a reminder said whole no longer asks "Когда?" (V-572)

"напомни в 11:00 позвонить маме" answered "Когда?" about an hour he had
just said. ReminderGrammar builds its slots by hand and the router ran no
extraction over a stage-0 decision, so HasTime was false however plainly
the hour was spoken; missingFor read the silence as absence.

The fix runs the stage-2 extractor over every stage-0 decision, filling
only the slots the grammar left empty. A matched value always wins: the
rule read a literal pattern, the extractor guesses. This is the same hole
the LLM path already had, so fillSlots and the new stage-0 call share one
fillMatchedSlots.

Enabled for all ten grammars rather than a chosen few, because for every
intent but reminder it is inert. Extract fills Time for a reminder, Fn for
an act and Key for a fact, and nothing at all for query, system, note or
chat — which is what the clock, agenda, feed, list, task, Praxis-adjacent
and narrative rules emit. The two act rules, wakeword-act and the Praxis
ones, already carry an Fn or they do not match, so the matcher has nothing
left to fill. Measured rather than asserted: benchmarked at 20000x, a
stage-0 query is 3.7µs against 3.9µs before and a clock or act rule is
0.7µs either way, both inside the noise. The reminder rule is the one that
gains, and its date parse is not new spend — actionReminder was already
running exactly that parse one layer down, and now skips it.

Slots.Text is deliberately not filled. Extract sets it to the raw
utterance, and a grammar that left it empty meant it: agendaQueryBuild
hands the query chain the sentence itself, and narrativeQueryBuild's Text
is the topic.

Fixture unchanged at 64/91 (70.3%) on TestONNXBaseline, no case regressed,
no new false clarify. What moved is the line the fixture calls "slots
deferred to daemon": 6 to 0.

Verified on homesrv: "напомни в 11:00 позвонить маме" now answers
"хорошо, напомню сегодня в 11:00."
This commit is contained in:
2026-08-06 01:08:38 +04:00
parent b6305f1b6e
commit 869580c913
3 changed files with 112 additions and 10 deletions
+39 -5
View File
@@ -80,6 +80,9 @@ func (r *Router) Route(ctx context.Context, utterance string, now time.Time) (De
continue // grammar matched shape but not content → fall through
}
d.Utterance = utterance
// The grammar decided the intent; the extractor fills the slots it did
// not match (V-572). See fillMatchedSlots for why every grammar gets it.
r.fillMatchedSlots(ctx, &d, now)
return d, nil
}
@@ -121,14 +124,37 @@ func (r *Router) Route(ctx context.Context, utterance string, now time.Time) (De
return d, nil
}
// fillSlots — run stage-2 extraction on an LLM decision and fill only the slots
// the model left empty. The LLM wins where it answered: it saw the sentence, the
// parsers are keyword tables. Extraction covers what the model cannot produce at
// all — a parsed reminder time and an allowlist fn.
// fillMatchedSlots — run stage-2 extraction over a decision some earlier
// claimant produced, and fill only the slots that claimant left empty. A
// matched value always wins: the claimant read the sentence, the extractor
// guesses from keyword tables.
//
// Shared by the stage-0 grammars and the LLM router, which had the same hole
// for the same reason. A grammar asserts an intent at confidence 1.0 and says
// nothing about the slots, so "напомни в 11:00 позвонить маме" arrived with
// HasTime false however plainly the hour was spoken, and the daemon read the
// silence as absence and asked "Когда?" (V-572). The alternative was ten
// grammars each re-implementing extraction.
//
// It is applied to every stage-0 decision rather than to a chosen few, because
// for every intent but reminder it is inert: Extract fills Time for a reminder,
// Fn for an act and Key for a fact, and nothing at all for query, system, note
// or chat, which is what the query, clock, agenda, feed, list, task and
// narrative rules emit. The act rules — wakeword-act and the Praxis ones —
// already carry an Fn or they do not match, so there is nothing left for the
// matcher to fill. The reminder rule is the one that gains, and its time parse
// is a cost the daemon was already paying one layer down in actionReminder.
//
// Slots.Text is deliberately NOT filled here. Extract sets it to the raw
// utterance, and a grammar that left it empty meant it: agendaQueryBuild hands
// the query chain the utterance itself, and narrativeQueryBuild's Text is the
// topic, not the sentence.
//
// If a reminder still has no time, leave it missing. The daemon then says it
// could not read the time; inventing one would set a wrong alarm.
func (r *Router) fillSlots(ctx context.Context, d *Decision, now time.Time) {
// Returns what the extractor read, so a caller that wants more of it does not
// pay for a second extraction — the reminder parser is the expensive one.
func (r *Router) fillMatchedSlots(ctx context.Context, d *Decision, now time.Time) Slots {
ex := r.extractor.Extract(ctx, d.Intent, d.Utterance, now)
if !d.Slots.HasTime && ex.HasTime {
d.Slots.Time, d.Slots.HasTime = ex.Time, ex.HasTime
@@ -139,6 +165,14 @@ func (r *Router) fillSlots(ctx context.Context, d *Decision, now time.Time) {
if !d.Slots.HasFn && ex.HasFn {
d.Slots.Fn, d.Slots.Args, d.Slots.HasFn = ex.Fn, ex.Args, ex.HasFn
}
return ex
}
// fillSlots — fillMatchedSlots for an LLM decision, plus the two backfills that
// only make sense there. The LLM wins where it answered: it saw the sentence,
// the parsers are keyword tables.
func (r *Router) fillSlots(ctx context.Context, d *Decision, now time.Time) {
ex := r.fillMatchedSlots(ctx, d, now)
// For an act the model returns the verb in Text ("restart nginx"), which is
// often cleaner than the raw utterance ("maven, could you restart nginx").
// Try it too when the utterance did not match the allowlist.
+67
View File
@@ -121,6 +121,73 @@ func TestStage0GrammarFiresThroughCyrillicWakeWord(t *testing.T) {
}
}
// TestStage0ReminderCarriesTheHourHeSaid — "напомни в 11:00 позвонить маме" is
// the commonest reminder there is, and it used to reach the daemon with HasTime
// false, because ReminderGrammar builds its slots by hand and the router ran no
// extraction over a stage-0 decision. The daemon read the silence as absence and
// asked "Когда?" about an hour he had just said (V-572).
func TestStage0ReminderCarriesTheHourHeSaid(t *testing.T) {
r := newTestRouter(t, 0.0)
r.grammars = append(r.grammars, ReminderGrammar())
d, err := r.Route(context.Background(), "напомни в 11:00 позвонить маме", refNow())
if err != nil {
t.Fatalf("route: %v", err)
}
if d.Stage != 0 || d.Intent != IntentReminder {
t.Fatalf("want stage0 reminder, got %+v", d)
}
if !d.Slots.HasTime {
t.Fatalf("the hour was spoken, so the slot must be filled: %+v", d.Slots)
}
if got, want := d.Slots.Time.Format("15:04"), "11:00"; got != want {
t.Errorf("fire time = %s, want %s", got, want)
}
// The subject is the grammar's, not the extractor's: Slots.Text is what she
// says at the hour, and Extract would have overwritten it with the sentence.
if d.Slots.Text != "в 11:00 позвонить маме" {
t.Errorf("Text = %q, want the grammar's capture", d.Slots.Text)
}
}
// TestStage0MatchedSlotBeatsTheExtractor — a grammar that matched a literal
// pattern outranks a parser that guessed. The wake-word act names its fn from
// the remainder after the wake token; extraction over the raw utterance must not
// be able to replace it.
func TestStage0MatchedSlotBeatsTheExtractor(t *testing.T) {
r := newTestRouter(t, 0.0)
d, err := r.Route(context.Background(), "maven, restart nginx", refNow())
if err != nil {
t.Fatalf("route: %v", err)
}
if d.Slots.Fn != "restart" || len(d.Slots.Args) != 1 || d.Slots.Args[0] != "nginx" {
t.Fatalf("matched fn was overwritten: %+v", d.Slots)
}
if d.Slots.Text != "restart nginx" {
t.Errorf("Text = %q, want the grammar's remainder", d.Slots.Text)
}
}
// TestStage0QueryKeepsAnEmptyText — agendaQueryBuild deliberately leaves Text
// empty so the query chain reads the utterance itself. Extraction fills Time,
// Key and Fn and never Text, or every stage-0 query would start carrying the
// whole sentence in a slot that means something narrower.
func TestStage0QueryKeepsAnEmptyText(t *testing.T) {
r := newTestRouter(t, 0.0)
r.grammars = append(r.grammars, AgendaQueryGrammars()...)
d, err := r.Route(context.Background(), "что у меня сегодня", refNow())
if err != nil {
t.Fatalf("route: %v", err)
}
if d.Stage != 0 || d.Intent != IntentQuery {
t.Fatalf("want stage0 query, got %+v", d)
}
if d.Slots.Text != "" {
t.Errorf("Text = %q, want it left empty", d.Slots.Text)
}
}
// ----------------------------- stage 1 ---------------------------------------
func TestStage1ClassifiesAct(t *testing.T) {
+6 -5
View File
@@ -88,11 +88,12 @@ func DefaultGrammars(actMatcher ActMatcher) []Grammar {
// non-reminder time queries toward it, and the verb+action overlap pushes
// actual reminders toward fact — a double contamination. Stage 0 fixes both.
//
// The grammar captures the part after "напомни"/"remind me" into Slots.Text
// so the daemon's time parser can extract the fire time from it. The grammar
// itself does NOT parse time — that's the extractor's job (stage 2), but
// stage 0 skips the extractor. The daemon's applyAction fallback calls the
// time parser for stage-0 reminders that arrive without HasTime.
// The grammar captures the part after "напомни"/"remind me" into Slots.Text
// what she says at the hour. The grammar itself does NOT parse time; that is
// the extractor's job, and since V-572 the router runs the extractor over a
// stage-0 decision too (fillMatchedSlots in router.go). Before that it did not,
// so "напомни в 11:00 позвонить маме" reached the daemon with HasTime false and
// was asked "Когда?" about an hour he had just said.
func ReminderGrammar() Grammar {
return Grammar{
Name: "reminder-wakeword",