Add Jina Embeddings v5 support

[hanxiao]

Summary

Add JinaDenseEmbedding as a new embedding provider, enabling Jina Embeddings v5 models for dense vector generation.

Two new files follow the existing provider pattern (OpenAI/Qwen):

• jina_function.py - Base class with Jina API client logic
• jina_embedding_function.py - JinaDenseEmbedding implementing DenseEmbeddingFunction Protocol

Features

• Task-specific embeddings via task parameter (retrieval.query, retrieval.passage, text-matching, classification, separation)
• Matryoshka dimension support (32, 64, 128, 256, 512, 768/1024)
• Uses OpenAI-compatible API (requires openai package, same as existing OpenAI provider)

Usage

from zvec.extension import JinaDenseEmbedding

# For retrieval: use different task types for queries vs documents
query_emb = JinaDenseEmbedding(task="retrieval.query")
doc_emb = JinaDenseEmbedding(task="retrieval.passage")

query_vector = query_emb.embed("What is machine learning?")
doc_vector = doc_emb.embed("Machine learning is a subset of artificial intelligence...")

# With custom dimension (Matryoshka)
emb = JinaDenseEmbedding(
    model="jina-embeddings-v5-text-small",
    dimension=256,
    task="text-matching",
)

Benchmarks

MMTEB scores vs model size. jina-v5-text models (red) outperform models 2-16x their size.

MTEB English v2 scores. v5-text-nano (239M) achieves 71.0, matching models with 2x+ parameters.

Both models are open-weight (Apache 2.0) and support Matryoshka dimension reduction, task-specific embeddings, and local deployment via GGUF/MLX.

Links

• Paper: arXiv:2602.15547
• Blog: jina.ai/blog/jina-embeddings-v5-text
• MTEB Leaderboard: huggingface.co/spaces/mteb/leaderboard
• HuggingFace: huggingface.co/jinaai

[CLAassistant]

All committers have signed the CLA.

[hanxiao]

@CLAassistant check