Initial prototype

[simonw]

Start with the jina-clip-v1-api model.

[simonw]

Jina API keys come with 1,000,000 free tokens. They charge about 2 cents per million tokens for more.

Images are charged at 1,000 tokens per 224x224 tile - I burned through the 1,000,000 token free trial with just a couple of images because I didn't resize them first!

Resizing to 896x896 before sending to the API uses a LOT less tokens - 16,000 per image. I've implemented that using Pillow.

[simonw]

I released this as an alpha, but there's a bunch of stuff I still need to do:

• Properly test it - build a collection of photos, run searches etc
• Consider if the resizing is the right thing to do, should I have an option for resizing to dimensions larger than 896x896?
• Add unit tests
• Implement support for all of Jina's other embedding models

[simonw]

Full list of their current models: https://jina.ai/embeddings/

• jina-embeddings-v3
• jina-colbert-v2
• jina-clip-v1
• jina-colbert-v1-en
• jina-embeddings-v2-base-es
• jina-embeddings-v2-base-code
• jina-embeddings-v2-base-de
• jina-embeddings-v2-base-zh
• jina-embeddings-v2-base-en

[simonw]

Some models have extra options, including query v.s. document for the Colbert ones: