Module embeddings 

Local-first visual embedding proxies for decentralized CBIR / near-duplicate search.

Honesty

This is not foundation CLIP (or any trained open-vocab embedder). It is a pure-Rust, zero-model-weight local CBIR proxy:

• perceptual_hash_u64 — aHash / dHash structure fingerprints
• color_hist_embed — fixed RGB joint histogram, L2-normalized into a caller buffer
• embed_distance — Hamming (hashes) + cosine (float embeds)

Place ONNX CLIP (or ResNet) under vendor/vision/embeddings/ when product needs semantic retrieval; keep these APIs as the cheap, always-on pre-filter.

Re-exports

pub use color_hist_embed::color_hist_embed_rgb;

pub use color_hist_embed::COLOR_HIST_BINS;

pub use color_hist_embed::COLOR_HIST_EMBED_DIM;

pub use embed_distance::cosine_distance;

pub use embed_distance::cosine_similarity;

pub use embed_distance::hamming_distance_u64;

pub use perceptual_hash_u64::ahash_u64;

pub use perceptual_hash_u64::dhash_u64;

pub use perceptual_hash_u64::AHASH_SIDE;

pub use perceptual_hash_u64::DHASH_HEIGHT;

pub use perceptual_hash_u64::DHASH_WIDTH;

Modules

color_hist_embed

Fixed-size RGB colour histogram → L2-normalized f32 embedding (caller buffer).

embed_distance

Distances for local visual embedding proxies (hash Hamming + float cosine).

perceptual_hash_u64

aHash / dHash perceptual hashes over a gray buffer.