Struct VectorIndex 

pub struct VectorIndex { /* private fields */ }

Vector index backed by HNSW (Hierarchical Navigable Small World) graphs.

Uses hnsw_rs for O(log N) approximate nearest-neighbour search. Deletions are handled via a soft-delete set that filters results; the index is compacted (rebuilt) when the deletion ratio exceeds 30%.

Implementations

impl VectorIndex

pub fn new(dimension: usize) -> Self

Create a new, empty index for vectors of the given dimension.

pub fn len(&self) -> usize

Number of live (non-deleted) entries.

pub fn is_empty(&self) -> bool

Whether the index contains zero live entries.

pub fn add(&mut self, chunk_id: String, embedding: Vec<f32>) -> Result<()>

Add an embedding to the index.

The embedding is L2-normalized at insert time so that the HNSW dot-product distance equals cosine distance.

pub fn remove(&mut self, chunk_id: &str)

Remove an embedding by chunk ID.

The entry is soft-deleted (filtered from search results). When the fraction of deleted entries exceeds 30 %, the HNSW graph is compacted automatically on the next search().

pub fn search(&self, query: &[f32], k: usize) -> Vec<(String, f32)>

Search for the k most similar embeddings to query.

Returns (chunk_id, cosine_similarity) pairs sorted by descending similarity.

pub fn clear(&mut self)

Clear the entire index.

pub fn compact(&mut self)

Compact the index: rebuild the HNSW graph, dropping deleted entries.

Called automatically when the deletion ratio exceeds the threshold, but can also be invoked manually.

pub fn cosine_similarity(a: &[f32], b: &[f32]) -> f32

Cosine similarity between two arbitrary vectors.

Normalizes both inputs before computing the dot product.

pub fn verify_index_integrity(&self) -> Vec<String>

Verify index integrity, returning a list of issues found.

Checks for:

• Mismatched embedding dimensions
• NaN or Inf values in vectors
• Duplicate chunk IDs
• Empty embedding vectors

pub fn check_health(&self) -> IndexHealth

Check overall health of the index.