iamsrikanth-dev/findly-smart-search

Findly Smarter Search

The on-device model behind Smarter Search in the Findly iPhone app. It lets people find their own photos by describing them, such as "red car at night". Everything runs on the iPhone, and no photo or search ever leaves it.

SmartSearch-1.aar is an Apple Archive containing:

• ImageEncoder.mlpackage: a 224×224 picture in, 768 numbers of length 1 out.
• TextEncoder.mlpackage: 64 token ids in, 768 numbers of length 1 out.
• tokenizer.json: the SentencePiece pieces and scores.
• LICENSE: the Apache License 2.0.

Changes from the original

This is a derivative of google/siglip-base-patch16-224, released by Google under the Apache License 2.0.

• Converted to Core ML.
• The outputs are normalized to length 1.
• The weights are compressed to 6-bit palettes, one palette for every 16 channels. Against the original model, embeddings stay above 0.99 cosine similarity.

The conversion script is Tools/SmartSearch/convert.py in the Findly repository.