refactor: output channel interface (payload/meta/type) + memory fixes

- Redesign output_send__ tools: content JSON string -> structured
  payload/meta/type params for LLM reliability
- executeOutputSendTool: route by type with capability check
- executeOutputSendHelp: show meta format + type enum
- Updated system prompt rules for new interface
- docToTriples: use jieba exact mode adjacent co-occurrence
- Unify vector space: Doc.Vector field, ContextToDoc vectorizer,
  ReindexWithVectorizer on startup

internal/memory/cut.go

@@ -117,3 +117,26 @@ func ExtractKeywords(text string) []string {
 	}
 	return keywords
 }
+
+// CutExact 精确模式分词：返回去停用词后的所有有义项（不限数量），用于 doc→graph 蒸馏
+func CutExact(text string) []string {
+	text = CleanTemplateText(text)
+	x := GetJieba()
+	if x == nil {
+		return nil
+	}
+	words := x.Cut(text, false)
+	var result []string
+	seen := make(map[string]bool)
+	for _, w := range words {
+		if stopWords[w] || seen[w] {
+			continue
+		}
+		if !validEntityName(w) {
+			continue
+		}
+		seen[w] = true
+		result = append(result, w)
+	}
+	return result
+}

internal/memory/cut_test.go

@@ -0,0 +1,107 @@
+package memory
+
+import (
+	"testing"
+)
+
+func needJieba(t *testing.T) {
+	t.Helper()
+	if GetJieba() == nil {
+		t.Skip("jieba dictionaries not found")
+	}
+}
+
+func TestCutExact(t *testing.T) {
+	needJieba(t)
+
+	tests := []struct {
+		name string
+		text string
+		min  int
+		not  []string
+	}{
+		{
+			name: "chinese_sentence",
+			text: "今天天气怎么样",
+			min:  2,
+			not:  nil,
+		},
+		{
+			name: "stop_words_removed",
+			text: "和天气地",
+			min:  1,
+			not:  []string{"的", "和"},
+		},
+		{
+			name: "short_words_filtered",
+			text: "今天好天气",
+			min:  1,
+			not:  []string{"好"},
+		},
+		{
+			name: "empty_text",
+			text: "",
+			min:  0,
+			not:  nil,
+		},
+		{
+			name: "qq_conversation",
+			text: "今天天气怎么样 → 今天天气很好",
+			min:  2,
+			not:  nil,
+		},
+		{
+			name: "all_stop_words",
+			text: "的了呢吗",
+			min:  0,
+			not:  nil,
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			got := CutExact(tt.text)
+			if len(got) < tt.min {
+				t.Errorf("CutExact(%q) = %v (len=%d), want at least %d terms", tt.text, got, len(got), tt.min)
+			}
+			for _, forbid := range tt.not {
+				for _, g := range got {
+					if g == forbid {
+						t.Errorf("CutExact(%q) = %v, should not contain %q", tt.text, got, forbid)
+					}
+				}
+			}
+		})
+	}
+}
+
+func TestCutExactNoDuplicates(t *testing.T) {
+	needJieba(t)
+
+	got := CutExact("天气天气天气")
+	if len(got) > 1 {
+		t.Errorf("expected deduplicated result, got %v (len=%d)", got, len(got))
+	}
+}
+
+func TestCutExactValidEntityName(t *testing.T) {
+	needJieba(t)
+
+	got := CutExact("a b c")
+	for _, g := range got {
+		if !validEntityName(g) {
+			t.Errorf("CutExact returned invalid entity name %q", g)
+		}
+	}
+}
+
+func TestCutExactRemoveTimestamp(t *testing.T) {
+	needJieba(t)
+
+	got := CutExact("[15:04] 今天天气不错")
+	for _, g := range got {
+		if g == "15" || g == "04" || g == "15:04" {
+			t.Errorf("timestamp should be removed by CleanTemplateText, got %q in %v", g, got)
+		}
+	}
+}

internal/memory/document/document.go

@@ -28,6 +28,7 @@ type Doc struct {
 	Meta        map[string]string `json:"meta,omitempty"`
 	AccessCount int               `json:"access_count"`    // 访问次数
 	LastAccess  time.Time         `json:"last_access"`     // 最后访问时间
+	Vector      vector.Vector     `json:"vector,omitempty"` // 预计算向量（与 context 同空间），nil 则用 TF-IDF 兜底
 }
 
 // Store — 文档记忆存储，包含向量索引
@@ -37,12 +38,31 @@ type Store struct {
 	veczer *vector.TFIDFVectorizer
 	mu     sync.RWMutex
 
-	docs      map[string]*Doc
-	summaries []string // 用于训练向量化器，最大 10000 条
+	docs       map[string]*Doc
+	summaries  []string // 用于训练向量化器，最大 10000 条
+	vectorizer vector.Vectorizer // 可选：与 context 同空间的向量化器
 
 	dirty bool
 }
 
+func (s *Store) SetVectorizer(v vector.Vectorizer) {
+	s.vectorizer = v
+}
+
+// ReindexWithVectorizer 用给定的向量化器重建所有文档的向量索引
+func (s *Store) ReindexWithVectorizer(v vector.Vectorizer) {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+
+	log.Printf("[document memory] reindex with vectorizer (%d docs)", len(s.docs))
+	s.vec = vector.NewStore()
+	for _, doc := range s.docs {
+		doc.Vector = v.Vectorize(doc.Summary + " " + doc.Content)
+		s.vec.Insert(doc.ID, doc.Summary, doc.Vector, doc.Meta)
+	}
+	log.Printf("[document memory] reindex with vectorizer complete (%d vectors)", s.vec.Size())
+}
+
 const maxSummaries = 10000
 
 func NewStore(dir string) *Store {
@@ -88,7 +108,10 @@ func (s *Store) Insert(doc *Doc) error {
 
 	// 增量训练向量化器并加入向量索引
 	s.addSummary(doc.Summary)
-	vec := s.veczer.Vectorize(doc.Summary + " " + doc.Content)
+	vec := doc.Vector
+	if vec == nil {
+		vec = s.veczer.Vectorize(doc.Summary + " " + doc.Content)
+	}
 	s.vec.Insert(doc.ID, doc.Summary, vec, doc.Meta)
 
 	// 立即写盘
@@ -101,7 +124,7 @@ func (s *Store) Insert(doc *Doc) error {
 }
 
 // ContextToDoc — 将一段上下文对话历史提炼为文档（带内容去重）
-func (s *Store) ContextToDoc(source string, entries []ContextEntry) (*Doc, error) {
+func (s *Store) ContextToDoc(source string, entries []ContextEntry, vec vector.Vectorizer) (*Doc, error) {
 	if len(entries) == 0 {
 		return nil, nil
 	}
@@ -140,6 +163,12 @@ func (s *Store) ContextToDoc(source string, entries []ContextEntry) (*Doc, error
 	}
 
 	id := fmt.Sprintf("doc_%d", time.Now().UnixNano())
+	var docVec vector.Vector
+	if vec != nil {
+		docVec = vec.Vectorize(summary + " " + content)
+	} else {
+		docVec = s.veczer.Vectorize(summary + " " + content)
+	}
 	doc := &Doc{
 		ID:          id,
 		Summary:     summary,
@@ -152,13 +181,13 @@ func (s *Store) ContextToDoc(source string, entries []ContextEntry) (*Doc, error
 		AccessCount: 1,
 		Source:      source,
 		Meta:        map[string]string{"content_hash": contentHash},
+		Vector:      docVec,
 	}
 	s.docs[id] = doc
 
-	// 增量训练向量化器并加入向量索引
+	// 加入向量索引
 	s.addSummary(summary)
-	vec := s.veczer.Vectorize(summary + " " + content)
-	s.vec.Insert(id, summary, vec, nil)
+	s.vec.Insert(id, summary, doc.Vector, nil)
 
 	s.dirty = true
 	s.mu.Unlock()
@@ -180,7 +209,7 @@ func (s *Store) Consume(text string, topK int) []*Doc {
 		topK = 5
 	}
 
-	vec := s.veczer.Vectorize(text)
+	vec := s.vectorizeQuery(text)
 	results := s.vec.Search(vec, topK)
 
 	var docs []*Doc
@@ -194,6 +223,14 @@ func (s *Store) Consume(text string, topK int) []*Doc {
 	return docs
 }
 
+// vectorizeQuery 用语义向量化器（首选）或 TF-IDF（兜底）处理查询文本
+func (s *Store) vectorizeQuery(text string) vector.Vector {
+	if s.vectorizer != nil {
+		return s.vectorizer.Vectorize(text)
+	}
+	return s.veczer.Vectorize(text)
+}
+
 // Query — 向量相似度查询文档
 func (s *Store) Query(text string, topK int) []*Doc {
 	s.mu.RLock()
@@ -203,7 +240,7 @@ func (s *Store) Query(text string, topK int) []*Doc {
 		topK = 5
 	}
 
-	vec := s.veczer.Vectorize(text)
+	vec := s.vectorizeQuery(text)
 	results := s.vec.Search(vec, topK)
 
 	var docs []*Doc
@@ -228,7 +265,10 @@ func (s *Store) Reindex() {
 
 	s.vec = vector.NewStore()
 	for _, doc := range s.docs {
-		vec := s.veczer.Vectorize(doc.Summary + " " + doc.Content)
+		vec := doc.Vector
+		if vec == nil {
+			vec = s.veczer.Vectorize(doc.Summary + " " + doc.Content)
+		}
 		s.vec.Insert(doc.ID, doc.Summary, vec, doc.Meta)
 	}
 
@@ -348,7 +388,10 @@ func (s *Store) loadAll() error {
 
 	// 重建向量索引
 	for _, doc := range s.docs {
-		vec := s.veczer.Vectorize(doc.Summary + " " + doc.Content)
+		vec := doc.Vector
+		if vec == nil {
+			vec = s.veczer.Vectorize(doc.Summary + " " + doc.Content)
+		}
 		s.vec.Insert(doc.ID, doc.Summary, vec, nil)
 	}
 

internal/memory/document/document_test.go

@@ -82,7 +82,7 @@ func TestContextToDoc(t *testing.T) {
 		{Timestamp: time.Now(), Source: "user", Content: "特别是Go语言", Response: "Go很棒"},
 	}
 
-	doc, err := s.ContextToDoc("test", entries)
+	doc, err := s.ContextToDoc("test", entries, nil)
 	if err != nil {
 		t.Fatal(err)
 	}