Jina

@dsh-ai-providers/jina is the Jina AI provider maintained by Firefox2100 as part of this project. It supplies reranking and embeddings. It is not an official DSH or Jina AI package.

Supported capabilities

Capability	Endpoint
Reranking	POST {base URL}/rerank
Embeddings	POST {base URL}/embeddings

The rerank API is the one Cohere introduced and Jina, vLLM and others copied, so a server that speaks it also works through this provider with a custom base URL. The transports live in the library package @dsh-ai-providers/jina-cohere, which this provider and the Cohere provider both use; neither depends on the other.

Configure a connection

Create one or more connections with an ID, name, base URL, and DSH credential reference in Settings > Providers > Jina connections. The default is:

Setting	Default
Base URL	https://api.jina.ai/v1
Credential reference	JINA_API_KEY
Rerank model	jina-reranker-v2-base-multilingual
Embedding model	jina-embeddings-v3
Timeout	60 seconds

Then open the Reranking or Embedding capability tab and select Jina as the active provider. Use separate connections when endpoints need different credentials or base URLs; a connection ID is configuration identity, so keep it stable once a selection refers to it. The credential reference names a DSH-managed secret, not the secret value. Self-hosted servers can leave it empty. The provider appends /rerank or /embeddings, so do not include them in the base URL.

Embeddings

The consumer's choice of query or document is sent as Jina's task: retrieval.query or retrieval.passage. Without a choice no task is sent. The task request option overrides the choice and accepts the model's other tasks (text-matching, classification, separation).

Setting	Request field	Default
Vector dimensions	dimensions	0: the model's own; also a request option chosen before a session's first request
Texts longer than the input	truncate	cut to fit
Batch size	texts per request	64

Vectors of different sizes or models cannot be compared, so choose the dimensions before the first request and keep the model.

Reranking

Requests contain model, query, documents, optional top_n, and return_documents: false. Responses may use either relevance_score or score. Token usage is read when present. The service rejects empty queries and documents, sorts results by descending score, and normalizes raw logits outside 0 to 1 through a logistic function, which keeps their order.

Jina's rerank API takes nothing beyond top_n and return_documents, so this provider describes no rerank request options. jina-reranker-m0 can also rank images, but the rerank contract carries text only, so documents are always text.

Every result must identify its input document by index. Indices outside the submitted list, missing scores, or non-numeric scores are malformed responses and fail the call. Jina and Cohere do not promise identical models, limits, or score calibration: keep candidate sets within the service's limits and compare scores only within one response.

Diagnosing compatibility

If a server rejects the request, verify the configured prefix and model, then compare its request shape with the fields above. Authentication, quota, and unknown-model responses become rejected; connectivity and server failures become unavailable; malformed caller input is invalid-input.

See Reranking and Embeddings for consumer usage.