Files
minstrel/internal/recsettings/service.go
T
bvandeusenandClaude Opus 5 799dab029a
test-go / test (push) Successful in 1m0s
test-go / integration (push) Failing after 4m55s
feat(discover): rank suggestions by taste-tag overlap — #2377 (server)
The payoff slice. Until now a candidate's only claim on a slot was "some
artist you play is adjacent to it in a similarity graph" — a fact that says
nothing about whether the music sounds like anything you like. Now the
candidate's own folksonomy tags (cached by slice 5) are compared against
the user's taste-profile tags, so the deck ranks on taste and can say WHY.

The blend is MULTIPLICATIVE — score × (1 + weight × overlap) — and that
choice carries the whole safety argument:

  - An untagged candidate has overlap 0, so its score is EXACTLY unchanged.
    Tag coverage is permanently partial (#2376); it must cost a candidate
    nothing, not sink it (rule #131).
  - Nothing can leapfrog on tags alone. An additive term with a large
    weight would let a near-zero-similarity artist outrank a strong match
    for sharing one popular tag, which reads as noise.
  - Weight 0 restores pure similarity order bit-for-bit, so the operator's
    knob has a real off position.

overlap = Σ(shared) candWeight × normalizedTasteWeight ÷ Σ(all) candWeight.
Normalizing the taste side by the user's strongest tag makes the score
comparable across users (taste weights accumulate with listening, so a
heavy listener's raw numbers dwarf a new user's while meaning the same
thing). Dividing by the candidate's own mass makes it comparable across
candidates, so a densely-tagged artist can't win on tag count alone.

Applied to the whole over-fetched pool BEFORE selectSuggestions, so the
rotation and diversity rules operate on blended scores — boosting only the
twelve already chosen by similarity would leave the re-ranking undone.

A query failure is returned, NOT degraded past. Graceful degradation is
for expected absence (no taste profile, no cached tags) and both are
handled explicitly as empty inputs; swallowing a real error would hide a
broken DB behind a subtly worse ranking that nothing reports.

Migration 0051 adds a FOURTH tuning scope rather than columns on
taste_tuning, because snooze_days lives here too and a snooze must never
be read as taste signal (#2374) — filing it under 'taste' would put it one
careless join from the leak that design forbids. Expanding
recommendation_tuning_audit's CHECK is in the same migration per rule #36,
and a test asserts the audit row lands, which is what would catch its
absence.

snooze_days moves out of a Go constant onto the tuning card (rule #25),
closing the deferral from #2374.

Tag-overlap tests use deliberately SKEWED fixtures: an evenly-matching pool
cannot exercise a re-ranking, since every candidate gets the same
multiplier and the order is unchanged whether the blend works or not.

Admin UI + client attribution follow in this batch — rule #27.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
2026-08-02 20:31:12 -04:00

501 lines
16 KiB
Go
Raw Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
// Package recsettings is the DB-backed home of the recommendation
// tuning knobs (#1250): the two scoring-weight profiles (radio /
// daily_mix) and the taste-profile build settings (engagement
// half-life + completion curve). It follows the coverart
// SettingsService pattern — boot reconcile seeds shipped defaults for
// missing rows, values are cached under a RWMutex, and every change
// takes effect live (the daily_mix profile + taste config are pushed
// into package playlists; radio reads the cache per request).
//
// Framing (decision #1247): this is the defaults-discovery lab. The
// operator turns knobs to FIND good values; found-good values get
// baked back into the Shipped* functions below as new shipped
// defaults. End users and other operators should never need the card.
package recsettings
import (
"context"
"encoding/json"
"fmt"
"log/slog"
"sort"
"sync"
"github.com/jackc/pgx/v5/pgxpool"
"git.fabledsword.com/bvandeusen/minstrel/internal/db/dbq"
"git.fabledsword.com/bvandeusen/minstrel/internal/playlists"
"git.fabledsword.com/bvandeusen/minstrel/internal/recommendation"
"git.fabledsword.com/bvandeusen/minstrel/internal/taste"
)
// Scope names — the three tunable groups. radio + daily_mix are weight
// profiles; taste is the profile-build settings singleton.
const (
ScopeRadio = "radio"
ScopeDailyMix = "daily_mix"
ScopeTaste = "taste"
// ScopeDiscover is the Discover request surface (#2377). Its own scope
// rather than columns on taste: SnoozeDays lives here, and a snooze must
// never be read as taste signal (#2374) — filing it under taste would put
// it one careless join from the leak that design forbids.
ScopeDiscover = "discover"
)
// TasteTuning is the tunable subset of taste.Config: the engagement
// half-life, the completion→engagement curve points, the enriched-tag
// weight (how much folksonomy tags count vs raw ID3 genre, #1520), and the
// era-facet weight (how strongly a decade-play imprints, #1530).
type TasteTuning struct {
HalfLifeDays float64
EngagementHardSkip float64
EngagementNeutral float64
EngagementFull float64
EnrichedTagScale float64
EraScale float64
MoodScale float64
}
// ShippedRadioWeights are the shipped radio-profile defaults (moved
// here from config.RecommendationConfig — YAML is bootstrap-only,
// rule: config in UI). Radio is seed-directed (the user picked a
// direction), so taste is a lighter nudge than in the daily mixes.
// ContextTimeWeight starts uniform (1.0) across both profiles pending
// trend data (#1531); split them once the metrics view justifies it.
func ShippedRadioWeights() recommendation.ScoringWeights {
return recommendation.ScoringWeights{
BaseWeight: 1.0,
LikeBoost: 2.0,
RecencyWeight: 1.0,
SkipPenalty: 1.0,
JitterMagnitude: 0.1,
ContextWeight: 2.0,
SimilarityWeight: 2.0,
TasteWeight: 1.0,
ContextTimeWeight: 1.0,
}
}
// ShippedDailyMixWeights are the shipped daily_mix-profile defaults.
// Must stay in sync with the pre-push literal in playlists/system.go.
func ShippedDailyMixWeights() recommendation.ScoringWeights {
return recommendation.ScoringWeights{
BaseWeight: 1.0,
LikeBoost: 2.0,
RecencyWeight: 1.0,
SkipPenalty: 2.0,
JitterMagnitude: 0.1,
ContextWeight: 0.5,
SimilarityWeight: 1.5,
TasteWeight: 1.5,
ContextTimeWeight: 1.0,
}
}
// DiscoverTuning is the tunable set for the Discover request surface (#2377).
type DiscoverTuning struct {
// TagOverlapWeight scales the taste-tag term: score × (1 + w × overlap).
// 0 disables it and restores pure similarity ranking.
TagOverlapWeight float64
// SnoozeDays is the default "not right now" duration (#2374).
SnoozeDays float64
}
// ShippedDiscoverTuning are the shipped Discover defaults.
//
// TagOverlapWeight 1.0 lets a perfect tag match at most double a candidate's
// similarity score — enough to reorder the deck meaningfully, not enough for a
// popular-tag coincidence to beat a genuinely strong similarity match. It is a
// starting point for the tuning lab, not a tuned value: the honest way to pick
// it is the metrics trend view after some real use.
//
// SnoozeDays 90 matches the operator's approved shape: long enough that a
// parked suggestion stops nagging, short enough that a taste shift brings it
// back on its own.
func ShippedDiscoverTuning() DiscoverTuning {
return DiscoverTuning{
TagOverlapWeight: 1.0,
SnoozeDays: 90,
}
}
// ShippedTasteTuning mirrors taste.DefaultConfig's tunable subset.
func ShippedTasteTuning() TasteTuning {
d := taste.DefaultConfig()
return TasteTuning{
HalfLifeDays: d.HalfLifeDays,
EngagementHardSkip: d.Engagement.HardSkip,
EngagementNeutral: d.Engagement.NeutralCompletion,
EngagementFull: d.Engagement.FullCompletion,
EnrichedTagScale: d.EnrichedTagScale,
EraScale: d.EraScale,
MoodScale: d.MoodScale,
}
}
// Service caches the tuning values and owns their DB persistence +
// audit trail. Construct with New at boot.
type Service struct {
pool *pgxpool.Pool
logger *slog.Logger
mu sync.RWMutex
profiles map[string]recommendation.ScoringWeights
taste TasteTuning
discover DiscoverTuning
}
// New boots the service: seeds shipped defaults for missing rows,
// loads the current values, and pushes the daily_mix weights + taste
// config into package playlists so the daily builds pick them up.
func New(ctx context.Context, pool *pgxpool.Pool, logger *slog.Logger) (*Service, error) {
s := &Service{
pool: pool,
logger: logger,
profiles: map[string]recommendation.ScoringWeights{},
}
if err := s.reconcile(ctx); err != nil {
return nil, fmt.Errorf("recsettings boot: %w", err)
}
return s, nil
}
// reconcile seeds missing rows with shipped defaults, reads everything
// back into the cache, and pushes the playlist-side values.
func (s *Service) reconcile(ctx context.Context) error {
q := dbq.New(s.pool)
for profile, w := range map[string]recommendation.ScoringWeights{
ScopeRadio: ShippedRadioWeights(),
ScopeDailyMix: ShippedDailyMixWeights(),
} {
if err := q.UpsertWeightProfileDefaults(ctx, upsertParams(profile, w)); err != nil {
return fmt.Errorf("seed profile %q: %w", profile, err)
}
}
st := ShippedTasteTuning()
if err := q.UpsertTasteTuningDefaults(ctx, dbq.UpsertTasteTuningDefaultsParams{
HalfLifeDays: st.HalfLifeDays,
EngagementHardSkip: st.EngagementHardSkip,
EngagementNeutral: st.EngagementNeutral,
EngagementFull: st.EngagementFull,
EnrichedTagScale: st.EnrichedTagScale,
EraScale: st.EraScale,
MoodScale: st.MoodScale,
}); err != nil {
return fmt.Errorf("seed taste tuning: %w", err)
}
sd := ShippedDiscoverTuning()
if err := q.UpsertDiscoverTuningDefaults(ctx, dbq.UpsertDiscoverTuningDefaultsParams{
TagOverlapWeight: sd.TagOverlapWeight,
SnoozeDays: sd.SnoozeDays,
}); err != nil {
return fmt.Errorf("seed discover tuning: %w", err)
}
rows, err := q.ListWeightProfiles(ctx)
if err != nil {
return fmt.Errorf("list weight profiles: %w", err)
}
tt, err := q.GetTasteTuning(ctx)
if err != nil {
return fmt.Errorf("get taste tuning: %w", err)
}
dt, err := q.GetDiscoverTuning(ctx)
if err != nil {
return fmt.Errorf("get discover tuning: %w", err)
}
s.mu.Lock()
s.profiles = map[string]recommendation.ScoringWeights{}
for _, r := range rows {
s.profiles[r.Profile] = weightsFromRow(r)
}
s.taste = TasteTuning{
HalfLifeDays: tt.HalfLifeDays,
EngagementHardSkip: tt.EngagementHardSkip,
EngagementNeutral: tt.EngagementNeutral,
EngagementFull: tt.EngagementFull,
EnrichedTagScale: tt.EnrichedTagScale,
EraScale: tt.EraScale,
MoodScale: tt.MoodScale,
}
s.discover = DiscoverTuning{
TagOverlapWeight: dt.TagOverlapWeight,
SnoozeDays: dt.SnoozeDays,
}
s.mu.Unlock()
s.push()
return nil
}
// push installs the daily-build values into package playlists (the
// coverart Configure() pattern). Radio needs no push — its handler
// reads Weights(ScopeRadio) per request.
func (s *Service) push() {
playlists.SetSystemMixWeights(s.Weights(ScopeDailyMix))
playlists.SetTasteConfig(s.TasteConfig())
}
// Weights returns the cached weights for a profile scope. Unknown
// scopes return the shipped radio defaults (defensive; callers pass
// the Scope* constants).
func (s *Service) Weights(profile string) recommendation.ScoringWeights {
s.mu.RLock()
defer s.mu.RUnlock()
if w, ok := s.profiles[profile]; ok {
return w
}
return ShippedRadioWeights()
}
// Taste returns the cached taste-tuning values.
func (s *Service) Taste() TasteTuning {
s.mu.RLock()
defer s.mu.RUnlock()
return s.taste
}
// Discover returns the cached Discover-tuning values. Read per request by the
// suggestions handler, so an admin change takes effect on the next refresh
// with no restart (rule #25).
func (s *Service) Discover() DiscoverTuning {
s.mu.RLock()
defer s.mu.RUnlock()
return s.discover
}
// TasteConfig assembles the full taste.Config the profile builder
// consumes: shipped non-tunable knobs (like bonuses, floors, caps)
// plus the tuned half-life and curve. WindowDays scales with the
// half-life at the shipped ratio (270/75 = 3.6 half-lives) — the
// window is a query-cost bound, not an independent knob.
func (s *Service) TasteConfig() taste.Config {
t := s.Taste()
cfg := taste.DefaultConfig()
cfg.HalfLifeDays = t.HalfLifeDays
cfg.WindowDays = t.HalfLifeDays * 3.6
cfg.Engagement = taste.EngagementParams{
HardSkip: t.EngagementHardSkip,
NeutralCompletion: t.EngagementNeutral,
FullCompletion: t.EngagementFull,
}
cfg.EnrichedTagScale = t.EnrichedTagScale
cfg.EraScale = t.EraScale
cfg.MoodScale = t.MoodScale
return cfg
}
// fieldChange is one entry of an audit row's changes array.
type fieldChange struct {
Field string `json:"field"`
Old float64 `json:"old"`
New float64 `json:"new"`
}
// UpdateProfile applies a partial update to one weight profile.
// Unknown fields and out-of-range values reject the whole patch. A
// no-op patch (all values equal to current) writes no audit row.
func (s *Service) UpdateProfile(ctx context.Context, profile string, patch map[string]float64) error {
if profile != ScopeRadio && profile != ScopeDailyMix {
return fmt.Errorf("%w: %q", ErrUnknownScope, profile)
}
current := s.Weights(profile)
next, changes, err := applyWeightPatch(current, patch)
if err != nil {
return err
}
if len(changes) == 0 {
return nil
}
return s.persistProfile(ctx, profile, next, "update", changes)
}
// UpdateTaste applies a partial update to the taste tuning singleton.
func (s *Service) UpdateTaste(ctx context.Context, patch map[string]float64) error {
current := s.Taste()
next, changes, err := applyTastePatch(current, patch)
if err != nil {
return err
}
if len(changes) == 0 {
return nil
}
return s.persistTaste(ctx, next, "update", changes)
}
// UpdateDiscover applies a partial update to the Discover tuning singleton.
func (s *Service) UpdateDiscover(ctx context.Context, patch map[string]float64) error {
current := s.Discover()
next, changes, err := applyDiscoverPatch(current, patch)
if err != nil {
return err
}
if len(changes) == 0 {
return nil
}
return s.persistDiscover(ctx, next, "update", changes)
}
// Reset restores a scope to its shipped defaults, with one audit row
// carrying the full diff. A scope already at defaults is a no-op.
func (s *Service) Reset(ctx context.Context, scope string) error {
switch scope {
case ScopeRadio, ScopeDailyMix:
shipped := ShippedRadioWeights()
if scope == ScopeDailyMix {
shipped = ShippedDailyMixWeights()
}
changes := diffWeights(s.Weights(scope), shipped)
if len(changes) == 0 {
return nil
}
return s.persistProfile(ctx, scope, shipped, "reset", changes)
case ScopeTaste:
shipped := ShippedTasteTuning()
changes := diffTaste(s.Taste(), shipped)
if len(changes) == 0 {
return nil
}
return s.persistTaste(ctx, shipped, "reset", changes)
case ScopeDiscover:
shipped := ShippedDiscoverTuning()
changes := diffDiscover(s.Discover(), shipped)
if len(changes) == 0 {
return nil
}
return s.persistDiscover(ctx, shipped, "reset", changes)
default:
return fmt.Errorf("%w: %q", ErrUnknownScope, scope)
}
}
// persistProfile writes the profile row + audit entry, refreshes the
// cache, and pushes daily-build values.
func (s *Service) persistProfile(
ctx context.Context, profile string, w recommendation.ScoringWeights,
action string, changes []fieldChange,
) error {
q := dbq.New(s.pool)
if _, err := q.UpdateWeightProfile(ctx, updateParams(profile, w)); err != nil {
return fmt.Errorf("update profile %q: %w", profile, err)
}
if err := s.audit(ctx, q, profile, action, changes); err != nil {
return err
}
s.mu.Lock()
s.profiles[profile] = w
s.mu.Unlock()
s.push()
return nil
}
func (s *Service) persistTaste(
ctx context.Context, t TasteTuning, action string, changes []fieldChange,
) error {
q := dbq.New(s.pool)
if _, err := q.UpdateTasteTuning(ctx, dbq.UpdateTasteTuningParams{
HalfLifeDays: t.HalfLifeDays,
EngagementHardSkip: t.EngagementHardSkip,
EngagementNeutral: t.EngagementNeutral,
EngagementFull: t.EngagementFull,
EnrichedTagScale: t.EnrichedTagScale,
EraScale: t.EraScale,
MoodScale: t.MoodScale,
}); err != nil {
return fmt.Errorf("update taste tuning: %w", err)
}
if err := s.audit(ctx, q, ScopeTaste, action, changes); err != nil {
return err
}
s.mu.Lock()
s.taste = t
s.mu.Unlock()
s.push()
return nil
}
// persistDiscover writes the discover row + audit entry and refreshes the
// cache. No push(): unlike taste and daily_mix, nothing precomputes from these
// — the suggestions handler reads Discover() per request.
func (s *Service) persistDiscover(
ctx context.Context, d DiscoverTuning, action string, changes []fieldChange,
) error {
q := dbq.New(s.pool)
if _, err := q.UpdateDiscoverTuning(ctx, dbq.UpdateDiscoverTuningParams{
TagOverlapWeight: d.TagOverlapWeight,
SnoozeDays: d.SnoozeDays,
}); err != nil {
return fmt.Errorf("update discover tuning: %w", err)
}
if err := s.audit(ctx, q, ScopeDiscover, action, changes); err != nil {
return err
}
s.mu.Lock()
s.discover = d
s.mu.Unlock()
return nil
}
// audit writes one recommendation_tuning_audit row. Changes are
// sorted by field so rows are deterministic and diff-friendly.
func (s *Service) audit(
ctx context.Context, q *dbq.Queries, scope, action string, changes []fieldChange,
) error {
sort.Slice(changes, func(i, j int) bool { return changes[i].Field < changes[j].Field })
payload, err := json.Marshal(changes)
if err != nil {
return fmt.Errorf("marshal audit changes: %w", err)
}
if err := q.InsertTuningAudit(ctx, dbq.InsertTuningAuditParams{
Scope: scope, Action: action, Changes: payload,
}); err != nil {
return fmt.Errorf("insert audit row: %w", err)
}
return nil
}
func upsertParams(profile string, w recommendation.ScoringWeights) dbq.UpsertWeightProfileDefaultsParams {
return dbq.UpsertWeightProfileDefaultsParams{
Profile: profile,
BaseWeight: w.BaseWeight,
LikeBoost: w.LikeBoost,
RecencyWeight: w.RecencyWeight,
SkipPenalty: w.SkipPenalty,
JitterMagnitude: w.JitterMagnitude,
ContextWeight: w.ContextWeight,
SimilarityWeight: w.SimilarityWeight,
TasteWeight: w.TasteWeight,
ContextTimeWeight: w.ContextTimeWeight,
}
}
func updateParams(profile string, w recommendation.ScoringWeights) dbq.UpdateWeightProfileParams {
return dbq.UpdateWeightProfileParams{
Profile: profile,
BaseWeight: w.BaseWeight,
LikeBoost: w.LikeBoost,
RecencyWeight: w.RecencyWeight,
SkipPenalty: w.SkipPenalty,
JitterMagnitude: w.JitterMagnitude,
ContextWeight: w.ContextWeight,
SimilarityWeight: w.SimilarityWeight,
TasteWeight: w.TasteWeight,
ContextTimeWeight: w.ContextTimeWeight,
}
}
func weightsFromRow(r dbq.RecommendationWeightProfile) recommendation.ScoringWeights {
return recommendation.ScoringWeights{
BaseWeight: r.BaseWeight,
LikeBoost: r.LikeBoost,
RecencyWeight: r.RecencyWeight,
SkipPenalty: r.SkipPenalty,
JitterMagnitude: r.JitterMagnitude,
ContextWeight: r.ContextWeight,
SimilarityWeight: r.SimilarityWeight,
TasteWeight: r.TasteWeight,
ContextTimeWeight: r.ContextTimeWeight,
}
}