newkirk
84a18f5866
- Added tier boundary descriptions (fast/standard/heavy) to the classifier system prompt so gpt-5.4-nano understands what each complexity band means - Added signal keywords for higher/lower ratings to reduce misclassification of simple requests as complex and vice versa
85 lines
2.9 KiB
Go
85 lines
2.9 KiB
Go
package router
|
|
|
|
import (
|
|
"context"
|
|
"fmt"
|
|
"strconv"
|
|
"strings"
|
|
|
|
"gophergate/internal/db"
|
|
)
|
|
|
|
const classifierSystemPrompt = `You are a task complexity classifier. Rate the following user message on a scale of 1 to %d, where:
|
|
1 = trivial/simple (basic facts, greetings, simple math)
|
|
%d = highly complex (multi-step reasoning, code generation, architecture design)
|
|
|
|
TIER BOUNDARIES:
|
|
1-3 (fast): Simple Q&A, classification, JSON parsing, short text, greetings, factual lookup, quick code snippets.
|
|
4-7 (standard): Summarize, explain, draft emails/docs, compare options, moderate coding, debug simple issues, translate.
|
|
8-10 (heavy): Multi-step reasoning, complex code generation, architecture design, agent orchestration, deep debugging, research analysis, long multi-file refactors.
|
|
|
|
SIGNALS for higher ratings: mentions of "architecture", "distributed", "agent", "multi-file", "system design", "think step by step", "race condition", "refactor entire", code blocks > 50 lines, requests asking to build/ship entire features.
|
|
SIGNALS for lower ratings: "what is", "how do I", "define", "list", single yes/no questions, requests under 20 words with no code.
|
|
|
|
Reply with ONLY the number. No explanation.`
|
|
|
|
func routeClassifier(ctx context.Context, classify ClassifierFunc, group db.ModelGroup, targets []string, routeCtx *RouteContext) (*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)
|
|
userMsg := ""
|
|
if routeCtx != nil {
|
|
userMsg = routeCtx.UserMessage
|
|
}
|
|
ratingStr, err := classify(ctx, getSelectorModel(group, targets), prompt, userMsg)
|
|
if err != nil {
|
|
// Classifier failed — fall back to heuristic
|
|
return routeHeuristic(group, targets, routeCtx)
|
|
}
|
|
|
|
rating, err := strconv.Atoi(strings.TrimSpace(ratingStr))
|
|
if err != nil || rating < 1 {
|
|
rating = 1
|
|
}
|
|
if rating > maxRating {
|
|
rating = maxRating
|
|
}
|
|
|
|
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",
|
|
Reason: fmt.Sprintf("complexity rating: %d/%d", rating, maxRating),
|
|
}, nil
|
|
}
|
|
|
|
func getSelectorModel(group db.ModelGroup, targets []string) string {
|
|
if group.SelectorModel != nil && *group.SelectorModel != "" {
|
|
return *group.SelectorModel
|
|
}
|
|
// Default: use the first (cheapest) target model as the selector
|
|
return targets[0]
|
|
}
|