Documentation
¶
Overview ¶
Package result provides result fusion and re-ranking capabilities for multi-source search results. It implements Reciprocal Rank Fusion (RRF) for combining results from different indexers and cosine similarity-based re-ranking for improving result quality.
Index ¶
- func Compress(limit int, llm goChatCore.Client, hits []core.Hit) ([]core.Hit, error)
- func Dedup(hits []core.Hit) ([]core.Hit, error)
- func RRF(sources ...FusionSource) ([]core.Hit, error)
- func RRFWithK(k int, sources ...FusionSource) ([]core.Hit, error)
- func Rerank(query *query.SemanticQuery, hits []core.Hit) ([]core.Hit, error)
- type FusionSource
Constants ¶
This section is empty.
Variables ¶
This section is empty.
Functions ¶
func Dedup ¶
Dedup 对搜索结果语义去重 使用 MinHash 算法估算文本 Jaccard 相似度，对同 DocID 下相似度超过阈值的 chunk 进行合并 合并规则：同 DocID 下，内容相似度 >= threshold 的保留分数最高的，丢弃其余
func RRF ¶
func RRF(sources ...FusionSource) ([]core.Hit, error)
RRF performs Reciprocal Rank Fusion on multiple sources using default k=60.
The RRF formula: score(doc) = Σ weight_s / (k + rank_s)
Parameters:
- sources: variable number of FusionSource inputs
Returns:
- []core.Hit: fused results sorted by score in descending order
- error: non-nil only if sources is empty (returns nil, nil)
func RRFWithK ¶
func RRFWithK(k int, sources ...FusionSource) ([]core.Hit, error)
RRFWithK performs RRF fusion with a custom smoothing parameter k. Recommended k range: 5-100. Larger values are more tolerant of lower-ranked results.
func Rerank ¶
Rerank re-ranks search results using cosine similarity between query vector and content vectors.
In RAG systems, Re-Ranking refers to re-scoring initial retrieval results using a more precise method rather than simply ordering by initial retrieval scores.
This implementation uses the pre-computed query vector from SemanticQuery.Vector() and computes cosine similarity with each hit's content vector (encoded on-the-fly). Results are returned in descending order of cosine similarity score.
Types ¶
type FusionSource ¶
type FusionSource struct {
Name string // 源标识（如 "vector"、"keyword"、"graph"）
Hits []core.Hit // 该源的搜索结果，顺序 = 排名
Weight float32 // 源权重，0 表示 1.0
}
FusionSource represents search results from a single source (e.g., vector, keyword, graph indexer). The Hits field maintains the internal ranking order from that source.
Source Files