293 lines
7.7 KiB
Go
293 lines
7.7 KiB
Go
package eino
|
||
|
||
import (
|
||
"context"
|
||
"errors"
|
||
"rag/dao"
|
||
"sort"
|
||
"time"
|
||
|
||
"github.com/cloudwego/eino/callbacks"
|
||
"github.com/cloudwego/eino/components/embedding"
|
||
"github.com/cloudwego/eino/components/retriever"
|
||
"github.com/cloudwego/eino/schema"
|
||
"github.com/gogf/gf/v2/frame/g"
|
||
"github.com/gogf/gf/v2/os/grpool"
|
||
"github.com/gogf/gf/v2/util/gconv"
|
||
"github.com/pgvector/pgvector-go"
|
||
)
|
||
|
||
type PGVectorRetrieverConfig struct {
|
||
Embedder embedding.Embedder
|
||
DefaultTopK int
|
||
DefaultIndex string
|
||
DSLInfo map[string]any
|
||
}
|
||
|
||
type PGVectorRetriever struct {
|
||
embedder embedding.Embedder
|
||
topK int
|
||
index string
|
||
dslInfo map[string]any
|
||
}
|
||
|
||
func NewPGVectorRetriever(config *PGVectorRetrieverConfig) (*PGVectorRetriever, error) {
|
||
if config.Embedder == nil {
|
||
return nil, errors.New("embedder is required")
|
||
}
|
||
if config.DefaultTopK <= 0 {
|
||
config.DefaultTopK = 5
|
||
}
|
||
|
||
return &PGVectorRetriever{
|
||
embedder: config.Embedder,
|
||
topK: config.DefaultTopK,
|
||
index: config.DefaultIndex,
|
||
dslInfo: config.DSLInfo,
|
||
}, nil
|
||
}
|
||
|
||
func (r *PGVectorRetriever) Retrieve(ctx context.Context, query string, opts ...retriever.Option) ([]*schema.Document, error) {
|
||
options := &retriever.Options{
|
||
Index: &r.index,
|
||
TopK: &r.topK,
|
||
DSLInfo: r.dslInfo,
|
||
Embedding: r.embedder,
|
||
}
|
||
options = retriever.GetCommonOptions(options, opts...)
|
||
|
||
// 安全保护：防止 nil 指针 panic
|
||
topK := 10
|
||
if options.TopK != nil {
|
||
topK = *options.TopK
|
||
}
|
||
|
||
ctx = callbacks.OnStart(ctx, &retriever.CallbackInput{
|
||
Query: query,
|
||
TopK: *options.TopK,
|
||
})
|
||
|
||
// ==========================================
|
||
// 🔥 优化版：grpool 并行双路检索（安全、健壮、无泄漏）
|
||
// ==========================================
|
||
var (
|
||
docsVector []*schema.Document
|
||
docsFulltext []*schema.Document
|
||
errVector error
|
||
errFulltext error
|
||
|
||
// 缓冲通道=2，确保无死锁等待
|
||
done = make(chan struct{}, 2)
|
||
)
|
||
|
||
// 上下文：超时 + 可取消双保障（建议5s超时，根据业务调整）
|
||
taskCtx, cancel := context.WithTimeout(ctx, 5*time.Second)
|
||
defer cancel()
|
||
|
||
// 封装并行任务函数，消除重复代码
|
||
runTask := func(task func() error, errTarget *error) {
|
||
defer func() {
|
||
// 任务结束必发信号，确保通道不阻塞
|
||
done <- struct{}{}
|
||
}()
|
||
|
||
// 捕获 panic + 执行业务逻辑
|
||
g.TryCatch(taskCtx, func(ctx context.Context) {
|
||
*errTarget = task()
|
||
}, func(ctx context.Context, panicErr error) {
|
||
*errTarget = panicErr
|
||
})
|
||
|
||
// 任务失败：立即取消另一个任务（快速失败）
|
||
if *errTarget != nil {
|
||
cancel()
|
||
}
|
||
}
|
||
|
||
// ----------------------
|
||
// 并行提交两个检索任务
|
||
// ----------------------
|
||
// 任务1：向量检索
|
||
grpool.Add(taskCtx, func(ctx context.Context) {
|
||
runTask(func() error {
|
||
docsVector, errVector = r.doRetrieveVector(ctx, query, options)
|
||
return errVector
|
||
}, &errVector)
|
||
})
|
||
|
||
// 任务2：全文检索
|
||
grpool.Add(taskCtx, func(ctx context.Context) {
|
||
runTask(func() error {
|
||
docsFulltext, errFulltext = r.doRetrieveMeilisearch(ctx, query, options)
|
||
return errFulltext
|
||
}, &errFulltext)
|
||
})
|
||
|
||
// ----------------------
|
||
// 安全等待所有任务完成
|
||
// ----------------------
|
||
<-done
|
||
<-done
|
||
|
||
// ----------------------
|
||
// 统一错误处理
|
||
// ----------------------
|
||
// 用 errors.Join 合并所有错误，不丢失信息
|
||
if err := errors.Join(errVector, errFulltext); err != nil {
|
||
return nil, err
|
||
}
|
||
|
||
// 合并 + 智能去重（保留最优分数）
|
||
docs := mergeAndDeduplicate(docsVector, docsFulltext)
|
||
|
||
// 排序：向量优先，同类型按距离升序
|
||
sort.Slice(docs, func(i, j int) bool {
|
||
//byI, okI := docs[i].MetaData["retrieve_by"].(string)
|
||
//byJ, okJ := docs[j].MetaData["retrieve_by"].(string)
|
||
//
|
||
//// 有类型标记的优先
|
||
//if okI && !okJ {
|
||
// return true
|
||
//}
|
||
//if !okI && okJ {
|
||
// return false
|
||
//}
|
||
//
|
||
//// 向量永远排前面
|
||
//if byI == "vector" && byJ == "fulltext" {
|
||
// return true
|
||
//}
|
||
//if byI == "fulltext" && byJ == "vector" {
|
||
// return false
|
||
//}
|
||
|
||
// 同类型按 distance 升序（越小越相似）
|
||
d1 := gconv.Float64(docs[i].MetaData["distance"])
|
||
d2 := gconv.Float64(docs[j].MetaData["distance"])
|
||
return d1 < d2
|
||
})
|
||
|
||
// 在Retrieve方法末尾，增加相关性校验
|
||
validDocs := make([]*schema.Document, 0)
|
||
for i, d := range docs {
|
||
// 过滤distance过大的垃圾结果（比如distance>0.8的直接丢弃）
|
||
if gconv.Float64(docs[i].MetaData["distance"]) < 0.8 {
|
||
validDocs = append(validDocs, d)
|
||
}
|
||
}
|
||
|
||
// 如果没有有效结果，返回空，让LLM回答「暂无相关信息」
|
||
if len(validDocs) == 0 {
|
||
callbacks.OnEnd(ctx, &retriever.CallbackOutput{Docs: validDocs})
|
||
return validDocs, nil
|
||
}
|
||
|
||
// 最多保留 topK
|
||
if len(validDocs) > topK {
|
||
validDocs = validDocs[:topK]
|
||
}
|
||
|
||
callbacks.OnEnd(ctx, &retriever.CallbackOutput{Docs: validDocs})
|
||
return validDocs, nil
|
||
}
|
||
|
||
// ==========================================
|
||
// 1. 向量检索（PG）
|
||
// ==========================================
|
||
func (r *PGVectorRetriever) doRetrieveVector(ctx context.Context, query string, opts *retriever.Options) ([]*schema.Document, error) {
|
||
vectors, err := opts.Embedding.EmbedStrings(ctx, []string{query})
|
||
if err != nil {
|
||
return nil, err
|
||
}
|
||
if len(vectors) == 0 {
|
||
return nil, errors.New("empty query vector")
|
||
}
|
||
|
||
queryVec := pgvector.NewVector(gconv.Float32s(vectors[0]))
|
||
topK := 10
|
||
if opts.TopK != nil {
|
||
topK = *opts.TopK
|
||
}
|
||
datasetIds := gconv.Int64s(opts.DSLInfo["dataset_ids"])
|
||
|
||
rows, err := dao.DocumentVector.GetAllByVector(ctx, datasetIds, queryVec, topK)
|
||
if err != nil {
|
||
return nil, err
|
||
}
|
||
|
||
docs := make([]*schema.Document, 0, len(rows))
|
||
for _, row := range rows {
|
||
docs = append(docs, &schema.Document{
|
||
ID: gconv.String(row["id"]),
|
||
Content: gconv.String(row["content"]),
|
||
MetaData: map[string]any{
|
||
"dataset_id": gconv.Int64(row["dataset_id"]),
|
||
"document_id": gconv.Int64(row["document_id"]),
|
||
"distance": gconv.Float64(row["distance"]),
|
||
"retrieve_by": "vector",
|
||
},
|
||
})
|
||
}
|
||
return docs, nil
|
||
}
|
||
|
||
// ==========================================
|
||
// 2. 全文检索（Meilisearch）🔥 新增
|
||
// ==========================================
|
||
func (r *PGVectorRetriever) doRetrieveMeilisearch(ctx context.Context, query string, opts *retriever.Options) ([]*schema.Document, error) {
|
||
topK := *opts.TopK
|
||
datasetIds := gconv.Int64s(opts.DSLInfo["dataset_ids"])
|
||
|
||
// 调用你已有的 Meilisearch DAO
|
||
rows, err := dao.DocumentVector.SearchByKeywords(ctx, query, datasetIds, topK)
|
||
if err != nil {
|
||
return nil, err
|
||
}
|
||
|
||
docs := make([]*schema.Document, 0, len(rows))
|
||
for _, row := range rows {
|
||
score := gconv.Float64(row["_rankingScore"])
|
||
distance := score
|
||
|
||
docs = append(docs, &schema.Document{
|
||
ID: gconv.String(row["id"]),
|
||
Content: gconv.String(row["content"]),
|
||
MetaData: map[string]any{
|
||
"dataset_id": gconv.Int64(row["dataset_id"]),
|
||
"document_id": gconv.Int64(row["document_id"]),
|
||
"distance": distance,
|
||
"retrieve_by": "fulltext",
|
||
},
|
||
})
|
||
}
|
||
return docs, nil
|
||
}
|
||
|
||
// ==========================================
|
||
// 合并去重（智能版：两路都命中时，保留向量结果 + 全文标记）
|
||
// ==========================================
|
||
func mergeAndDeduplicate(vecDocs, fullDocs []*schema.Document) []*schema.Document {
|
||
idMap := make(map[string]*schema.Document)
|
||
|
||
// 先存入向量结果
|
||
for _, d := range vecDocs {
|
||
idMap[d.ID] = d
|
||
}
|
||
|
||
// 再处理全文：不存在则添加；存在则标记“双路命中”，不覆盖向量分数
|
||
for _, d := range fullDocs {
|
||
if existDoc, ok := idMap[d.ID]; ok {
|
||
// 标记同时被向量和全文检索到
|
||
existDoc.MetaData["retrieve_by"] = "both"
|
||
} else {
|
||
idMap[d.ID] = d
|
||
}
|
||
}
|
||
|
||
merged := make([]*schema.Document, 0, len(idMap))
|
||
for _, d := range idMap {
|
||
merged = append(merged, d)
|
||
}
|
||
return merged
|
||
}
|