openpay/ (+3 crates, ~8400 LOC, 162 tests)
op-iso8583/ ISO 8583 messages for direct acquirer / card-network
connectivity. MTIs, primary+secondary bitmaps, full
DE catalogue, BCD/ASCII/EBCDIC/LLVAR/LLLVAR codecs,
five dialects (Visa Base I, Mastercard MDS, Amex GNS,
Discover, JCB), network-management messages
(sign-on/echo/cutover), DE 55 EMV TLV.
op-mobile-wallets/ Apple Pay (ECDH-P256 + AES-256-GCM), Google Pay ECv2
(ECDH + HKDF + AES-256-CTR + HMAC), Samsung Pay. PAN
never leaves the decryption boundary as a String;
lands in VaultRef. No-PAN-leakage assertions per
wallet.
op-routing/ Least-cost routing + MCC-aware routing + intelligent
retry / soft-decline recovery. 231-MCC catalogue.
Real ISO 8583 DE-39 soft/hard decline taxonomies.
smart-byte/smart-byte-rs/ (+3 crates, ~7100 LOC, 96 tests)
smart-byte-vc/ W3C Verifiable Credentials 2.0: VCDM, presentations,
Data Integrity proofs (eddsa-jcs-2022), VC-JWT,
SD-JWT (selective disclosure), Bitstring Status
List 2021. Cargo-bridge maps VCs to Smart Byte
envelopes with SAID stability.
smart-byte-did/ W3C DID resolution: did:key (Ed25519/P-256/secp256k1),
did:web (HTTPS), did:peer (numalgo 2+4), did:jwk,
did:ion (feature-gated stub). DID URL dereferencer.
smart-byte-pq/ Post-quantum signatures: ML-DSA-44/65/87 (FIPS 204),
SLH-DSA across 12 parameter sets (FIPS 205), FN-DSA
(draft FIPS 206, feature-gated). Hybrid Ed25519 +
ML-DSA pattern for the transition. Secret-key
zeroize-on-drop.
eoc/eoc-rs/ (+3 crates, ~6450 LOC, 76 tests)
eoc-vendor-api/ NeuralBackend implementations: Anthropic, OpenAI,
Google, Mistral, Cohere, Groq, Together, Fireworks.
SSE streaming. Prompt-caching where supported.
Joule estimates per published energy coefficient
(12-model curated table).
eoc-embeddings/ 21 embedding models behind one Embedder trait:
OpenAI v3 (3), Cohere v3 (3), Voyage (5), Mistral,
Jina, plus local ONNX (BGE x3, nomic, mxbai, GTE,
E5) behind 'local' feature. Matryoshka-style
dimension projection. Per-content-hash cache.
eoc-route-learned/ RouteLLM-style matrix-factorisation router +
multinomial-logreg classifier + LinUCB / Thompson
bandits for online learning. LearnedCascade wraps
the rule-based cascade with a confidence-threshold
skip policy. Synthetic-eval accuracy: MF 0.935,
LogReg 0.980.
Quality:
- cargo check across all three workspaces: clean
- cargo clippy --all-targets -- -D warnings on every new crate: clean
- 334 new tests total; all passing
- #![forbid(unsafe_code)] across every new crate
- No unwrap/expect outside #[cfg(test)]
Rust 1.95.0 edition 2024 (inherited via root rust-toolchain.toml).
105 lines
3.1 KiB
Rust
105 lines
3.1 KiB
Rust
//! Optional in-memory embedding cache.
|
|
//!
|
|
//! Embedders that hit the network (or run expensive local ONNX) benefit
|
|
//! from a content-addressed cache so the same string isn't re-embedded.
|
|
//! Keys are blake3 hashes of the (model, text) pair; values are the
|
|
//! produced vectors.
|
|
|
|
use std::num::NonZeroUsize;
|
|
use std::sync::Mutex;
|
|
|
|
use lru::LruCache;
|
|
|
|
/// Content-addressed cache key: blake3 hash of `(model, text)`.
|
|
#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
|
|
pub struct ContentHash(pub [u8; 32]);
|
|
|
|
impl ContentHash {
|
|
/// Hash a `(model, text)` pair.
|
|
pub fn of(model: &str, text: &str) -> Self {
|
|
let mut h = blake3::Hasher::new();
|
|
h.update(model.as_bytes());
|
|
h.update(&[0u8]);
|
|
h.update(text.as_bytes());
|
|
let mut out = [0u8; 32];
|
|
out.copy_from_slice(h.finalize().as_bytes());
|
|
Self(out)
|
|
}
|
|
}
|
|
|
|
/// Bounded LRU cache of embedding vectors.
|
|
pub struct EmbeddingCache {
|
|
inner: Mutex<LruCache<ContentHash, Vec<f32>>>,
|
|
}
|
|
|
|
impl EmbeddingCache {
|
|
/// Construct with the given capacity. Panics only at construction time
|
|
/// if `capacity == 0` (impossible at runtime since the type enforces it).
|
|
pub fn new(capacity: usize) -> Self {
|
|
// `NonZeroUsize::new(1)` is infallible at construction.
|
|
let cap = NonZeroUsize::new(capacity.max(1)).unwrap_or_else(|| {
|
|
// Provably unreachable — the above max(1) guarantees ≥ 1.
|
|
#[allow(clippy::expect_used)]
|
|
NonZeroUsize::new(1).expect("1 is non-zero")
|
|
});
|
|
Self {
|
|
inner: Mutex::new(LruCache::new(cap)),
|
|
}
|
|
}
|
|
|
|
/// Look up an embedding by `(model, text)`.
|
|
pub fn get(&self, model: &str, text: &str) -> Option<Vec<f32>> {
|
|
let key = ContentHash::of(model, text);
|
|
let mut guard = self.inner.lock().ok()?;
|
|
guard.get(&key).cloned()
|
|
}
|
|
|
|
/// Store an embedding for `(model, text)`.
|
|
pub fn put(&self, model: &str, text: &str, vector: Vec<f32>) {
|
|
let key = ContentHash::of(model, text);
|
|
if let Ok(mut guard) = self.inner.lock() {
|
|
guard.put(key, vector);
|
|
}
|
|
}
|
|
|
|
/// Current number of entries.
|
|
pub fn len(&self) -> usize {
|
|
self.inner.lock().map(|g| g.len()).unwrap_or(0)
|
|
}
|
|
|
|
/// Whether the cache is empty.
|
|
pub fn is_empty(&self) -> bool {
|
|
self.len() == 0
|
|
}
|
|
}
|
|
|
|
#[cfg(test)]
|
|
mod tests {
|
|
use super::*;
|
|
|
|
#[test]
|
|
fn hash_is_deterministic() {
|
|
let a = ContentHash::of("m", "hello");
|
|
let b = ContentHash::of("m", "hello");
|
|
let c = ContentHash::of("m", "world");
|
|
assert_eq!(a, b);
|
|
assert_ne!(a, c);
|
|
}
|
|
|
|
#[test]
|
|
fn cache_round_trip() {
|
|
let c = EmbeddingCache::new(8);
|
|
assert!(c.get("m", "hi").is_none());
|
|
c.put("m", "hi", vec![0.1, 0.2, 0.3]);
|
|
assert_eq!(c.get("m", "hi"), Some(vec![0.1, 0.2, 0.3]));
|
|
assert_eq!(c.len(), 1);
|
|
}
|
|
|
|
#[test]
|
|
fn cache_zero_capacity_clamps_to_one() {
|
|
let c = EmbeddingCache::new(0);
|
|
c.put("m", "x", vec![1.0]);
|
|
// Size 1: store then retrieve.
|
|
assert!(c.get("m", "x").is_some());
|
|
}
|
|
}
|