Crate nusy_graph_query 

nusy-graph-query

Graph-native semantic search for Arrow RecordBatches — embeddings, traversal, hybrid ranking, and caching.

This crate provides the building blocks for combining structural graph queries with semantic similarity search over Apache Arrow data.

Quick Example

use nusy_graph_query::{HashEmbeddingProvider, EmbeddingProvider, cosine_similarity};

let provider = HashEmbeddingProvider::new(384);
let vecs = provider.embed_batch(&[
    "Alice knows Bob".to_string(),
    "Cat is an animal".to_string(),
]).unwrap();

let sim = cosine_similarity(&vecs[0], &vecs[1]);
assert!(sim >= -1.0 && sim <= 1.0);

Modules

• embedding — EmbeddingProvider trait, hash provider, cosine similarity
• traversal — Generic BFS/DFS over Arrow edge RecordBatches
• [hybrid_rank] — Combine structural + semantic scores
• cache — Content-hash embedding cache with Parquet persistence
• [subprocess] — Python sentence-transformers provider (feature: subprocess)
• [fastembed_provider] — Local ONNX embedding provider (feature: fastembed)

Feature Flags

Flag	Description
subprocess	Enable Python sentence-transformers provider
fastembed	Enable local ONNX embedding via fastembed-rs (~2ms/chunk)

Re-exports

pub use cache::EmbeddingCache;

pub use embedding::EmbeddedItem;

pub use embedding::EmbeddingError;

pub use embedding::EmbeddingProvider;

pub use embedding::HashEmbeddingProvider;

pub use embedding::SearchResult;

pub use embedding::cosine_similarity;

pub use embedding::hash_to_vector;

pub use embedding::semantic_search;

pub use hybrid_rank::HybridConfig;

pub use hybrid_rank::RankCandidate;

pub use hybrid_rank::RankedResult;

pub use hybrid_rank::hybrid_rank;

pub use traversal::Direction;

pub use traversal::EdgeSchema;

pub use traversal::TraversalNode;

pub use traversal::bfs;

pub use traversal::bfs_with_adjacency;

pub use traversal::build_adjacency;

pub use traversal::build_adjacency_from_list;

Modules

cache

Embedding cache — content-hash invalidation with Parquet persistence.

embedding

Embedding infrastructure — provider trait, hash provider, cosine similarity.

hybrid_rank

Hybrid ranking — combine structural and semantic scores.

traversal

Graph traversal — generic BFS/DFS over Arrow edge RecordBatches.