| New file |
| | |
| | | package search |
| | | |
| | | import ( |
| | | "regexp" |
| | | "strings" |
| | | |
| | | "github.com/aisim/kb-cli/internal/index" |
| | | ) |
| | | |
| | | // headingRe 1-4 级 Markdown 标题 |
| | | var headingRe = regexp.MustCompile(`^#{1,4} .+$`) |
| | | |
| | | // splitParagraphs 按 1-4 级标题切段。无标题的文档整体为一段。 |
| | | func splitParagraphs(content string) []string { |
| | | lines := strings.Split(content, "\n") |
| | | var paras []string |
| | | var cur []string |
| | | flush := func() { |
| | | if len(cur) > 0 { |
| | | paras = append(paras, strings.TrimRight(strings.Join(cur, "\n"), "\n")+"\n") |
| | | cur = nil |
| | | } |
| | | } |
| | | for _, l := range lines { |
| | | if headingRe.MatchString(l) { |
| | | flush() |
| | | } |
| | | cur = append(cur, l) |
| | | } |
| | | flush() |
| | | return paras |
| | | } |
| | | |
| | | // extractRelevantParagraphs 返回命中关键词的段落(整段不截半句)。 |
| | | // 文档总长 <= budget 时整篇输出;无命中段落时输出空串; |
| | | // 预算约束:累计超预算的段落丢弃(不截半段)。 |
| | | func extractRelevantParagraphs(content string, keywords []string, budget int) string { |
| | | if budget <= 0 || len(content) <= budget { |
| | | return content |
| | | } |
| | | var out []string |
| | | for _, p := range splitParagraphs(content) { |
| | | for _, kw := range keywords { |
| | | if strings.Contains(p, kw) { |
| | | out = append(out, p) |
| | | break |
| | | } |
| | | } |
| | | } |
| | | // 预算约束:累计超预算的段落丢弃(不截半段) |
| | | var total int |
| | | kept := []string{} |
| | | for _, p := range out { |
| | | if total+len(p) > budget { |
| | | break |
| | | } |
| | | total += len(p) |
| | | kept = append(kept, p) |
| | | } |
| | | return strings.Join(kept, "") |
| | | } |
| | | |
| | | // extractWithBudget extractRelevantParagraphs 的预算版(budget<=0 视为无预算) |
| | | func extractWithBudget(content string, keywords []string, budget int) string { |
| | | if budget <= 0 { |
| | | budget = 100000 |
| | | } |
| | | return extractRelevantParagraphs(content, keywords, budget) |
| | | } |
| | | |
| | | // ExploreOptions explore 参数 |
| | | type ExploreOptions struct { |
| | | Budget int // 字节预算(0 = 用配置默认) |
| | | TopN int // 0 = 用配置默认 |
| | | HardBudget int // 字节硬上限(0 = 用代码缺省 32000) |
| | | } |
| | | |
| | | // ExploredDoc 入选文档及其输出正文 |
| | | type ExploredDoc struct { |
| | | Path string `json:"path"` |
| | | Title string `json:"title"` |
| | | Section string `json:"section"` |
| | | Score int `json:"score"` |
| | | Body string `json:"body"` |
| | | } |
| | | |
| | | // ExploreResult explore 结果 |
| | | type ExploreResult struct { |
| | | Docs []ExploredDoc `json:"docs"` |
| | | Related map[string][]string `json:"related"` |
| | | UnresolvedLinks []string `json:"unresolved_links"` |
| | | } |
| | | |
| | | // Explore 一次调用返回相关文档原文 + 关联清单 + 悬空链接 |
| | | func Explore(store *index.Store, keywords []string, cfg ExploreOptions) (*ExploreResult, error) { |
| | | if len(keywords) == 0 { |
| | | return nil, nil |
| | | } |
| | | // 段落命中用原词(精确),FTS/LIKE 检索用展开后的 bigram |
| | | searchKeywords := ExpandCJKKeywords(keywords) |
| | | budget := cfg.Budget |
| | | if budget <= 0 { |
| | | budget = 16000 // 代码缺省(config 读取在 cmd 层完成) |
| | | } |
| | | if budget > 32000 { |
| | | budget = 32000 // 硬上限缺省 |
| | | } |
| | | // 配置显式给出的硬上限优先(cmd 层读 config.yaml 的 explore.hard_budget) |
| | | if cfg.HardBudget > 0 && budget > cfg.HardBudget { |
| | | budget = cfg.HardBudget |
| | | } |
| | | topN := cfg.TopN |
| | | if topN <= 0 { |
| | | topN = 5 |
| | | } |
| | | |
| | | opts := SearchOptions{TopN: topN} |
| | | results, err := Search(store, searchKeywords, opts) |
| | | if err != nil { |
| | | return nil, err |
| | | } |
| | | if len(results) == 0 { |
| | | return &ExploreResult{Related: map[string][]string{}}, nil |
| | | } |
| | | |
| | | // 按分数降序分配预算:每文档至少 800 字节 |
| | | res := &ExploreResult{Related: map[string][]string{}} |
| | | perDoc := budget / len(results) |
| | | if perDoc < 800 { |
| | | perDoc = 800 |
| | | } |
| | | for _, r := range results { |
| | | content, _, _, err := store.GetNodeContent(r.ID) |
| | | if err != nil || content == "" { |
| | | continue |
| | | } |
| | | // 两级回退:先用原词提取(精确);原词在任何段落都不连续出现时 |
| | | //(复合 CJK 词常见:bigram 召回了文档,但原词不连续)回退到展开词提取, |
| | | // 避免大文档被 continue 静默丢弃、最终输出「未找到相关文档」。 |
| | | body := extractRelevantParagraphs(content, keywords, perDoc) |
| | | if body == "" { |
| | | body = extractRelevantParagraphs(content, searchKeywords, perDoc) |
| | | } |
| | | if body == "" { |
| | | // 无命中段落但文档入选 → 整篇(若放得下),否则跳过 |
| | | if len(content) <= perDoc { |
| | | body = content |
| | | } else { |
| | | continue |
| | | } |
| | | } |
| | | res.Docs = append(res.Docs, ExploredDoc{ |
| | | Path: r.Path, Title: r.Title, Section: r.Section, |
| | | Score: r.Score, Body: body, |
| | | }) |
| | | // 关联清单 |
| | | if links, err := store.GetNodeLinks(r.ID); err == nil { |
| | | res.Related[r.Path] = links |
| | | } |
| | | } |
| | | |
| | | // 悬空链接提示:入选文档的 wikilinks 中未解析的 |
| | | for _, d := range res.Docs { |
| | | links, err := store.GetUnresolvedLinks(d.Path) |
| | | if err == nil { |
| | | res.UnresolvedLinks = append(res.UnresolvedLinks, links...) |
| | | } |
| | | } |
| | | return res, nil |
| | | } |