Jina Ai

Jina AI provides three highly practical primitives for agent RAG pipelines: (1) embeddings — generate multilingual vector embeddings (jina-embeddings-v3: 8192 token context, 89 languages, 1024 dims) via a clean OpenAI-compatible API; (2) reranking — given a query and N retrieved documents, rerank by cross-encoder relevance score (jina-reranker-v2) to improve RAG precision; (3) web reader — GET r.jina.ai/{url} returns clean markdown from any webpage, zero auth required. Free tier: 1M embedding tokens/month. All three primitives compose naturally with standard RAG architectures. Confidence is docs-derived.

Embeddings API: POST https://api.jina.ai/v1/embeddings. Body: {"model": "jina-embeddings-v3", "input": ["text1", "text2", ...], "task": "retrieval.passage"|"retrieval.query"|"text-matching"|"classification"|"separation", "dimensions": 1024}. Response: {data: [{index: 0, embedding: [0.1, 0.2, ...], object: "embedding"}], usage: {total_tokens: N}}. Reranking API: POST https://api.jina.ai/v1/rerank. Body: {"model": "jina-reranker-v2-base-multilingual", "query": "user question", "documents": ["doc1 text", "doc2 text", ...], "top_n": 5}. Response: {results: [{index: 0, document: {text: "..."}, relevance_score: 0.92}]}. Web Reader: GET https://r.jina.ai/{url} — returns clean markdown (no auth required, but API key adds higher rate limits). Deep Search: POST https://s.jina.ai/ with X-With-Generated-Alt: true for image alt text generation.

API key auth: Authorization: Bearer {jina_api_key}. Keys from jina.ai → API Keys page. HTTPS enforced. Web Reader: works without API key (free, lower rate limits) or with key (higher limits). No OAuth2. Rate limits: free tier (1M embedding tokens/month, ~60 req/min); paid tiers scale. OpenAI-compatible API: embeddings endpoint is drop-in compatible with OpenAI embeddings client (change base_url to https://api.jina.ai/v1). No per-request auth delegation — single API key per project.

Batch embeddings: send up to 2048 texts per request for efficiency; API handles batching server-side. Long texts: jina-embeddings-v3 supports 8192 tokens — no chunking needed for most documents. Rate limiting: 429 with Retry-After header; implement exponential backoff. Reranking: O(n) latency where n is number of documents — keep reranked candidates ≤ 100 for low latency. Web Reader reliability: dependent on target URL accessibility; JavaScript-heavy SPAs may not render fully. Error responses: standard HTTP status codes with {detail: "..."} body. Embedding dimensions: support for Matryoshka representation — reduce to 256/512 dims for storage efficiency.

jina.ai/api-dashboard covers Embeddings API, Reranker API, Web Reader, and usage dashboard. Getting started: get API key, POST /v1/embeddings — first embedding in under 2 minutes. Free tier (1M tokens/month). GitHub: jina-ai organization (30k+ stars across repos). Community via Discord (active). Good API documentation; Jina AI has one of the most straightforward embedding API setups. Web Reader is a standout utility for agent web extraction.