mirror of
https://github.com/Ed1s0nZ/CyberStrikeAI.git
synced 2026-10-11 02:08:51 +02:00
Add files via upload
This commit is contained in:
23 files changed
+1695
-1413
No files matched your search
+174
-545
@@ -8,23 +8,34 @@ import (
|
||||
"math"
|
||||
"sort"
|
||||
"strings"
|
||||
"sync"
|
||||
|
||||
"cyberstrike-ai/internal/config"
|
||||
|
||||
"github.com/cloudwego/eino/components/retriever"
|
||||
"github.com/cloudwego/eino/schema"
|
||||
"go.uber.org/zap"
|
||||
)
|
||||
|
||||
// Retriever 检索器
|
||||
// Retriever 检索器:SQLite 存向量 + Eino 嵌入,**纯向量检索**(余弦相似度、TopK、阈值),
|
||||
// 实现语义与 [retriever.Retriever] 适配层 [VectorEinoRetriever] 一致。
|
||||
type Retriever struct {
|
||||
db *sql.DB
|
||||
embedder *Embedder
|
||||
config *RetrievalConfig
|
||||
logger *zap.Logger
|
||||
|
||||
rerankMu sync.RWMutex
|
||||
reranker DocumentReranker
|
||||
}
|
||||
|
||||
// RetrievalConfig 检索配置
|
||||
type RetrievalConfig struct {
|
||||
TopK int
|
||||
SimilarityThreshold float64
|
||||
HybridWeight float64
|
||||
// SubIndexFilter 非空时仅检索 sub_indexes 包含该标签(逗号分隔之一)的行;空 sub_indexes 的旧行仍保留以兼容。
|
||||
SubIndexFilter string
|
||||
PostRetrieve config.PostRetrieveConfig
|
||||
}
|
||||
|
||||
// NewRetriever 创建新的检索器
|
||||
@@ -38,18 +49,41 @@ func NewRetriever(db *sql.DB, embedder *Embedder, config *RetrievalConfig, logge
|
||||
}
|
||||
|
||||
// UpdateConfig 更新检索配置
|
||||
func (r *Retriever) UpdateConfig(config *RetrievalConfig) {
|
||||
if config != nil {
|
||||
r.config = config
|
||||
r.logger.Info("检索器配置已更新",
|
||||
zap.Int("top_k", config.TopK),
|
||||
zap.Float64("similarity_threshold", config.SimilarityThreshold),
|
||||
zap.Float64("hybrid_weight", config.HybridWeight),
|
||||
)
|
||||
func (r *Retriever) UpdateConfig(cfg *RetrievalConfig) {
|
||||
if cfg != nil {
|
||||
r.config = cfg
|
||||
if r.logger != nil {
|
||||
r.logger.Info("检索器配置已更新",
|
||||
zap.Int("top_k", cfg.TopK),
|
||||
zap.Float64("similarity_threshold", cfg.SimilarityThreshold),
|
||||
zap.String("sub_index_filter", cfg.SubIndexFilter),
|
||||
zap.Int("post_retrieve_prefetch_top_k", cfg.PostRetrieve.PrefetchTopK),
|
||||
zap.Int("post_retrieve_max_context_chars", cfg.PostRetrieve.MaxContextChars),
|
||||
zap.Int("post_retrieve_max_context_tokens", cfg.PostRetrieve.MaxContextTokens),
|
||||
)
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
// cosineSimilarity 计算余弦相似度
|
||||
// SetDocumentReranker 注入可选重排器(并发安全);nil 表示禁用。
|
||||
func (r *Retriever) SetDocumentReranker(rr DocumentReranker) {
|
||||
if r == nil {
|
||||
return
|
||||
}
|
||||
r.rerankMu.Lock()
|
||||
defer r.rerankMu.Unlock()
|
||||
r.reranker = rr
|
||||
}
|
||||
|
||||
func (r *Retriever) documentReranker() DocumentReranker {
|
||||
if r == nil {
|
||||
return nil
|
||||
}
|
||||
r.rerankMu.RLock()
|
||||
defer r.rerankMu.RUnlock()
|
||||
return r.reranker
|
||||
}
|
||||
|
||||
func cosineSimilarity(a, b []float32) float64 {
|
||||
if len(a) != len(b) {
|
||||
return 0.0
|
||||
@@ -69,608 +103,203 @@ func cosineSimilarity(a, b []float32) float64 {
|
||||
return dotProduct / (math.Sqrt(normA) * math.Sqrt(normB))
|
||||
}
|
||||
|
||||
// bm25Score 计算 BM25 分数(带缓存的改进版本)
|
||||
// 注意:由于缺少全局文档统计,使用简化 IDF 计算
|
||||
func (r *Retriever) bm25Score(query, text string) float64 {
|
||||
queryTerms := strings.Fields(strings.ToLower(query))
|
||||
if len(queryTerms) == 0 {
|
||||
return 0.0
|
||||
// Search 搜索知识库。统一经 [VectorEinoRetriever](Eino retriever.Retriever 边界)。
|
||||
func (r *Retriever) Search(ctx context.Context, req *SearchRequest) ([]*RetrievalResult, error) {
|
||||
if req == nil {
|
||||
return nil, fmt.Errorf("请求不能为空")
|
||||
}
|
||||
|
||||
textLower := strings.ToLower(text)
|
||||
textTerms := strings.Fields(textLower)
|
||||
if len(textTerms) == 0 {
|
||||
return 0.0
|
||||
q := strings.TrimSpace(req.Query)
|
||||
if q == "" {
|
||||
return nil, fmt.Errorf("查询不能为空")
|
||||
}
|
||||
|
||||
// BM25 参数(标准值)
|
||||
k1 := 1.2 // 词频饱和度参数(标准范围 1.2-2.0)
|
||||
b := 0.75 // 长度归一化参数(标准值)
|
||||
avgDocLength := 150.0 // 估算的平均文档长度(基于典型知识块大小)
|
||||
docLength := float64(len(textTerms))
|
||||
|
||||
// 计算词频映射
|
||||
textTermFreq := make(map[string]int, len(textTerms))
|
||||
for _, term := range textTerms {
|
||||
textTermFreq[term]++
|
||||
opts := r.einoRetrieverOptions(req)
|
||||
docs, err := NewVectorEinoRetriever(r).Retrieve(ctx, q, opts...)
|
||||
if err != nil {
|
||||
return nil, err
|
||||
}
|
||||
|
||||
score := 0.0
|
||||
matchedQueryTerms := 0
|
||||
|
||||
for _, term := range queryTerms {
|
||||
termFreq, exists := textTermFreq[term]
|
||||
if !exists || termFreq == 0 {
|
||||
continue
|
||||
}
|
||||
matchedQueryTerms++
|
||||
|
||||
// BM25 TF 计算公式
|
||||
tf := float64(termFreq)
|
||||
lengthNorm := 1 - b + b*(docLength/avgDocLength)
|
||||
tfScore := tf / (tf + k1*lengthNorm)
|
||||
|
||||
// 改进的 IDF 计算:使用词长度和出现频率估算
|
||||
// 短词(2-3 字符)通常更重要,长词 IDF 略低
|
||||
idfWeight := 1.0
|
||||
termLen := len(term)
|
||||
if termLen <= 2 {
|
||||
// 极短词(如 go, js)给予更高权重
|
||||
idfWeight = 1.2 + math.Log(1.0+float64(termFreq)/20.0)
|
||||
} else if termLen <= 4 {
|
||||
// 短词(4 字符)标准权重
|
||||
idfWeight = 1.0 + math.Log(1.0+float64(termFreq)/15.0)
|
||||
} else {
|
||||
// 长词稍微降低权重
|
||||
idfWeight = 0.9 + math.Log(1.0+float64(termFreq)/10.0)
|
||||
}
|
||||
|
||||
score += tfScore * idfWeight
|
||||
}
|
||||
|
||||
// 归一化:考虑匹配的查询词比例
|
||||
if len(queryTerms) > 0 {
|
||||
// 使用匹配比例作为额外因子
|
||||
matchRatio := float64(matchedQueryTerms) / float64(len(queryTerms))
|
||||
score = (score / float64(len(queryTerms))) * (1 + matchRatio) / 2
|
||||
}
|
||||
|
||||
return math.Min(score, 1.0)
|
||||
return documentsToRetrievalResults(docs)
|
||||
}
|
||||
|
||||
// Search 搜索知识库
|
||||
func (r *Retriever) Search(ctx context.Context, req *SearchRequest) ([]*RetrievalResult, error) {
|
||||
func (r *Retriever) einoRetrieverOptions(req *SearchRequest) []retriever.Option {
|
||||
var opts []retriever.Option
|
||||
if req.TopK > 0 {
|
||||
opts = append(opts, retriever.WithTopK(req.TopK))
|
||||
}
|
||||
dsl := map[string]any{}
|
||||
if strings.TrimSpace(req.RiskType) != "" {
|
||||
dsl[DSLRiskType] = strings.TrimSpace(req.RiskType)
|
||||
}
|
||||
if req.Threshold > 0 {
|
||||
dsl[DSLSimilarityThreshold] = req.Threshold
|
||||
}
|
||||
if strings.TrimSpace(req.SubIndexFilter) != "" {
|
||||
dsl[DSLSubIndexFilter] = strings.TrimSpace(req.SubIndexFilter)
|
||||
}
|
||||
if len(dsl) > 0 {
|
||||
opts = append(opts, retriever.WithDSLInfo(dsl))
|
||||
}
|
||||
return opts
|
||||
}
|
||||
|
||||
// EinoRetrieve 直接返回 [schema.Document],供 Eino Graph / Chain 使用。
|
||||
func (r *Retriever) EinoRetrieve(ctx context.Context, query string, opts ...retriever.Option) ([]*schema.Document, error) {
|
||||
return NewVectorEinoRetriever(r).Retrieve(ctx, query, opts...)
|
||||
}
|
||||
|
||||
func (r *Retriever) knowledgeEmbeddingSelectSQL(riskType, subIndexFilter string) (string, []interface{}) {
|
||||
q := `SELECT e.id, e.item_id, e.chunk_index, e.chunk_text, e.embedding, e.embedding_model, e.embedding_dim, i.category, i.title
|
||||
FROM knowledge_embeddings e
|
||||
JOIN knowledge_base_items i ON e.item_id = i.id
|
||||
WHERE 1=1`
|
||||
var args []interface{}
|
||||
if strings.TrimSpace(riskType) != "" {
|
||||
q += ` AND TRIM(i.category) = TRIM(?) COLLATE NOCASE`
|
||||
args = append(args, riskType)
|
||||
}
|
||||
if tag := strings.TrimSpace(subIndexFilter); tag != "" {
|
||||
tag = strings.ToLower(strings.ReplaceAll(tag, " ", ""))
|
||||
q += ` AND (TRIM(COALESCE(e.sub_indexes,'')) = '' OR INSTR(',' || LOWER(REPLACE(e.sub_indexes,' ','')) || ',', ',' || ? || ',') > 0)`
|
||||
args = append(args, tag)
|
||||
}
|
||||
return q, args
|
||||
}
|
||||
|
||||
// vectorSearch 纯向量检索:余弦相似度排序,按相似度阈值与 TopK 截断(无 BM25、无混合分、无邻块扩展)。
|
||||
func (r *Retriever) vectorSearch(ctx context.Context, req *SearchRequest) ([]*RetrievalResult, error) {
|
||||
if req.Query == "" {
|
||||
return nil, fmt.Errorf("查询不能为空")
|
||||
}
|
||||
|
||||
topK := req.TopK
|
||||
if topK <= 0 {
|
||||
if topK <= 0 && r.config != nil {
|
||||
topK = r.config.TopK
|
||||
}
|
||||
if topK == 0 {
|
||||
if topK <= 0 {
|
||||
topK = 5
|
||||
}
|
||||
|
||||
threshold := req.Threshold
|
||||
if threshold <= 0 {
|
||||
if threshold <= 0 && r.config != nil {
|
||||
threshold = r.config.SimilarityThreshold
|
||||
}
|
||||
if threshold == 0 {
|
||||
if threshold <= 0 {
|
||||
threshold = 0.7
|
||||
}
|
||||
|
||||
// 向量化查询(如果提供了risk_type,也包含在查询文本中,以便更好地匹配)
|
||||
queryText := req.Query
|
||||
if req.RiskType != "" {
|
||||
// 将risk_type信息包含到查询中,格式与索引时保持一致
|
||||
queryText = fmt.Sprintf("[风险类型: %s] %s", req.RiskType, req.Query)
|
||||
subIdxFilter := strings.TrimSpace(req.SubIndexFilter)
|
||||
if subIdxFilter == "" && r.config != nil {
|
||||
subIdxFilter = strings.TrimSpace(r.config.SubIndexFilter)
|
||||
}
|
||||
|
||||
queryText := FormatQueryEmbeddingText(req.RiskType, req.Query)
|
||||
queryEmbedding, err := r.embedder.EmbedText(ctx, queryText)
|
||||
if err != nil {
|
||||
return nil, fmt.Errorf("向量化查询失败: %w", err)
|
||||
}
|
||||
|
||||
// 查询所有向量(或按风险类型过滤)
|
||||
// 使用精确匹配(=)以提高性能和准确性
|
||||
// 由于系统提供了内置工具来获取风险类型列表,用户应该使用准确的category名称
|
||||
// 同时,向量嵌入中已包含category信息,即使SQL过滤不完全匹配,向量相似度也能帮助匹配
|
||||
var rows *sql.Rows
|
||||
if req.RiskType != "" {
|
||||
// 使用精确匹配(=),性能更好且更准确
|
||||
// 使用 COLLATE NOCASE 实现大小写不敏感匹配,提高容错性
|
||||
// 注意:如果用户输入的risk_type与category不完全一致,可能匹配不到
|
||||
// 建议用户先调用相应的内置工具获取准确的category名称
|
||||
rows, err = r.db.Query(`
|
||||
SELECT e.id, e.item_id, e.chunk_index, e.chunk_text, e.embedding, i.category, i.title
|
||||
FROM knowledge_embeddings e
|
||||
JOIN knowledge_base_items i ON e.item_id = i.id
|
||||
WHERE TRIM(i.category) = TRIM(?) COLLATE NOCASE
|
||||
`, req.RiskType)
|
||||
} else {
|
||||
rows, err = r.db.Query(`
|
||||
SELECT e.id, e.item_id, e.chunk_index, e.chunk_text, e.embedding, i.category, i.title
|
||||
FROM knowledge_embeddings e
|
||||
JOIN knowledge_base_items i ON e.item_id = i.id
|
||||
`)
|
||||
queryDim := len(queryEmbedding)
|
||||
expectedModel := ""
|
||||
if r.embedder != nil {
|
||||
expectedModel = r.embedder.EmbeddingModelName()
|
||||
}
|
||||
|
||||
sqlStr, sqlArgs := r.knowledgeEmbeddingSelectSQL(strings.TrimSpace(req.RiskType), subIdxFilter)
|
||||
rows, err := r.db.QueryContext(ctx, sqlStr, sqlArgs...)
|
||||
if err != nil {
|
||||
return nil, fmt.Errorf("查询向量失败: %w", err)
|
||||
}
|
||||
defer rows.Close()
|
||||
|
||||
// 计算相似度
|
||||
type candidate struct {
|
||||
chunk *KnowledgeChunk
|
||||
item *KnowledgeItem
|
||||
similarity float64
|
||||
bm25Score float64
|
||||
hasStrongKeywordMatch bool
|
||||
hybridScore float64 // 混合分数,用于最终排序
|
||||
chunk *KnowledgeChunk
|
||||
item *KnowledgeItem
|
||||
similarity float64
|
||||
}
|
||||
|
||||
candidates := make([]candidate, 0)
|
||||
|
||||
rowNum := 0
|
||||
for rows.Next() {
|
||||
var chunkID, itemID, chunkText, embeddingJSON, category, title string
|
||||
var chunkIndex int
|
||||
rowNum++
|
||||
if rowNum%48 == 0 {
|
||||
select {
|
||||
case <-ctx.Done():
|
||||
return nil, ctx.Err()
|
||||
default:
|
||||
}
|
||||
}
|
||||
|
||||
if err := rows.Scan(&chunkID, &itemID, &chunkIndex, &chunkText, &embeddingJSON, &category, &title); err != nil {
|
||||
var chunkID, itemID, chunkText, embeddingJSON, category, title, rowModel string
|
||||
var chunkIndex, rowDim int
|
||||
|
||||
if err := rows.Scan(&chunkID, &itemID, &chunkIndex, &chunkText, &embeddingJSON, &rowModel, &rowDim, &category, &title); err != nil {
|
||||
r.logger.Warn("扫描向量失败", zap.Error(err))
|
||||
continue
|
||||
}
|
||||
|
||||
// 解析向量
|
||||
var embedding []float32
|
||||
if err := json.Unmarshal([]byte(embeddingJSON), &embedding); err != nil {
|
||||
r.logger.Warn("解析向量失败", zap.Error(err))
|
||||
continue
|
||||
}
|
||||
|
||||
// 计算余弦相似度
|
||||
similarity := cosineSimilarity(queryEmbedding, embedding)
|
||||
|
||||
// 计算BM25分数(考虑chunk文本、category和title)
|
||||
// category和title是结构化字段,完全匹配时应该被优先考虑
|
||||
chunkBM25 := r.bm25Score(req.Query, chunkText)
|
||||
categoryBM25 := r.bm25Score(req.Query, category)
|
||||
titleBM25 := r.bm25Score(req.Query, title)
|
||||
|
||||
// 检查category或title是否有显著匹配(这对于结构化字段很重要)
|
||||
hasStrongKeywordMatch := categoryBM25 > 0.3 || titleBM25 > 0.3
|
||||
|
||||
// 综合BM25分数(用于后续排序)
|
||||
bm25Score := math.Max(math.Max(chunkBM25, categoryBM25), titleBM25)
|
||||
|
||||
// 收集所有候选(先不严格过滤,以便后续智能处理跨语言情况)
|
||||
// 只过滤掉相似度极低的结果(< 0.1),避免噪音
|
||||
if similarity < 0.1 {
|
||||
if rowDim > 0 && len(embedding) != rowDim {
|
||||
r.logger.Debug("跳过维度不一致的向量行", zap.String("chunkId", chunkID), zap.Int("rowDim", rowDim), zap.Int("got", len(embedding)))
|
||||
continue
|
||||
}
|
||||
if queryDim > 0 && len(embedding) != queryDim {
|
||||
r.logger.Debug("跳过与查询维度不一致的向量", zap.String("chunkId", chunkID), zap.Int("queryDim", queryDim), zap.Int("got", len(embedding)))
|
||||
continue
|
||||
}
|
||||
if expectedModel != "" && strings.TrimSpace(rowModel) != "" && strings.TrimSpace(rowModel) != expectedModel {
|
||||
r.logger.Debug("跳过嵌入模型不一致的行", zap.String("chunkId", chunkID), zap.String("rowModel", rowModel), zap.String("expected", expectedModel))
|
||||
continue
|
||||
}
|
||||
|
||||
chunk := &KnowledgeChunk{
|
||||
ID: chunkID,
|
||||
ItemID: itemID,
|
||||
ChunkIndex: chunkIndex,
|
||||
ChunkText: chunkText,
|
||||
Embedding: embedding,
|
||||
}
|
||||
|
||||
item := &KnowledgeItem{
|
||||
ID: itemID,
|
||||
Category: category,
|
||||
Title: title,
|
||||
}
|
||||
|
||||
similarity := cosineSimilarity(queryEmbedding, embedding)
|
||||
candidates = append(candidates, candidate{
|
||||
chunk: chunk,
|
||||
item: item,
|
||||
similarity: similarity,
|
||||
bm25Score: bm25Score,
|
||||
hasStrongKeywordMatch: hasStrongKeywordMatch,
|
||||
chunk: &KnowledgeChunk{
|
||||
ID: chunkID,
|
||||
ItemID: itemID,
|
||||
ChunkIndex: chunkIndex,
|
||||
ChunkText: chunkText,
|
||||
Embedding: embedding,
|
||||
},
|
||||
item: &KnowledgeItem{
|
||||
ID: itemID,
|
||||
Category: category,
|
||||
Title: title,
|
||||
},
|
||||
similarity: similarity,
|
||||
})
|
||||
}
|
||||
|
||||
// 先按相似度排序(使用更高效的排序)
|
||||
sort.Slice(candidates, func(i, j int) bool {
|
||||
return candidates[i].similarity > candidates[j].similarity
|
||||
})
|
||||
|
||||
// 智能过滤策略:优先保留关键词匹配的结果,对跨语言查询使用更宽松的阈值
|
||||
filteredCandidates := make([]candidate, 0)
|
||||
|
||||
// 检查是否有任何关键词匹配(用于判断是否是跨语言查询)
|
||||
hasAnyKeywordMatch := false
|
||||
for _, cand := range candidates {
|
||||
if cand.hasStrongKeywordMatch {
|
||||
hasAnyKeywordMatch = true
|
||||
break
|
||||
filtered := make([]candidate, 0, len(candidates))
|
||||
for _, c := range candidates {
|
||||
if c.similarity >= threshold {
|
||||
filtered = append(filtered, c)
|
||||
}
|
||||
}
|
||||
|
||||
// 检查最高相似度,用于判断是否确实有相关内容
|
||||
maxSimilarity := 0.0
|
||||
if len(candidates) > 0 {
|
||||
maxSimilarity = candidates[0].similarity
|
||||
if len(filtered) > topK {
|
||||
filtered = filtered[:topK]
|
||||
}
|
||||
|
||||
// 应用智能过滤
|
||||
// 如果用户设置了高阈值(>=0.8),更严格地遵守阈值,减少自动放宽
|
||||
strictMode := threshold >= 0.8
|
||||
|
||||
// 根据是否有关键词匹配,采用不同的阈值策略
|
||||
// 严格模式下,禁用跨语言放宽策略,严格遵守用户设置的阈值
|
||||
effectiveThreshold := threshold
|
||||
if !strictMode && !hasAnyKeywordMatch {
|
||||
// 非严格模式下,没有关键词匹配,可能是跨语言查询,适度放宽阈值
|
||||
// 但即使跨语言,也不能无脑降低阈值,需要保证最低相关性
|
||||
// 跨语言阈值设为0.6,确保返回的结果至少有一定相关性
|
||||
effectiveThreshold = math.Max(threshold*0.85, 0.6)
|
||||
r.logger.Debug("检测到可能的跨语言查询,使用放宽的阈值",
|
||||
zap.Float64("originalThreshold", threshold),
|
||||
zap.Float64("effectiveThreshold", effectiveThreshold),
|
||||
)
|
||||
} else if strictMode {
|
||||
// 严格模式下,即使没有关键词匹配,也严格遵守阈值
|
||||
r.logger.Debug("严格模式:严格遵守用户设置的阈值",
|
||||
zap.Float64("threshold", threshold),
|
||||
zap.Bool("hasKeywordMatch", hasAnyKeywordMatch),
|
||||
)
|
||||
}
|
||||
for _, cand := range candidates {
|
||||
if cand.similarity >= effectiveThreshold {
|
||||
// 达到阈值,直接通过
|
||||
filteredCandidates = append(filteredCandidates, cand)
|
||||
} else if !strictMode && cand.hasStrongKeywordMatch {
|
||||
// 非严格模式下,有关键词匹配但相似度略低于阈值,适当放宽
|
||||
// 严格模式下,即使有关键词匹配,也严格遵守阈值
|
||||
relaxedThreshold := math.Max(effectiveThreshold*0.85, 0.55)
|
||||
if cand.similarity >= relaxedThreshold {
|
||||
filteredCandidates = append(filteredCandidates, cand)
|
||||
}
|
||||
}
|
||||
// 如果既没有关键词匹配,相似度又低于阈值,则过滤掉
|
||||
}
|
||||
|
||||
// 智能兜底策略:只有在最高相似度达到合理水平时,才考虑返回结果
|
||||
// 如果最高相似度都很低(<0.55),说明确实没有相关内容,应该返回空
|
||||
// 严格模式下(阈值>=0.8),禁用兜底策略,严格遵守用户设置的阈值
|
||||
if len(filteredCandidates) == 0 && len(candidates) > 0 && !strictMode {
|
||||
// 即使没有通过阈值过滤,如果最高相似度还可以(>=0.55),可以考虑返回Top-K
|
||||
// 但这是最后的兜底,只在确实有一定相关性时才使用
|
||||
// 严格模式下不使用兜底策略
|
||||
minAcceptableSimilarity := 0.55
|
||||
if maxSimilarity >= minAcceptableSimilarity {
|
||||
r.logger.Debug("过滤后无结果,但最高相似度可接受,返回Top-K结果",
|
||||
zap.Int("totalCandidates", len(candidates)),
|
||||
zap.Float64("maxSimilarity", maxSimilarity),
|
||||
zap.Float64("effectiveThreshold", effectiveThreshold),
|
||||
)
|
||||
maxResults := topK
|
||||
if len(candidates) < maxResults {
|
||||
maxResults = len(candidates)
|
||||
}
|
||||
// 只返回相似度 >= 0.55 的结果
|
||||
for _, cand := range candidates {
|
||||
if cand.similarity >= minAcceptableSimilarity && len(filteredCandidates) < maxResults {
|
||||
filteredCandidates = append(filteredCandidates, cand)
|
||||
}
|
||||
}
|
||||
} else {
|
||||
r.logger.Debug("过滤后无结果,且最高相似度过低,返回空结果",
|
||||
zap.Int("totalCandidates", len(candidates)),
|
||||
zap.Float64("maxSimilarity", maxSimilarity),
|
||||
zap.Float64("minAcceptableSimilarity", minAcceptableSimilarity),
|
||||
)
|
||||
}
|
||||
} else if len(filteredCandidates) == 0 && strictMode {
|
||||
// 严格模式下,如果过滤后无结果,直接返回空,不使用兜底策略
|
||||
r.logger.Debug("严格模式:过滤后无结果,严格遵守阈值,返回空结果",
|
||||
zap.Float64("threshold", threshold),
|
||||
zap.Float64("maxSimilarity", maxSimilarity),
|
||||
)
|
||||
}
|
||||
|
||||
// 统一在最终返回前严格限制 Top-K 数量
|
||||
if len(filteredCandidates) > topK {
|
||||
// 如果过滤后结果太多,只取Top-K
|
||||
filteredCandidates = filteredCandidates[:topK]
|
||||
}
|
||||
|
||||
candidates = filteredCandidates
|
||||
|
||||
// 混合排序(向量相似度 + BM25)
|
||||
// 注意:hybridWeight可以是0.0(纯关键词检索),所以不设置默认值
|
||||
// 如果配置文件中未设置,应该在配置加载时使用默认值
|
||||
hybridWeight := r.config.HybridWeight
|
||||
// 如果未设置,使用默认值0.7(偏重向量检索)
|
||||
if hybridWeight < 0 || hybridWeight > 1 {
|
||||
r.logger.Warn("混合权重超出范围,使用默认值0.7",
|
||||
zap.Float64("provided", hybridWeight))
|
||||
hybridWeight = 0.7
|
||||
}
|
||||
|
||||
// 先计算混合分数并存储在candidate中,用于排序
|
||||
for i := range candidates {
|
||||
normalizedBM25 := math.Min(candidates[i].bm25Score, 1.0)
|
||||
candidates[i].hybridScore = hybridWeight*candidates[i].similarity + (1-hybridWeight)*normalizedBM25
|
||||
|
||||
// 调试日志:记录前几个候选的分数计算(仅在debug级别)
|
||||
if i < 3 {
|
||||
r.logger.Debug("混合分数计算",
|
||||
zap.Int("index", i),
|
||||
zap.Float64("similarity", candidates[i].similarity),
|
||||
zap.Float64("bm25Score", candidates[i].bm25Score),
|
||||
zap.Float64("normalizedBM25", normalizedBM25),
|
||||
zap.Float64("hybridWeight", hybridWeight),
|
||||
zap.Float64("hybridScore", candidates[i].hybridScore))
|
||||
}
|
||||
}
|
||||
|
||||
// 根据混合分数重新排序(这才是真正的混合检索)
|
||||
sort.Slice(candidates, func(i, j int) bool {
|
||||
return candidates[i].hybridScore > candidates[j].hybridScore
|
||||
})
|
||||
|
||||
// 转换为结果
|
||||
results := make([]*RetrievalResult, len(candidates))
|
||||
for i, cand := range candidates {
|
||||
results := make([]*RetrievalResult, len(filtered))
|
||||
for i, c := range filtered {
|
||||
results[i] = &RetrievalResult{
|
||||
Chunk: cand.chunk,
|
||||
Item: cand.item,
|
||||
Similarity: cand.similarity,
|
||||
Score: cand.hybridScore,
|
||||
Chunk: c.chunk,
|
||||
Item: c.item,
|
||||
Similarity: c.similarity,
|
||||
Score: c.similarity,
|
||||
}
|
||||
}
|
||||
|
||||
// 上下文扩展:为每个匹配的chunk添加同一文档中的相关chunk
|
||||
// 这可以防止文本描述和payload被分开切分时,只返回描述而丢失payload的问题
|
||||
results = r.expandContext(ctx, results)
|
||||
|
||||
return results, nil
|
||||
}
|
||||
|
||||
// expandContext 扩展检索结果的上下文
|
||||
// 对于每个匹配的chunk,自动包含同一文档中的相关chunk(特别是包含代码块、payload的chunk)
|
||||
func (r *Retriever) expandContext(ctx context.Context, results []*RetrievalResult) []*RetrievalResult {
|
||||
if len(results) == 0 {
|
||||
return results
|
||||
}
|
||||
|
||||
// 收集所有匹配到的文档ID
|
||||
itemIDs := make(map[string]bool)
|
||||
for _, result := range results {
|
||||
itemIDs[result.Item.ID] = true
|
||||
}
|
||||
|
||||
// 为每个文档加载所有chunk
|
||||
itemChunksMap := make(map[string][]*KnowledgeChunk)
|
||||
for itemID := range itemIDs {
|
||||
chunks, err := r.loadAllChunksForItem(itemID)
|
||||
if err != nil {
|
||||
r.logger.Warn("加载文档chunk失败", zap.String("itemId", itemID), zap.Error(err))
|
||||
continue
|
||||
}
|
||||
itemChunksMap[itemID] = chunks
|
||||
}
|
||||
|
||||
// 按文档分组结果,每个文档只扩展一次
|
||||
resultsByItem := make(map[string][]*RetrievalResult)
|
||||
for _, result := range results {
|
||||
itemID := result.Item.ID
|
||||
resultsByItem[itemID] = append(resultsByItem[itemID], result)
|
||||
}
|
||||
|
||||
// 扩展每个文档的结果
|
||||
expandedResults := make([]*RetrievalResult, 0, len(results))
|
||||
processedChunkIDs := make(map[string]bool) // 避免重复添加
|
||||
|
||||
for itemID, itemResults := range resultsByItem {
|
||||
// 获取该文档的所有chunk
|
||||
allChunks, exists := itemChunksMap[itemID]
|
||||
if !exists {
|
||||
// 如果无法加载chunk,直接添加原始结果
|
||||
for _, result := range itemResults {
|
||||
if !processedChunkIDs[result.Chunk.ID] {
|
||||
expandedResults = append(expandedResults, result)
|
||||
processedChunkIDs[result.Chunk.ID] = true
|
||||
}
|
||||
}
|
||||
continue
|
||||
}
|
||||
|
||||
// 添加原始结果
|
||||
for _, result := range itemResults {
|
||||
if !processedChunkIDs[result.Chunk.ID] {
|
||||
expandedResults = append(expandedResults, result)
|
||||
processedChunkIDs[result.Chunk.ID] = true
|
||||
}
|
||||
}
|
||||
|
||||
// 为该文档的匹配chunk收集需要扩展的相邻chunk
|
||||
// 策略:只对混合分数最高的前3个匹配chunk进行扩展,避免扩展过多
|
||||
// 先按混合分数排序,只扩展前3个(使用混合分数而不是相似度)
|
||||
sortedItemResults := make([]*RetrievalResult, len(itemResults))
|
||||
copy(sortedItemResults, itemResults)
|
||||
sort.Slice(sortedItemResults, func(i, j int) bool {
|
||||
return sortedItemResults[i].Score > sortedItemResults[j].Score
|
||||
})
|
||||
|
||||
// 只扩展前3个(或所有,如果少于3个)
|
||||
maxExpandFrom := 3
|
||||
if len(sortedItemResults) < maxExpandFrom {
|
||||
maxExpandFrom = len(sortedItemResults)
|
||||
}
|
||||
|
||||
// 使用map去重,避免同一个chunk被多次添加
|
||||
relatedChunksMap := make(map[string]*KnowledgeChunk)
|
||||
|
||||
for i := 0; i < maxExpandFrom; i++ {
|
||||
result := sortedItemResults[i]
|
||||
// 查找相关chunk(上下各2个,排除已处理的chunk)
|
||||
relatedChunks := r.findRelatedChunks(result.Chunk, allChunks, processedChunkIDs)
|
||||
for _, relatedChunk := range relatedChunks {
|
||||
// 使用chunk ID作为key去重
|
||||
if !processedChunkIDs[relatedChunk.ID] {
|
||||
relatedChunksMap[relatedChunk.ID] = relatedChunk
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
// 限制每个文档最多扩展的chunk数量(避免扩展过多)
|
||||
// 策略:最多扩展8个chunk,无论匹配了多少个chunk
|
||||
// 这样可以避免当多个匹配chunk分散在文档不同位置时,扩展出过多chunk
|
||||
maxExpandPerItem := 8
|
||||
|
||||
// 将相关chunk转换为切片并按索引排序,优先选择距离匹配chunk最近的
|
||||
relatedChunksList := make([]*KnowledgeChunk, 0, len(relatedChunksMap))
|
||||
for _, chunk := range relatedChunksMap {
|
||||
relatedChunksList = append(relatedChunksList, chunk)
|
||||
}
|
||||
|
||||
// 计算每个相关chunk到最近匹配chunk的距离,按距离排序
|
||||
sort.Slice(relatedChunksList, func(i, j int) bool {
|
||||
// 计算到最近匹配chunk的距离
|
||||
minDistI := len(allChunks)
|
||||
minDistJ := len(allChunks)
|
||||
for _, result := range itemResults {
|
||||
distI := abs(relatedChunksList[i].ChunkIndex - result.Chunk.ChunkIndex)
|
||||
distJ := abs(relatedChunksList[j].ChunkIndex - result.Chunk.ChunkIndex)
|
||||
if distI < minDistI {
|
||||
minDistI = distI
|
||||
}
|
||||
if distJ < minDistJ {
|
||||
minDistJ = distJ
|
||||
}
|
||||
}
|
||||
return minDistI < minDistJ
|
||||
})
|
||||
|
||||
// 限制数量
|
||||
if len(relatedChunksList) > maxExpandPerItem {
|
||||
relatedChunksList = relatedChunksList[:maxExpandPerItem]
|
||||
}
|
||||
|
||||
// 添加去重后的相关chunk
|
||||
// 使用该文档中混合分数最高的结果作为参考
|
||||
maxScore := 0.0
|
||||
maxSimilarity := 0.0
|
||||
for _, result := range itemResults {
|
||||
if result.Score > maxScore {
|
||||
maxScore = result.Score
|
||||
}
|
||||
if result.Similarity > maxSimilarity {
|
||||
maxSimilarity = result.Similarity
|
||||
}
|
||||
}
|
||||
|
||||
// 计算扩展chunk的混合分数(使用相同的混合权重)
|
||||
hybridWeight := r.config.HybridWeight
|
||||
expandedSimilarity := maxSimilarity * 0.8 // 相关chunk的相似度略低
|
||||
// 对于扩展的chunk,BM25分数设为0(因为它们是上下文扩展,不是直接匹配)
|
||||
expandedBM25 := 0.0
|
||||
expandedScore := hybridWeight*expandedSimilarity + (1-hybridWeight)*expandedBM25
|
||||
|
||||
for _, relatedChunk := range relatedChunksList {
|
||||
expandedResult := &RetrievalResult{
|
||||
Chunk: relatedChunk,
|
||||
Item: itemResults[0].Item, // 使用第一个结果的Item信息
|
||||
Similarity: expandedSimilarity,
|
||||
Score: expandedScore, // 使用正确的混合分数
|
||||
}
|
||||
expandedResults = append(expandedResults, expandedResult)
|
||||
processedChunkIDs[relatedChunk.ID] = true
|
||||
}
|
||||
}
|
||||
|
||||
return expandedResults
|
||||
}
|
||||
|
||||
// loadAllChunksForItem 加载文档的所有chunk
|
||||
func (r *Retriever) loadAllChunksForItem(itemID string) ([]*KnowledgeChunk, error) {
|
||||
rows, err := r.db.Query(`
|
||||
SELECT id, item_id, chunk_index, chunk_text, embedding
|
||||
FROM knowledge_embeddings
|
||||
WHERE item_id = ?
|
||||
ORDER BY chunk_index
|
||||
`, itemID)
|
||||
if err != nil {
|
||||
return nil, fmt.Errorf("查询chunk失败: %w", err)
|
||||
}
|
||||
defer rows.Close()
|
||||
|
||||
var chunks []*KnowledgeChunk
|
||||
for rows.Next() {
|
||||
var chunkID, itemID, chunkText, embeddingJSON string
|
||||
var chunkIndex int
|
||||
|
||||
if err := rows.Scan(&chunkID, &itemID, &chunkIndex, &chunkText, &embeddingJSON); err != nil {
|
||||
r.logger.Warn("扫描chunk失败", zap.Error(err))
|
||||
continue
|
||||
}
|
||||
|
||||
// 解析向量(可选,这里不需要)
|
||||
var embedding []float32
|
||||
if embeddingJSON != "" {
|
||||
json.Unmarshal([]byte(embeddingJSON), &embedding)
|
||||
}
|
||||
|
||||
chunk := &KnowledgeChunk{
|
||||
ID: chunkID,
|
||||
ItemID: itemID,
|
||||
ChunkIndex: chunkIndex,
|
||||
ChunkText: chunkText,
|
||||
Embedding: embedding,
|
||||
}
|
||||
chunks = append(chunks, chunk)
|
||||
}
|
||||
|
||||
return chunks, nil
|
||||
}
|
||||
|
||||
// findRelatedChunks 查找与给定chunk相关的其他chunk
|
||||
// 策略:只返回上下各2个相邻的chunk(共最多4个)
|
||||
// 排除已处理的chunk,避免重复添加
|
||||
func (r *Retriever) findRelatedChunks(targetChunk *KnowledgeChunk, allChunks []*KnowledgeChunk, processedChunkIDs map[string]bool) []*KnowledgeChunk {
|
||||
related := make([]*KnowledgeChunk, 0)
|
||||
|
||||
// 查找上下各2个相邻chunk
|
||||
for _, chunk := range allChunks {
|
||||
if chunk.ID == targetChunk.ID {
|
||||
continue
|
||||
}
|
||||
|
||||
// 检查是否已经被处理过(可能已经在检索结果中)
|
||||
if processedChunkIDs[chunk.ID] {
|
||||
continue
|
||||
}
|
||||
|
||||
// 检查是否是相邻chunk(索引相差不超过2,且不为0)
|
||||
indexDiff := chunk.ChunkIndex - targetChunk.ChunkIndex
|
||||
if indexDiff >= -2 && indexDiff <= 2 && indexDiff != 0 {
|
||||
related = append(related, chunk)
|
||||
}
|
||||
}
|
||||
|
||||
// 按索引距离排序,优先选择最近的
|
||||
sort.Slice(related, func(i, j int) bool {
|
||||
diffI := abs(related[i].ChunkIndex - targetChunk.ChunkIndex)
|
||||
diffJ := abs(related[j].ChunkIndex - targetChunk.ChunkIndex)
|
||||
return diffI < diffJ
|
||||
})
|
||||
|
||||
// 限制最多返回4个(上下各2个)
|
||||
if len(related) > 4 {
|
||||
related = related[:4]
|
||||
}
|
||||
|
||||
return related
|
||||
}
|
||||
|
||||
// abs 返回整数的绝对值
|
||||
func abs(x int) int {
|
||||
if x < 0 {
|
||||
return -x
|
||||
}
|
||||
return x
|
||||
// AsEinoRetriever 将纯向量检索暴露为 Eino [retriever.Retriever]。
|
||||
func (r *Retriever) AsEinoRetriever() retriever.Retriever {
|
||||
return NewVectorEinoRetriever(r)
|
||||
}
|
||||
Reference in new issue
Block a user