xinmi/internal/workbench/manuscript.go

package workbench

import (
	"context"
	"fmt"
	"strings"

	"github.com/Tencent/WeKnora/internal/accesspolicy"
	"github.com/Tencent/WeKnora/internal/agent/token"
	"github.com/Tencent/WeKnora/internal/models/chat"
	"github.com/Tencent/WeKnora/internal/types"
)

const qualitySystem = `你是内置跨稿质检员。审查已采用正文的逻辑、先修依赖、衔接、重复和论据一致性。严重逻辑断裂或重复为 blocking，措辞建议为 suggestion。不得编造问题或审查范围。严格输出 JSON {"passed":true,"summary":"中文总结","issues":[{"segment_ids":["实际段落编号"],"severity":"blocking 或 suggestion","problem":"具体问题","suggestion":"具体修改建议"}]}。无 blocking 时 passed 必须为 true，否则 false。`

func manuscriptPrompt(segments []Segment, advice string) string {
	return "本次仅审查以下实际采用的正文。用户可以主动跳过内容段，编号不连续或段数少于原计划本身不是缺陷，不得要求恢复已跳过段；若确有未完成承诺，应建议在已采用段内补足或修正预告。\n已采用正文：" + compact(segments) + "\n用户补充意见：" + advice
}

// Every accepted segment is reviewed in full, and adjacent windows overlap by
// one complete segment. A single segment that cannot fit still fails explicitly.
func manuscriptWindows(c *completionRequest, segments []Segment, advice string) ([][]Segment, error) {
	fits := func(rows []Segment) bool {
		messages := append(c.messages[:0:0], c.messages...)
		messages[1].Content = manuscriptPrompt(rows, advice)
		return c.estimator.EstimateMessages(messages) <= c.inputLimit
	}
	if fits(segments) {
		return [][]Segment{segments}, nil
	}
	windows := [][]Segment{}
	current := []Segment{}
	for _, segment := range segments {
		if !fits([]Segment{segment}) {
			return nil, fmt.Errorf("%w: 模型上下文容量不足以容纳单个内容段，请为质检选择更大上下文模型或拆分该段", ErrInvalid)
		}
		if len(current) > 0 && !fits(append(append([]Segment(nil), current...), segment)) {
			windows = append(windows, current)
			current = []Segment{current[len(current)-1]}
			if !fits(append(current, segment)) {
				return nil, fmt.Errorf("%w: 相邻正文的上下文容量不足，无法完整检查衔接，请选择更大上下文模型", ErrInvalid)
			}
		}
		current = append(current, segment)
	}
	if len(current) > 0 {
		windows = append(windows, current)
	}
	return windows, nil
}

func mergeQualityReports(reports []QualityReport, accepted []Segment) (*QualityReport, error) {
	merged := &QualityReport{Passed: true, Issues: []QualityIssue{}}
	seen := map[string]bool{}
	for i, report := range reports {
		merged.Summary += fmt.Sprintf("检查 %d：%s\n", i+1, report.Summary)
		for _, issue := range report.Issues {
			key := compact(issue)
			if !seen[key] {
				seen[key] = true
				merged.Issues = append(merged.Issues, issue)
			}
			if issue.Severity == "blocking" {
				merged.Passed = false
			}
		}
	}
	if err := validateQuality(merged, accepted); err != nil {
		return nil, err
	}
	return merged, nil
}

func (s *Service) prepareQualityCompletion(ctx context.Context, agentID string) (*completionRequest, error) {
	if !accesspolicy.Allowed(ctx, "content.review") {
		return nil, ErrForbidden
	}
	snapshot, frozen := snapshotFromContext(ctx, agentID)
	if !frozen || snapshot.ModelID == "" {
		return s.prepareCompletion(ctx, agentID, qualitySystem, "")
	}
	// Use the task's recorded model for review, without loading private skills,
	// persona, memory, or knowledge that the reviewer cannot otherwise access.
	metadata, err := s.models.GetModelByID(ctx, snapshot.ModelID)
	if err != nil {
		return nil, err
	}
	if metadata == nil {
		return nil, fmt.Errorf("%w: 任务的质检模型已不可用，请更新任务配置", ErrInvalid)
	}
	estimator, err := token.NewEstimator()
	if err != nil {
		return nil, err
	}
	contextSize := types.AgentMaxContextTokens(0, metadata.Parameters.ContextWindow)
	if snapshot.ContextWindow > 0 {
		contextSize = min(contextSize, snapshot.ContextWindow)
	}
	outputLimit := snapshot.OutputLimit
	if outputLimit <= 0 {
		outputLimit = types.DefaultMaxCompletionTokens(types.AgentModeQuickAnswer, "")
	}
	if metadata.Parameters.MaxOutputTokens > 0 {
		outputLimit = min(outputLimit, metadata.Parameters.MaxOutputTokens)
	}
	agent := &types.CustomAgent{ID: agentID, Config: types.CustomAgentConfig{ModelID: snapshot.ModelID, Temperature: snapshot.Temperature, Thinking: snapshot.Thinking}}
	c := &completionRequest{agent: agent, estimator: estimator, contextSize: contextSize,
		inputLimit: contextSize - contextSize*15/100 - outputLimit, outputLimit: outputLimit,
		messages: []chat.Message{{Role: "system", Content: qualitySystem}, {Role: "user"}}, jsonOutput: true}
	return c, c.validateBudget()
}

func (s *Service) reviewManuscript(ctx context.Context, agentID string, accepted []Segment, advice string) (*QualityReport, error) {
	c, err := s.prepareQualityCompletion(ctx, agentID)
	if err != nil {
		return nil, err
	}
	windows, err := manuscriptWindows(c, accepted, advice)
	if err != nil {
		return nil, err
	}
	reports := []QualityReport{}
	for i, window := range windows {
		if lease, ok := ctx.Value(leaseContextKey{}).(jobLease); ok {
			if err := s.progress(ctx, lease.ID, 40+40*i/len(windows), fmt.Sprintf("跨稿质检：第 %d/%d 组，完整检查正文及相邻衔接", i+1, len(windows))); err != nil {
				return nil, err
			}
		}
		c.messages[1].Content = manuscriptPrompt(window, advice)
		c.jsonOutput = true
		text, err := s.runCompletion(ctx, c)
		if err != nil {
			return nil, err
		}
		report, err := parseJSON[QualityReport](text)
		if err != nil {
			return nil, err
		}
		if err := validateQuality(&report, window); err != nil {
			return nil, err
		}
		reports = append(reports, report)
	}
	report, err := mergeQualityReports(reports, accepted)
	if err != nil {
		return nil, err
	}
	if len(windows) > 1 {
		report.Summary = fmt.Sprintf("共 %d 组，全部已采用正文逐组完整质检，相邻组保留完整衔接段；远距离语义一致性仍需结合整稿复核。\n", len(windows)) + report.Summary
	}
	return report, nil
}

func composeAccepted(accepted []Segment) string {
	parts := make([]string, 0, len(accepted))
	for _, segment := range accepted {
		parts = append(parts, segment.Content)
	}
	return strings.Join(parts, "\n\n")
}

func (s *Service) composeManuscript(ctx context.Context, agentID string, accepted []Segment, advice string) (string, error) {
	if strings.TrimSpace(advice) != "" {
		return "", fmt.Errorf("%w: 组合成稿保留原文；请先在对应内容段应用修改建议并重新质检", ErrInvalid)
	}
	return composeAccepted(accepted), nil
}