post-cortex-embeddings 0.5.0

Embedding engines (Model2Vec static + BERT via Candle + static-hash fallback) and HNSW vector database for post-cortex. Self-contained ML stack — depend on this crate when you need vector search without the full post-cortex orchestrator.
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37

// Copyright (c) 2025, 2026 Julius ML
// Licensed under the MIT License. See LICENSE at the workspace root.

//! HNSW vector database — types + impl.
//!
//! [`types`] is dependency-light (only `serde`, `chrono`, `uuid`,
//! `std::sync::atomic`). The HNSW index + product-quantization codebook
//! live in private modules and are reachable through the [`VectorDB`]
//! entry point. Storage backends that only need to round-trip
//! [`VectorMetadata`] and [`SearchMatch`] can depend on
//! `post-cortex-embeddings` with `default-features = false`, which
//! compiles only this `types` module.

pub mod types;

// HNSW impl modules. Behind the `bert` feature today via the rest of the
// pipeline; we keep them unconditional so non-BERT consumers can still
// build the HNSW path with a custom embedding backend.
mod common;
pub mod config;
pub mod core;
mod hnsw_index;
pub mod product_quantization;
pub mod similarity;
mod text_index;

#[cfg(test)]
mod tests;

pub use config::{SearchMode, SearchQualityPreset, VectorDbConfig};
pub use core::VectorDB;
pub use product_quantization::ProductQuantizationCodebook;
pub use similarity::{cosine_similarity, cosine_with_norms, dot, l2_norm};
pub use types::{
    Corpus, MESSAGE_CONTENT_TYPE, SearchMatch, StoredVector, VectorDbStats, VectorDbStatsSnapshot,
    VectorMetadata,
};