Struct BruteForceVectorIndex 

pub struct BruteForceVectorIndex { /* private fields */ }

A cosine-similarity brute-force vector index.

Stores L2-normalised f32 vectors in a flat row-major buffer for cache-friendly linear scans. search(q, k) is O(n * dim) in time and O(n) in allocations; a min-heap optimisation is not worth the complexity at the corpus sizes this impl targets (see module docs).

Implementations

impl BruteForceVectorIndex

pub fn empty(model: impl Into<String>, dim: u32) -> Self

Construct an empty index for (model, dim).

Agents who want to stream insert rather than build from a repo can start here. The repo-scan path (Self::build_from_repo) is the common case.

pub fn model(&self) -> &str

Model identifier this index is bound to (e.g. "openai:text-embedding-3-small"). Exposed so downstream consumers (e.g. the KNN-edge derivation in mnem-http’s GraphCache) can tag their derived artefacts with the same model string the vectors were indexed under.

pub const fn dim(&self) -> u32

Dimensionality of every stored vector. 0 iff the index was empty()-constructed and never inserted into.

pub fn is_empty(&self) -> bool

true iff no vectors were indexed.

pub fn points_iter(&self) -> impl Iterator<Item = (NodeId, &[f32])> + '_

Iterate (node_id, unit_vector_slice) pairs in build order (canonical Prolly-key order at build time). The returned slice is borrowed from the flat row-major buffer; every row is already L2-normalised so cosine == dot product.

Used by mnem-http’s GraphCache KNN-edge fallback to derive a deterministic KNN-edge substrate when the authored-edges adjacency is empty (experiment E0 wire activation). Returning a borrowed slice avoids the per-row to_vec() clone the HNSW variant pays.

pub fn try_insert(&mut self, node_id: NodeId, embed: &Embedding) -> bool

Insert one node’s embedding. The node’s embedding MUST match self.model and self.dim; mismatched entries are silently skipped so callers can feed a heterogeneous stream.

Returns true if the vector was indexed, false if it was skipped (wrong model, wrong dim, absent, or undecodable).

pub fn build_from_repo(repo: &ReadonlyRepo, model: &str) -> Result<Self, Error>

Build an index over every node at the repo head whose embedding under model is present in the per-commit sidecar (Commit.embeddings Prolly tree, keyed by NodeCid). Nodes without a sidecar entry under model are silently skipped.

The sidecar is the only source of truth: dense vectors live in a separate tree so nondeterministic producers (e.g. ORT thread-count drift) cannot leak into NodeCid and break federated dedup. Operators with repos authored before the sidecar shipped must run mnem reindex to lift inline vectors into the sidecar; until then those vectors are invisible to retrieval.

Errors

• RepoError::Uninitialized if the repo has no head commit.
• Store / codec errors walking the node tree, decoding nodes, or walking the embedding sidecar.
• crate::error::ObjectError::EmbeddingSizeMismatch if a node carries an embedding whose vector.len() contradicts dim * bytes_per_dtype(dtype).

Trait Implementations

impl Clone for BruteForceVectorIndex

fn clone(&self) -> BruteForceVectorIndex

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for BruteForceVectorIndex

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl VectorIndex for BruteForceVectorIndex

fn model(&self) -> &str

Embedding model this index was built for.

fn dim(&self) -> u32

Vector dimension the index accepts on queries.

fn search(&self, query: &[f32], k: usize) -> Result<Vec<VectorHit>, Error>

Nearest-neighbour lookup. Returns up to k hits.

fn len(&self) -> usize

Number of indexed vectors.

fn is_empty(&self) -> bool

true iff no vectors were indexed.