Jina AI

What Jina AI provides

Jina AI develops search and RAG foundation models, operates hosted APIs at api.jina.ai, and offers cloud-marketplace or self-hosted options. Embeddings turn text or multimodal input into vectors, rerankers refine candidate order, and Reader and Search are separate hosted products. Jina is not a general chat-model catalog.

Direct versus gateway authentication

Direct Search Foundation calls use a Jina-issued key, Jina base URL, and Authorization: Bearer. This site uses the same header syntax but requires this site's base URL, a site-issued key, and an exact marketplace model ID. Its api_paths list only the adapted /v1/embeddings and /v1/rerank entry points; it does not promise Reader, Search, every Jina model, or every native extension field.

Embedding input, task, and input_type

Depending on the model, input can contain text, text objects, or lists; selected multimodal models also accept image or PDF URLs and Base64 bytes. task is not a generic input type: on jina-embeddings-v3/v4 it selects adapters such as retrieval.query, retrieval.passage, text-matching, classification, or separation. input_type=query/document belongs to selected ColBERT-style multi-vector protocols. Contract-test fields, truncation, and content types against the exact model and gateway route.

Rerank after retrieval

A typical RAG pipeline first retrieves a broad candidate set with embeddings or lexical search, then sends query, documents, and the desired top_n to the reranker and restores order using returned scores and original indexes. A reranker does not store the corpus or replace a vector database, authorization filters, or citation handling. Candidate count and text length directly affect tokens, latency, and cost.

Why dimension changes require rebuilding

Selected Jina models support Matryoshka dimensions, but allowed values, defaults, and useful minimums differ by model. The vector-store schema must match output size; vectors created by another model, task, dimension, or normalization setting cannot be assumed comparable. Build a new index, re-embed the corpus, and regression-test retrieval and ranking before switching traffic.

Limits, keys, and data safety

Jina's hosted API enforces RPM, TPM, concurrency, and IP limits that vary by key tier; inspect current response headers and 429 details instead of hard-coding figures. Jina's product FAQ and legal page state that customer requests, inputs, prompts, and uploaded content are not used to train its models, while current processing is governed by the referenced Elastic terms. Keep API keys server-side and revoke exposed keys; minimize sensitive inputs, verify lawful sources and the current DPA/privacy terms, and never use Reader to bypass site access controls.