Files
Maven/internal/memory/store.go
T
kami f02f3b55b6 memory: keep voiceprints out of note and fact recall
Speaker profiles share the vector table with notes and facts. The doc
comment said reading them through Catalog is what keeps recall from
ranking a voiceprint. It is not. Catalog controls how speaker code reads
its own rows and says nothing about Search, which scanned every row.
What actually hid them was cosine returning 0 on a width mismatch, so a
192-dim ECAPA row scored 0 against a 384-dim query. Some x-vector
exports are 384-dim, and one of those would have surfaced speaker:kami
as a recall hit carrying the name of a person.

Both backends now skip the prefix in Search, and the prefix is one
constant in internal/memory so the store layer can filter on it without
importing internal/speaker.

Two more differences between the backends closed here. ByPrefix on the
in-memory store returned the stored metadata map by reference, so a
caller editing a returned Record edited the row, while the persistent
one unmarshals fresh. And the append to upsert change in Insert is a fix
in its own right, not only a speaker concern: any re-indexed id used to
leave a second stale copy searchable.

Found in review of #74.
2026-08-01 14:12:43 +04:00

188 lines
5.4 KiB
Go

package memory
import (
"context"
"sort"
"strings"
"sync"
)
// NonRecallPrefix — rows whose id starts with this are excluded from Search by
// every backend. Speaker voiceprints live in the same vector table as notes and
// facts (internal/speaker writes them under this prefix), and they are not
// recall material: a voiceprint has no text to read back and surfacing one as a
// note hit leaks a name attached to a biometric.
//
// Reading them through Catalog.ByPrefix was documented as what keeps them out
// of recall. It is not. It controls how speaker code reads its own rows and
// says nothing about what Search scores. What actually kept them out was that
// cosine returns 0 on a width mismatch, so a 192-dim voiceprint scored 0
// against a 384-dim query. That is a coincidence of two model choices — some
// x-vector exports are 384-dim — and not an invariant. This is the invariant.
const NonRecallPrefix = "speaker:"
// Result is a single search hit.
type Result struct {
ID string
Score float64
Meta map[string]string
}
// Store is a vector memory interface.
type Store interface {
Insert(ctx context.Context, id string, vec []float32, meta map[string]string) error
Search(ctx context.Context, vec []float32, topK int) ([]Result, error)
}
// Record is a stored vector read back whole — id, vector and metadata — as
// opposed to Result, which is a search hit and carries a score instead of the
// vector.
type Record struct {
ID string
Vec []float32
Meta map[string]string
}
// Catalog is a Store that can also be enumerated by id prefix and deleted from.
//
// Search is not enough for every user of the vector table. Speaker profiles
// (internal/speaker) need to list exactly their own rows without scoring
// anything, because listing enrolled voices is not a similarity question, and
// they need Delete because a voiceprint is data about a person and "forget this
// voice" has to actually remove it. Note and fact recall use plain Store and are
// unaffected.
type Catalog interface {
Store
// ByPrefix returns every row whose id starts with prefix, in no particular
// order. An empty prefix returns everything.
ByPrefix(ctx context.Context, prefix string) ([]Record, error)
// Delete removes one row by id. Deleting a row that is not there is not an
// error: the caller asked for it to be gone and it is gone.
Delete(ctx context.Context, id string) error
}
// item is a single stored vector with metadata.
type item struct {
id string
vec []float32
meta map[string]string
}
// InMemoryStore implements Store with cosine similarity search.
type InMemoryStore struct {
mu sync.RWMutex
items []item
}
// compile-time check: InMemoryStore satisfies Catalog.
var _ Catalog = (*InMemoryStore)(nil)
func NewInMemoryStore() *InMemoryStore {
return &InMemoryStore{}
}
// Insert upserts by id, matching the persistent store.MemoryStore: a repeated
// id replaces the prior row rather than accumulating a second copy. Re-indexing
// a note is an update, and re-enrolling a voice must replace the old voiceprint
// rather than leave it searchable.
func (s *InMemoryStore) Insert(_ context.Context, id string, vec []float32, meta map[string]string) error {
s.mu.Lock()
defer s.mu.Unlock()
for i := range s.items {
if s.items[i].id == id {
s.items[i] = item{id: id, vec: vec, meta: meta}
return nil
}
}
s.items = append(s.items, item{id: id, vec: vec, meta: meta})
return nil
}
// ByPrefix implements Catalog.
func (s *InMemoryStore) ByPrefix(_ context.Context, prefix string) ([]Record, error) {
s.mu.RLock()
defer s.mu.RUnlock()
var out []Record
for _, it := range s.items {
if !strings.HasPrefix(it.id, prefix) {
continue
}
// Copy the metadata too. Returning it.meta by reference let a caller
// mutating the returned map edit the stored row, and the persistent
// backend unmarshals fresh, so the two disagreed.
meta := make(map[string]string, len(it.meta))
for k, v := range it.meta {
meta[k] = v
}
out = append(out, Record{ID: it.id, Vec: append([]float32(nil), it.vec...), Meta: meta})
}
return out, nil
}
// Delete implements Catalog.
func (s *InMemoryStore) Delete(_ context.Context, id string) error {
s.mu.Lock()
defer s.mu.Unlock()
for i := range s.items {
if s.items[i].id == id {
s.items = append(s.items[:i], s.items[i+1:]...)
return nil
}
}
return nil
}
func (s *InMemoryStore) Search(_ context.Context, vec []float32, topK int) ([]Result, error) {
s.mu.RLock()
defer s.mu.RUnlock()
if topK <= 0 {
topK = 10
}
type scored struct {
id string
score float64
meta map[string]string
}
scores := make([]scored, 0, len(s.items))
for _, it := range s.items {
if strings.HasPrefix(it.id, NonRecallPrefix) {
continue
}
score := cosine(vec, it.vec)
scores = append(scores, scored{id: it.id, score: score, meta: it.meta})
}
sort.Slice(scores, func(i, j int) bool {
return scores[i].score > scores[j].score // descending
})
if topK > len(scores) {
topK = len(scores)
}
out := make([]Result, topK)
for i := 0; i < topK; i++ {
out[i] = Result{
ID: scores[i].id,
Score: scores[i].score,
Meta: scores[i].meta,
}
}
return out, nil
}
// cosine similarity (dot product, assumes L2-normalized vectors).
func cosine(a, b []float32) float64 {
if len(a) != len(b) || len(a) == 0 {
return 0
}
var dot float64
for i := range a {
dot += float64(a[i]) * float64(b[i])
}
return dot
}