Struct FannAdapter 

pub struct FannAdapter { /* private fields */ }

Adapter that lets ruv-FANN store and retrieve layer-weight embeddings.

Each neural-network layer can be fingerprinted as a flat f32 embedding (e.g. the flattened weight matrix or its PCA projection). Storing these fingerprints in RuVector enables fast recall of “similar layers” across model checkpoints.

Example

use ruvector_core::integration::FannAdapter;

let mut adapter = FannAdapter::new(128, "./fann_index.db").unwrap();
adapter.store_layer("model_v1/layer_0", &[0.1f32; 128], None).unwrap();
let similar = adapter.find_similar_layers(&[0.1f32; 128], 5).unwrap();

Implementations

impl FannAdapter

pub fn new(dimensions: usize, storage_path: impl Into<String>) -> Result<Self>

Create a new adapter backed by a RuVector database.

dimensions must match the size of the layer embeddings you intend to store. Cosine distance is used because weight embeddings are typically meaningful up to scale.

pub fn store_layer( &self, layer_id: impl Into<String>, embedding: &[f32], metadata: Option<HashMap<String, Value>>, ) -> Result<String>

Store a layer embedding identified by layer_id.

metadata can carry arbitrary JSON-serialisable key-value pairs (e.g. model name, checkpoint step, layer type).

pub fn find_similar_layers( &self, query: &[f32], k: usize, ) -> Result<Vec<SearchResult>>

Find the k most similar layer embeddings to query.

Returns results sorted by ascending cosine distance.

pub fn find_similar_layers_filtered( &self, query: &[f32], k: usize, filter: HashMap<String, Value>, ) -> Result<Vec<SearchResult>>

Find similar layers with a filter on metadata fields.

Only results where every (key, value) in filter matches are returned.

pub fn delete_layer(&self, layer_id: &str) -> Result<bool>

Delete a layer embedding by ID.

pub fn len(&self) -> Result<usize>

Total number of stored layer embeddings.

pub fn is_empty(&self) -> Result<bool>

Returns true if no embeddings have been stored yet.