feat: classifier bucket mapping + dispatcher seed group
CI / Lint (push) Has been cancelled
CI / Test (push) Has been cancelled
CI / Build (push) Has been cancelled

Classifier: When complexity_threshold is set (e.g. 10), uses it as the
rating scale and maps ratings proportionally to target buckets instead
of 1:1. Formula: idx = rating * len(targets) / (threshold + 1).

With threshold=10 and 3 targets: 1-3→target[0], 4-7→target[1], 8-10→target[2].

Seed: Added 'dispatcher' group (classifier, threshold=10, selector=deepseek-v4-flash)
that auto-routes to fast-flow/standard-pro/heavy-logic by complexity score.

Combined with hierarchical routing, this enables two-level dispatch:
  dispatcher scores 1-10 → routes to tier group → tier picks concrete model.
This commit is contained in:
2026-05-07 13:18:35 -04:00
parent 7517307c11
commit 3c0b59622e
2 changed files with 39 additions and 11 deletions
+20 -1
View File
@@ -16,11 +16,19 @@ const classifierSystemPrompt = `You are a task complexity classifier. Rate the f
Reply with ONLY the number. No explanation.`
func routeClassifier(ctx context.Context, classify ClassifierFunc, group db.ModelGroup, targets []string, userMessage string) (*Decision, error) {
// Determine the rating scale
maxRating := len(targets)
if maxRating < 2 {
maxRating = 2
}
// When complexity_threshold is set, use it as a wider scale (e.g., 1-10)
// and map ratings proportionally to target buckets.
bucketMode := group.ComplexityThreshold != nil && *group.ComplexityThreshold > 0
if bucketMode {
maxRating = *group.ComplexityThreshold
}
prompt := fmt.Sprintf(classifierSystemPrompt, maxRating, maxRating)
ratingStr, err := classify(ctx, getSelectorModel(group, targets), prompt, userMessage)
if err != nil {
@@ -36,7 +44,18 @@ func routeClassifier(ctx context.Context, classify ClassifierFunc, group db.Mode
rating = maxRating
}
idx := rating - 1 // 0-based index into targets
var idx int
if bucketMode {
// Proportional mapping: wider scale → N target buckets
// e.g., threshold=10, 3 targets: 1-3→0, 4-7→1, 8-10→2
idx = rating * len(targets) / (maxRating + 1)
if idx >= len(targets) {
idx = len(targets) - 1
}
} else {
idx = rating - 1 // 1:1 mapping
}
return &Decision{
SelectedModel: targets[idx],
Strategy: "classifier",