Skip to main content

Advanced

This section is the knobs and dials of the AI proxy: which models it can use, the rules each one has, and the extra data some of them return. You don't need any of it to get started — but when you want faces, audio, code, or fine control, it's all here.

The one idea to hold onto: model-aware stores​

Ahnlich DB stores and compares vectors. The AI proxy's whole job is turning your raw input (text, images, or audio) into those vectors using a machine-learning model. So every AI store remembers two models:

  • an index model — used to embed the items you store, and
  • a query model — used to embed the things you search with.
When you store (SET)an itemtext·img·audioIndex modelML model[0.12, 0.98, …]When you search (GETSIMN)your query"a red bike"Query modelML model[0.10, 0.95, …]must share one dimensionAhnlich DBstore · compare
A store remembers two models: one to embed what you store, one to embed what you search with. As long as they output vectors of the same size, the query and the stored items live in one space and can be compared.

There's exactly one rule: both models must produce vectors of the same size (the "embedding dimension"). If they do, your query and your stored items live in the same space and can be compared. That's why the model tables always list an Embedding dim — matching it is how you know two models are compatible.

# same model both sides — always compatibleCREATESTORE my_store QUERYMODEL all-minilm-l6-v2 INDEXMODEL all-minilm-l6-v2

In this section​

  • Models & input types — every built-in model, grouped by family, plus the raw input types and how to pick one.
  • Model constraints — the input rules for audio and face models, and which pairs are cross-modal.
  • Code search — build a natural-language + code-to-code search engine with Jina Code v2.
  • Model parameters — per-request tuning (model_params), used today by the face models.
  • Embedding metadata — bounding boxes, confidence, and age/gender returned by face models.