From 5a2be87f658d01bc5b3b7ecdba92ac69bf09dcf7 Mon Sep 17 00:00:00 2001
From: ai_xiaopei <xiaopei@aisim.cn>
Date: Thu, 03 Sep 2026 15:15:37 +0800
Subject: [PATCH] fix(search): 搜索准确性四连修——FTS5 MATCH 引号包裹(连字符不再被解析成 MINUS)；KeywordSearch 截断按命中关键词数排序(高频 bigram 不再挤出强相关文档)；原词奖励档(完整短语压过泛化 bigram)；tags 字段真实参与评分(此前误用 Section)
---
internal/index/fts.go | 181 +++++++++++++++++++++++++++++++++++++++++----
1 files changed, 164 insertions(+), 17 deletions(-)
diff --git a/internal/index/fts.go b/internal/index/fts.go
index c06b250..83e7838 100644
--- a/internal/index/fts.go
+++ b/internal/index/fts.go
@@ -4,6 +4,7 @@
"database/sql"
"encoding/json"
"sort"
+ "strings"
)
// FTSResult 全文搜索结果
type FTSResult struct {
@@ -12,29 +13,51 @@
Title string
Section string
Rank float64
+ Aliases []string // 别名（JSON 解析）
+ Tags []string // 标签（JSON 解析）
+ Status string // 状态（草稿/待确认/跟进中 等参与降权）
}
-// CreateFTS 创建 FTS5 虚拟表
+// CreateFTS 创建 FTS5 虚拟表（external-content 模式，由触发器增量维护）
+// 注意：external-content 模式下 FTS 列名必须与 content 表列名一致，
+// 故 content 列对应 nodes 表的 content_fts 列
func (s *Store) CreateFTS() error {
_, err := s.db.Exec(`
CREATE VIRTUAL TABLE IF NOT EXISTS nodes_fts USING fts5(
- title,
- content,
- tags,
- entities,
- content='nodes',
- content_rowid='id'
- )
- `)
- return err
+ title, content_fts, tags, aliases,
+ content='nodes', content_rowid='id'
+ )`)
+ if err != nil {
+ return err
+ }
+ // 触发器：nodes 的增删改同步维护 FTS
+ triggers := []string{
+ `CREATE TRIGGER IF NOT EXISTS nodes_fts_ai AFTER INSERT ON nodes BEGIN
+ INSERT INTO nodes_fts(rowid, title, content_fts, tags, aliases)
+ VALUES (new.id, new.title, new.content_fts, new.tags, new.aliases);
+ END`,
+ `CREATE TRIGGER IF NOT EXISTS nodes_fts_ad AFTER DELETE ON nodes BEGIN
+ INSERT INTO nodes_fts(nodes_fts, rowid, title, content_fts, tags, aliases)
+ VALUES ('delete', old.id, old.title, old.content_fts, old.tags, old.aliases);
+ END`,
+ `CREATE TRIGGER IF NOT EXISTS nodes_fts_au AFTER UPDATE ON nodes BEGIN
+ INSERT INTO nodes_fts(nodes_fts, rowid, title, content_fts, tags, aliases)
+ VALUES ('delete', old.id, old.title, old.content_fts, old.tags, old.aliases);
+ INSERT INTO nodes_fts(rowid, title, content_fts, tags, aliases)
+ VALUES (new.id, new.title, new.content_fts, new.tags, new.aliases);
+ END`,
+ }
+ for _, t := range triggers {
+ if _, err := s.db.Exec(t); err != nil {
+ return err
+ }
+ }
+ return nil
}
-// PopulateFTS 填充 FTS 索引
+// PopulateFTS 全量重灌 FTS（external-content 模式专用语法）
func (s *Store) PopulateFTS() error {
- _, err := s.db.Exec(`
- INSERT INTO nodes_fts(rowid, title, content, tags, entities)
- SELECT id, title, content_fts, tags, entities FROM nodes
- `)
+ _, err := s.db.Exec(`INSERT INTO nodes_fts(nodes_fts) VALUES('rebuild')`)
return err
}
@@ -51,7 +74,7 @@
for _, kw := range keywords {
rows, err := s.db.Query(`
- SELECT n.id, n.path, n.title, n.section, fts.rank
+ SELECT n.id, n.path, n.title, n.section, fts.rank, n.aliases, n.status
FROM nodes_fts fts
JOIN nodes n ON n.id = fts.rowid
WHERE nodes_fts MATCH ?
@@ -65,10 +88,14 @@
for rows.Next() {
var r FTSResult
- if err := rows.Scan(&r.ID, &r.Path, &r.Title, &r.Section, &r.Rank); err != nil {
+ var aliasesJSON string
+ if err := rows.Scan(&r.ID, &r.Path, &r.Title, &r.Section, &r.Rank, &aliasesJSON, &r.Status); err != nil {
rows.Close()
continue
}
+ if aliasesJSON != "" {
+ _ = json.Unmarshal([]byte(aliasesJSON), &r.Aliases)
+ }
// 去重
if !seen[r.ID] {
seen[r.ID] = true
@@ -91,6 +118,126 @@
return allResults, nil
}
+// containsCJK 是否含 CJK 统一表意文字
+func containsCJK(s string) bool {
+ for _, r := range s {
+ if r >= 0x4E00 && r <= 0x9FFF {
+ return true
+ }
+ }
+ return false
+}
+
+// KeywordSearch 双通道关键词检索：ASCII 词走 FTS5 MATCH，CJK 词走 LIKE（title/aliases/content/tags）。
+// FTS5 unicode61 把连续中文当整串单 token，多字符 CJK 词 MATCH 匹配不到，必须走 LIKE。
+// 结果按"首命中顺序"排列（FTS 通道按 rank，LIKE 通道补在尾部），engine 再按双信号重排。
+func (s *Store) KeywordSearch(keywords []string, limit int) ([]FTSResult, error) {
+ if len(keywords) == 0 {
+ return nil, nil
+ }
+ seen := make(map[int64]*FTSResult)
+ hitCount := make(map[int64]int) // 每个节点被多少个关键词命中（截断排序用）
+ firstIdx := make(map[int64]int) // 首命中位置（截断平手时的次序依据）
+ var order []int64
+ addCandidate := func(r *FTSResult) {
+ if _, ok := seen[r.ID]; !ok {
+ seen[r.ID] = r
+ firstIdx[r.ID] = len(order)
+ order = append(order, r.ID)
+ }
+ hitCount[r.ID]++
+ }
+
+ // 通道 1: FTS（ASCII 词）
+ var ascii []string
+ for _, kw := range keywords {
+ if !containsCJK(kw) {
+ ascii = append(ascii, kw)
+ }
+ }
+ for _, kw := range ascii {
+ // FTS5 查询语法里 "-" 是 MINUS 运算符、"(" 等是语法符，
+ // 不引号包裹会报 "no such column" 被静默跳过（如 tag "fill-curve" 搜不到）。
+ // 用双引号包裹成字符串字面量，内部双引号转义。
+ quoted := `"` + strings.ReplaceAll(kw, `"`, `""`) + `"`
+ rows, err := s.db.Query(`
+ SELECT n.id, n.path, n.title, n.section, fts.rank, n.aliases, n.tags, n.status
+ FROM nodes_fts fts JOIN nodes n ON n.id = fts.rowid
+ WHERE nodes_fts MATCH ? ORDER BY fts.rank LIMIT ?`, quoted, limit)
+ if err != nil {
+ continue // 语法异常词跳过
+ }
+ for rows.Next() {
+ var r FTSResult
+ var aliasesJSON, tagsJSON string
+ if err := rows.Scan(&r.ID, &r.Path, &r.Title, &r.Section, &r.Rank, &aliasesJSON, &tagsJSON, &r.Status); err != nil {
+ continue
+ }
+ if aliasesJSON != "" {
+ _ = json.Unmarshal([]byte(aliasesJSON), &r.Aliases)
+ }
+ if tagsJSON != "" {
+ _ = json.Unmarshal([]byte(tagsJSON), &r.Tags)
+ }
+ addCandidate(&r)
+ }
+ rows.Close()
+ }
+
+ // 通道 2: LIKE（CJK 词）
+ for _, kw := range keywords {
+ if !containsCJK(kw) {
+ continue
+ }
+ pat := "%" + kw + "%"
+ rows, err := s.db.Query(`
+ SELECT id, path, title, section, aliases, tags, status FROM nodes
+ WHERE title LIKE ? OR aliases LIKE ? OR content_fts LIKE ? OR tags LIKE ?`,
+ pat, pat, pat, pat)
+ if err != nil {
+ continue
+ }
+ for rows.Next() {
+ var r FTSResult
+ var aliasesJSON, tagsJSON string
+ if err := rows.Scan(&r.ID, &r.Path, &r.Title, &r.Section, &aliasesJSON, &tagsJSON, &r.Status); err != nil {
+ continue
+ }
+ if aliasesJSON != "" {
+ _ = json.Unmarshal([]byte(aliasesJSON), &r.Aliases)
+ }
+ if tagsJSON != "" {
+ _ = json.Unmarshal([]byte(tagsJSON), &r.Tags)
+ }
+ r.Rank = 0 // LIKE 无 rank
+ addCandidate(&r)
+ }
+ rows.Close()
+ }
+
+ // 截断排序：命中关键词多的节点优先保留，平手按首命中先后。
+ // 旧实现按首命中顺序直接截断，高频泛词（如 bigram "电子" 命中 486 行）
+ // 会把排在后面但与多词强相关的文档挤出候选集。
+ ranked := make([]int64, len(order))
+ copy(ranked, order)
+ sort.SliceStable(ranked, func(i, j int) bool {
+ a, b := ranked[i], ranked[j]
+ if hitCount[a] != hitCount[b] {
+ return hitCount[a] > hitCount[b]
+ }
+ return firstIdx[a] < firstIdx[b]
+ })
+
+ results := make([]FTSResult, 0, len(order))
+ for _, id := range ranked {
+ results = append(results, *seen[id])
+ }
+ if limit > 0 && len(results) > limit {
+ results = results[:limit]
+ }
+ return results, nil
+}
+
// GetNodeContent 获取节点内容
func (s *Store) GetNodeContent(id int64) (string, []string, []string, error) {
var content, tagsJSON, entitiesJSON string
--
Gitblit v1.10.0