glpi-neural-brain/internal/engine/article_cpu_quality.go

package engine

import (
	"context"
	"errors"
	"fmt"
	"math"
	"strings"
	"time"

	"github.com/local/glpi-neural-brain/internal/articlequality"
	"github.com/local/glpi-neural-brain/internal/model"
	"github.com/local/glpi-neural-brain/internal/sourceagent"
)

type articleCPUQualityExecution struct {
	Result         articlequality.Result
	Offloaded      bool
	AgentID        string
	ComputeMS      int64
	FallbackReason string
}

func buildArticleCPUQualityRequest(draft model.KnowledgeArticleDraft, articleType string, sources []articleSource, research []model.ResearchResult, language string) articlequality.Request {
	docs := make([]articlequality.Document, 0, len(sources)+len(research))
	for _, src := range sources {
		text := articleQualitySample(strings.TrimSpace(src.Node.Label+"\n"+src.Content), 20000)
		if text != "" {
			docs = append(docs, articlequality.Document{ID: src.Node.ID, Text: text})
		}
	}
	for i, r := range research {
		text := strings.TrimSpace(r.Content)
		if text == "" {
			text = strings.TrimSpace(r.Snippet)
		}
		if text == "" {
			continue
		}
		id := fmt.Sprintf("R%d", i+1)
		if strings.TrimSpace(r.URL) != "" {
			id = r.URL
		}
		docs = append(docs, articlequality.Document{ID: id, Text: articleQualitySample(strings.TrimSpace(r.Title+"\n"+text), 20000)})
	}
	return articlequality.Request{ArticleType: articleType, Title: draft.Title, Problem: draft.Text, Answer: formatArticleAnswer(draft, language), Prerequisites: draft.Prerequisites, Validation: draft.Validation, Troubleshoot: draft.Troubleshooting, Categories: draft.Categories, Keywords: draft.Keywords, Sources: docs}
}

func articleQualitySample(value string, maxRunes int) string {
	value = strings.TrimSpace(value)
	if maxRunes <= 0 {
		maxRunes = 20000
	}
	r := []rune(value)
	if len(r) <= maxRunes {
		return value
	}
	// Preserve context from the beginning, middle and end instead of silently
	// biasing the lexical metric toward only the first section of long sources.
	head := maxRunes * 2 / 5
	middle := maxRunes * 3 / 10
	tail := maxRunes - head - middle
	midStart := len(r)/2 - middle/2
	if midStart < head {
		midStart = head
	}
	if midStart+middle > len(r)-tail {
		midStart = len(r) - tail - middle
	}
	return strings.TrimSpace(string(r[:head]) + "\n[… Stichprobe Mitte …]\n" + string(r[midStart:midStart+middle]) + "\n[… Stichprobe Ende …]\n" + string(r[len(r)-tail:]))
}

func validateArticleCPUQualityResult(r articlequality.Result) error {
	if r.Algorithm != articlequality.Algorithm {
		return fmt.Errorf("unexpected article quality algorithm %q", r.Algorithm)
	}
	for n, v := range map[string]float64{"score": r.Score, "lexical_diversity": r.LexicalDiversity, "redundancy": r.Redundancy, "evidence_alignment": r.EvidenceAlignment, "source_utilization": r.SourceUtilization, "technical_specificity": r.TechnicalSpecificity, "type_depth_score": r.TypeDepthScore} {
		if math.IsNaN(v) || math.IsInf(v, 0) || v < 0 || v > 1.000001 {
			return fmt.Errorf("invalid article quality metric %s=%v", n, v)
		}
	}
	if r.WordCount < 0 || r.SectionCount < 0 || r.ParagraphCount < 0 {
		return errors.New("invalid article quality counters")
	}
	return nil
}
func (e *Engine) evaluateArticleCPUQuality(ctx context.Context, articleRunID string, draft model.KnowledgeArticleDraft, articleType string, sources []articleSource, research []model.ResearchResult) (articleCPUQualityExecution, error) {
	payload := buildArticleCPUQualityRequest(draft, articleType, sources, research, e.Cfg.ArticleLanguage)
	if !e.Cfg.ArticleCPUQualityEnabled {
		r := articlequality.Evaluate(payload)
		r.Passed = true
		return articleCPUQualityExecution{Result: r, FallbackReason: "disabled_gate_observation_only"}, nil
	}
	if !e.Cfg.ArticleCPUQualityAgentOffload || e.SourceInbox == nil {
		r := articlequality.Evaluate(payload)
		return articleCPUQualityExecution{Result: r}, validateArticleCPUQualityResult(r)
	}
	checkCtx, cancel := context.WithTimeout(ctx, 2*time.Second)
	hasAgent, checkErr := e.SourceInbox.HasOnlineComputeAgent(checkCtx, sourceagent.ComputeKindArticleQuality, 3*time.Minute)
	cancel()
	if checkErr != nil || !hasAgent {
		reason := "no_article_quality_compute_agent"
		if checkErr != nil {
			reason = "article_quality_agent_check_failed: " + checkErr.Error()
		}
		if e.Cfg.ArticleCPUQualityAgentRequired {
			return articleCPUQualityExecution{FallbackReason: reason}, errors.New(reason)
		}
		r := articlequality.Evaluate(payload)
		return articleCPUQualityExecution{Result: r, FallbackReason: reason}, validateArticleCPUQualityResult(r)
	}
	jobCtx, cancelJob := context.WithTimeout(ctx, e.Cfg.ArticleCPUQualityAgentWait)
	result, err := e.SourceInbox.SubmitArticleQualityJob(jobCtx, sourceagent.ArticleQualityComputeRequest{Payload: payload})
	cancelJob()
	if err == nil {
		// Reconstruct pass/fail and recommendations locally. The agent contributes
		// bounded math only; it cannot inject arbitrary author instructions.
		result.Quality = articlequality.NormalizeResult(payload, result.Quality)
		err = validateArticleCPUQualityResult(result.Quality)
	}
	if err != nil {
		reason := err.Error()
		if e.Cfg.ArticleCPUQualityAgentRequired {
			return articleCPUQualityExecution{FallbackReason: reason}, err
		}
		r := articlequality.Evaluate(payload)
		return articleCPUQualityExecution{Result: r, FallbackReason: reason}, validateArticleCPUQualityResult(r)
	}
	return articleCPUQualityExecution{Result: result.Quality, Offloaded: true, AgentID: result.AgentID, ComputeMS: result.DurationMS}, nil
}

func articleCPUQualityBetter(a, b articlequality.Result) bool {
	if a.Passed != b.Passed {
		return a.Passed
	}
	if len(a.HardFailures) != len(b.HardFailures) {
		return len(a.HardFailures) < len(b.HardFailures)
	}
	if math.Abs(a.Score-b.Score) > 0.000001 {
		return a.Score > b.Score
	}
	if a.WordCount != b.WordCount {
		return a.WordCount > b.WordCount
	}
	if a.SectionCount != b.SectionCount {
		return a.SectionCount > b.SectionCount
	}
	return a.SourceUtilization > b.SourceUtilization
}