post-cortex-embeddings

Embedding engines + in-memory vector database for post-cortex.

Self-contained ML stack — no dependency on any other post-cortex crate, so anyone needing a local text embedder or a nearest-neighbour index can depend on this crate without pulling the daemon or orchestrator.

• LocalEmbeddingEngine selects an EmbeddingBackend from EmbeddingConfig::model_type. The default is PotionMultilingual — Model2Vec static embeddings (minishlab/potion-multilingual-128M, 256 dimensions, multilingual, no GPU). PotionCode (potion-code-16M) is the other Model2Vec model. The BERT models (MiniLM, MultilingualMiniLM, TinyBERT, BGESmall, via Candle) need the opt-in bert feature. StaticSimilarityMRL is a hash-based fallback that does not produce semantic vectors.
• VectorDB is a DashMap-backed in-memory index with per-source (session) and per-content-id indexes. Up to exact_search_threshold vectors (default 10 000) searches are exact bounded top-k scans and no HNSW graph is kept; past it the HNSW graph is built once in bulk and maintained incrementally. search_in_sources always scans only the given sessions' vectors. Quantization and product quantization are off by default.

Install

[dependencies]
post-cortex-embeddings = "0.4"                                     # Model2Vec + VectorDB (default)
post-cortex-embeddings = { version = "0.4", features = ["bert"] }  # + BERT via Candle
# or without any model backend (VectorDB, VectorMetadata, SearchMatch, hash fallback only):
post-cortex-embeddings = { version = "0.4", default-features = false }

Features

With both model2vec and bert off the crate still compiles VectorDB and its types, which is how post-cortex-storage uses it. Selecting a model whose backend is not compiled in fails at LocalEmbeddingEngine::new with an error naming the missing feature.

Example

use post_cortex_embeddings::{
    EmbeddingConfig, LocalEmbeddingEngine, VectorDB, VectorDbConfig, VectorMetadata,
};

async fn run() -> Result<(), Box<dyn std::error::Error>> {
    // Downloads potion-multilingual-128M (~50 MB) on first use.
    let engine = LocalEmbeddingEngine::new(EmbeddingConfig::default()).await?;

    // The index dimension must match the model (256 for the default model).
    let db = VectorDB::new(VectorDbConfig {
        dimension: engine.embedding_dimension(),
        ..VectorDbConfig::default()
    })?;

    let vector = engine.encode_text("hello world").await?;
    let meta = VectorMetadata::new(
        "doc-1".into(),       // id
        "hello world".into(), // text
        "session-1".into(),   // source (session id)
        "update".into(),      // content type
    );
    db.add_vector(vector.clone(), meta)?;

    for hit in db.search_in_sources(&vector, 5, &["session-1"])? {
        println!("{} (similarity {:.3})", hit.metadata.text, hit.similarity);
    }
    Ok(())
}

License

MIT — see LICENSE.