Vector Store

Crate: nis2-vectordb | Files: embed.rs, lance.rs, store.rs | Tests: 1

What it does

Indexes legal text chunks as 384-dimensional vectors using BGE-Small-EN-v1.5 (via fastembed/ONNX Runtime) and stores them in LanceDB with an HNSW index. Provides semantic search to retrieve the most relevant provisions for a natural-language query.

Components

• Embedder wraps fastembed::TextEmbedding behind a Mutex (fastembed requires &mut self)
• LanceStore manages the LanceDB connection and provides index() and search() methods

Arrow schema

Column	Type
id	Utf8
directive	Utf8
article	Utf8
paragraph	Utf8
letter	Utf8 (nullable)
text	Utf8
vector	FixedSizeList(Float32, 384)

Binaries

• cargo run --bin indexer -- populates data/lancedb/ from data/sources/
• cargo run --bin search -- "<query>" -- returns the 5 most similar chunks with similarity scores

Test

• embedding_dimensions_and_semantic_similarity -- verifies 384 dimensions and that semantically similar texts score higher