//! Alter-Ego serving system. //! //! The production home of the dual-model system: a blank-slate student (SAN) //! whose every experience is validated by an independent alter-ego judge before //! it is allowed to become permanent. This crate implements inference, the //! tri-memory system, the trust-score gate, the digestion pipeline, recall, //! LoRA-based online learning, and the real-time ecosystem loop (the "skull"). //! The student checkpoint format is identical to the one produced by the Python //! Kaggle training pipeline (`training/kernels/train.py`), so a model trained //! there loads directly here. pub mod bundle; pub mod calibration; pub mod bench; pub mod checkpoint; pub mod digestion; pub mod ecosystem; pub mod error; pub mod export; pub mod lora; #[cfg(feature = "embedder-jina")] pub mod jina_embedder; pub mod memory; pub mod model; #[cfg(feature = "og-backend")] pub mod og_backend; pub mod recall; pub mod server; pub mod tokenizer_bpe; pub mod trust; pub mod vector_index; #[cfg(needle_ffi)] pub mod needle_ffi; #[cfg(needle_ffi)] pub use needle_ffi::NeedleModel; pub use error::{Error, Result}; /// Shared builders (tests today; serving bootstrap later). pub mod main_helpers { use crate::model::{ParamMap, SanConfig}; use ndarray::Array2; use std::collections::HashMap; fn rand_arr(rows: usize, cols: usize) -> Array2 { let n = rows * cols; let mut s: u64 = 0x853C49E6748FEA9B ^ (n as u64).rotate_left(17); let data: Vec = (0..n) .map(|_| { s ^= s << 13; s ^= s >> 7; s ^= s << 17; ((s >> 33) as f32 / (u32::MAX as f32) - 0.5) * 0.1 }) .collect(); Array2::from_shape_vec((rows, cols), data).expect("shape") } /// Deterministic random base weights for a config (frozen-base material). pub fn random_student(cfg: &SanConfig) -> crate::model::San { let d = cfg.d_model; let ffn = cfg.ffn_ratio * d; let mut p = ParamMap::new(); p.insert("tok_emb.weight".into(), rand_arr(cfg.vocab, d)); p.insert("pos".into(), rand_arr(cfg.max_seq_len, d)); p.insert("head.weight".into(), rand_arr(cfg.vocab, d)); for i in 0..cfg.n_layers { p.insert(format!("blocks.{i}.ln1.weight"), rand_arr(1, d)); p.insert(format!("blocks.{i}.ln1.bias"), rand_arr(1, d)); p.insert(format!("blocks.{i}.attn.qkv.weight"), rand_arr(3 * d, d)); p.insert(format!("blocks.{i}.attn.o.weight"), rand_arr(d, d)); p.insert(format!("blocks.{i}.ln2.weight"), rand_arr(1, d)); p.insert(format!("blocks.{i}.ln2.bias"), rand_arr(1, d)); p.insert(format!("blocks.{i}.mlp.0.weight"), rand_arr(ffn, d)); p.insert(format!("blocks.{i}.mlp.0.bias"), rand_arr(1, ffn)); p.insert(format!("blocks.{i}.mlp.2.weight"), rand_arr(d, ffn)); p.insert(format!("blocks.{i}.mlp.2.bias"), rand_arr(1, d)); } p.insert("ln_f.weight".into(), rand_arr(1, d)); p.insert("ln_f.bias".into(), rand_arr(1, d)); let cfg2 = cfg.clone(); crate::model::San::from_params(p, cfg2).expect("valid params") } }