Struct VecPool 

pub struct VecPool<'a> { /* private fields */ }

Flat store of quantized vectors.

The slot bytes are an overlay of a borrowed base and an owned tail: the owned path (new/push/from_parts) keeps base = &[] with every slot in tail, byte-for-byte unchanged; the borrowed/overlay path (from_parts_borrowed/from_parts_overlay) maps an mmap’d section as base and appends new slots to tail. Because dead slots are dropped by maintain and never rewritten in place, a slot is wholly in base or wholly in tail, so a heap opened over a multi-gigabyte mmap grows without cloning it — reads dispatch on one comparison per slot.

Implementations

impl<'a> VecPool<'a>

pub fn new(dim: usize, max_bytes: usize) -> Self

Creates an empty pool for dim-dimensional vectors (dim == 0 leaves the layer inert — the pool stays empty).

pub fn stride(&self) -> usize

Byte stride of one slot.

pub fn len(&self) -> usize

Number of stored slots.

pub fn is_empty(&self) -> bool

true when no vector is stored.

pub fn pool_bytes(&self) -> usize

Total bytes held.

pub fn slot_fact(&self, i: usize) -> u32

The owning fact of slot i.

pub fn push(&mut self, fact: FactId, v: &[f32]) -> Result<u32, Error>

Quantizes v and appends it as fact’s slot, returning the slot index. Ids are not stored sorted — the fact record keeps the index.

Errors

Error::DimMismatch/Error::Invalid from quantization, or Error::CapacityExceeded at the byte ceiling.

pub fn quantize_query( &self, v: &[f32], scratch: &mut VecScratch, ) -> Result<(), Error>

Quantizes a query vector into scratch.query (sized to stride, fact left as 0). Reused across searches.

pub fn cosine_slots(&self, a: u32, b: u32) -> f32

Quantized cosine of slots a and b (similar-detection uses it on two stored facts). Returns 0.0 if either index is out of range.

pub fn search( &self, query: &[f32], k: usize, admit: &mut dyn FnMut(FactId) -> bool, scratch: &mut VecScratch, out: &mut Vec<(FactId, f32)>, ) -> Result<(), Error>

Flat two-phase search: Hamming prefilter on signatures, then exact quantized-cosine rescore of the best max(4·k, 64) candidates. admit filters by the shared recall rule; writes the top k (fact, cosine) into out, descending.

Trait Implementations

impl<'a> Debug for VecPool<'a>

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

Formats the value using the given formatter.