clip-v1 | 8,192 | 768 | Text + Images |

Supported Input Formats

Text Embeddings

• Array of strings: • const strings = ["text1", "text2"]

Multimodal Embeddings

• Text objects: • const text = [{ text: "Your text here" }] • Image objects: • const image = [{ image: "https://example.com/image.jpg" }] • const image = [{ image: "base64-encoded-image" }] • Mixed arrays: • const mixed = [{ text: "object text" }, { image: "image-url" }, {image: "base64-encoded-image"}]

[!TIP] You can pass base64 encoded image to image property.

Authors

• patelvivekdev