Initial QSfera import
This commit is contained in:
+26
@@ -0,0 +1,26 @@
|
||||
// Copyright (c) 2025 Couchbase, Inc.
|
||||
//
|
||||
// Licensed under the Apache License, Version 2.0 (the "License");
|
||||
// you may not use this file except in compliance with the License.
|
||||
// You may not use this file except in compliance with the License.
|
||||
// You may obtain a copy of the License at
|
||||
//
|
||||
// http://www.apache.org/licenses/LICENSE-2.0
|
||||
//
|
||||
// Unless required by applicable law or agreed to in writing, software
|
||||
// distributed under the License is distributed on an "AS IS" BASIS,
|
||||
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
// See the License for the specific language governing permissions and
|
||||
// limitations under the License.
|
||||
|
||||
package fusion
|
||||
|
||||
import (
|
||||
"github.com/blevesearch/bleve/v2/search"
|
||||
)
|
||||
|
||||
type FusionResult struct {
|
||||
Hits search.DocumentMatchCollection
|
||||
Total uint64
|
||||
MaxScore float64
|
||||
}
|
||||
+143
@@ -0,0 +1,143 @@
|
||||
// Copyright (c) 2025 Couchbase, Inc.
|
||||
//
|
||||
// Licensed under the Apache License, Version 2.0 (the "License");
|
||||
// you may not use this file except in compliance with the License.
|
||||
// You may not use this file except in compliance with the License.
|
||||
// You may obtain a copy of the License at
|
||||
//
|
||||
// http://www.apache.org/licenses/LICENSE-2.0
|
||||
//
|
||||
// Unless required by applicable law or agreed to in writing, software
|
||||
// distributed under the License is distributed on an "AS IS" BASIS,
|
||||
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
// See the License for the specific language governing permissions and
|
||||
// limitations under the License.
|
||||
|
||||
package fusion
|
||||
|
||||
import (
|
||||
"fmt"
|
||||
|
||||
"github.com/blevesearch/bleve/v2/search"
|
||||
)
|
||||
|
||||
// formatRRFMessage builds the explanation string for a single component of the
|
||||
// Reciprocal Rank Fusion calculation.
|
||||
func formatRRFMessage(weight float64, rank int, rankConstant int) string {
|
||||
return fmt.Sprintf("rrf score (weight=%.3f, rank=%d, rank_constant=%d), normalized score of", weight, rank, rankConstant)
|
||||
}
|
||||
|
||||
// ReciprocalRankFusion applies Reciprocal Rank Fusion across the primary FTS
|
||||
// results and each KNN sub-query. Ranks are limited to `windowSize` per source,
|
||||
// weighted, and combined into a single fused score, with optional explanation
|
||||
// details.
|
||||
func ReciprocalRankFusion(hits search.DocumentMatchCollection, weights []float64, rankConstant int, windowSize int, numKNNQueries int, explain bool) *FusionResult {
|
||||
nHits := len(hits)
|
||||
if nHits == 0 || windowSize == 0 {
|
||||
return &FusionResult{
|
||||
Hits: search.DocumentMatchCollection{},
|
||||
Total: 0,
|
||||
MaxScore: 0.0,
|
||||
}
|
||||
}
|
||||
|
||||
limit := min(nHits, windowSize)
|
||||
|
||||
// precompute rank+scores to prevent additional division ops later
|
||||
rankReciprocals := make([]float64, limit)
|
||||
for i := range rankReciprocals {
|
||||
rankReciprocals[i] = 1.0 / float64(rankConstant+i+1)
|
||||
}
|
||||
|
||||
// init explanations if required
|
||||
var fusionExpl map[*search.DocumentMatch][]*search.Explanation
|
||||
if explain {
|
||||
fusionExpl = make(map[*search.DocumentMatch][]*search.Explanation, nHits)
|
||||
}
|
||||
|
||||
// The code here mainly deals with obtaining rank/score for fts hits.
|
||||
// First sort hits by score
|
||||
sortDocMatchesByScore(hits)
|
||||
|
||||
// Calculate fts rank+scores
|
||||
ftsWeight := weights[0]
|
||||
for i := 0; i < nHits; i++ {
|
||||
if i < windowSize {
|
||||
hit := hits[i]
|
||||
|
||||
// No fts scores from this hit onwards, break loop
|
||||
if hit.Score == 0.0 {
|
||||
break
|
||||
}
|
||||
|
||||
contrib := ftsWeight * rankReciprocals[i]
|
||||
hit.Score = contrib
|
||||
|
||||
if explain {
|
||||
expl := getFusionExplAt(
|
||||
hit,
|
||||
0,
|
||||
contrib,
|
||||
formatRRFMessage(ftsWeight, i+1, rankConstant),
|
||||
)
|
||||
fusionExpl[hit] = append(fusionExpl[hit], expl)
|
||||
}
|
||||
} else {
|
||||
// These FTS hits are not counted in the results, so set to 0
|
||||
hits[i].Score = 0.0
|
||||
}
|
||||
}
|
||||
|
||||
// Code from here is to calculate knn ranks and scores
|
||||
// iterate over each knn query and calculate knn rank+scores
|
||||
for queryIdx := 0; queryIdx < numKNNQueries; queryIdx++ {
|
||||
knnWeight := weights[queryIdx+1]
|
||||
// Sorts hits in decreasing order of hit.ScoreBreakdown[i]
|
||||
sortDocMatchesByBreakdown(hits, queryIdx)
|
||||
|
||||
for i := 0; i < nHits; i++ {
|
||||
// break if score breakdown doesn't exist (sort function puts these hits at the end)
|
||||
// or if we go past the windowSize
|
||||
_, scoreBreakdownExists := scoreBreakdownForQuery(hits[i], queryIdx)
|
||||
if i >= windowSize || !scoreBreakdownExists {
|
||||
break
|
||||
}
|
||||
|
||||
hit := hits[i]
|
||||
contrib := knnWeight * rankReciprocals[i]
|
||||
hit.Score += contrib
|
||||
|
||||
if explain {
|
||||
expl := getFusionExplAt(
|
||||
hit,
|
||||
queryIdx+1,
|
||||
contrib,
|
||||
formatRRFMessage(knnWeight, i+1, rankConstant),
|
||||
)
|
||||
fusionExpl[hit] = append(fusionExpl[hit], expl)
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
var maxScore float64
|
||||
for _, hit := range hits {
|
||||
if explain {
|
||||
finalizeFusionExpl(hit, fusionExpl[hit])
|
||||
}
|
||||
hit.ScoreBreakdown = nil
|
||||
|
||||
if hit.Score > maxScore {
|
||||
maxScore = hit.Score
|
||||
}
|
||||
}
|
||||
|
||||
sortDocMatchesByScore(hits)
|
||||
if nHits > windowSize {
|
||||
hits = hits[:windowSize]
|
||||
}
|
||||
return &FusionResult{
|
||||
Hits: hits,
|
||||
Total: uint64(len(hits)),
|
||||
MaxScore: maxScore,
|
||||
}
|
||||
}
|
||||
+164
@@ -0,0 +1,164 @@
|
||||
// Copyright (c) 2025 Couchbase, Inc.
|
||||
//
|
||||
// Licensed under the Apache License, Version 2.0 (the "License");
|
||||
// you may not use this file except in compliance with the License.
|
||||
// You may obtain a copy of the License at
|
||||
//
|
||||
// http://www.apache.org/licenses/LICENSE-2.0
|
||||
//
|
||||
// Unless required by applicable law or agreed to in writing, software
|
||||
// distributed under the License is distributed on an "AS IS" BASIS,
|
||||
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
// See the License for the specific language governing permissions and
|
||||
// limitations under the License.
|
||||
|
||||
package fusion
|
||||
|
||||
import (
|
||||
"fmt"
|
||||
|
||||
"github.com/blevesearch/bleve/v2/search"
|
||||
)
|
||||
|
||||
// formatRSFMessage builds the explanation string associated with a single
|
||||
// component of the Relative Score Fusion calculation.
|
||||
func formatRSFMessage(weight float64, normalizedScore float64, minScore float64, maxScore float64) string {
|
||||
return fmt.Sprintf("rsf score (weight=%.3f, normalized=%.6f, min=%.6f, max=%.6f), normalized score of",
|
||||
weight, normalizedScore, minScore, maxScore)
|
||||
}
|
||||
|
||||
// RelativeScoreFusion normalizes the best-scoring documents from the primary
|
||||
// FTS query and each KNN query, scales those normalized values by the supplied
|
||||
// weights, and combines them into a single fused score. Only the top
|
||||
// `windowSize` documents per source are considered, and explanations are
|
||||
// materialized lazily when requested.
|
||||
func RelativeScoreFusion(hits search.DocumentMatchCollection, weights []float64, windowSize int, numKNNQueries int, explain bool) *FusionResult {
|
||||
nHits := len(hits)
|
||||
if nHits == 0 || windowSize == 0 {
|
||||
return &FusionResult{
|
||||
Hits: search.DocumentMatchCollection{},
|
||||
Total: 0,
|
||||
MaxScore: 0.0,
|
||||
}
|
||||
}
|
||||
|
||||
// init explanations if required
|
||||
var fusionExpl map[*search.DocumentMatch][]*search.Explanation
|
||||
if explain {
|
||||
fusionExpl = make(map[*search.DocumentMatch][]*search.Explanation, nHits)
|
||||
}
|
||||
|
||||
// Code here for calculating fts results
|
||||
// Sort by fts scores
|
||||
sortDocMatchesByScore(hits)
|
||||
|
||||
// ftsLimit holds the total number of fts hits to consider for rsf
|
||||
ftsLimit := 0
|
||||
for _, hit := range hits {
|
||||
if hit.Score == 0.0 {
|
||||
break
|
||||
}
|
||||
ftsLimit++
|
||||
}
|
||||
ftsLimit = min(ftsLimit, windowSize)
|
||||
|
||||
// calculate fts scores
|
||||
if ftsLimit > 0 {
|
||||
max := hits[0].Score
|
||||
min := hits[ftsLimit-1].Score
|
||||
denom := max - min
|
||||
weight := weights[0]
|
||||
|
||||
for i := 0; i < ftsLimit; i++ {
|
||||
hit := hits[i]
|
||||
norm := 1.0
|
||||
if denom > 0 {
|
||||
norm = (hit.Score - min) / denom
|
||||
}
|
||||
contrib := weight * norm
|
||||
if explain {
|
||||
expl := getFusionExplAt(
|
||||
hit,
|
||||
0,
|
||||
norm,
|
||||
formatRSFMessage(weight, norm, min, max),
|
||||
)
|
||||
fusionExpl[hit] = append(fusionExpl[hit], expl)
|
||||
}
|
||||
hit.Score = contrib
|
||||
}
|
||||
for i := ftsLimit; i < nHits; i++ {
|
||||
// These FTS hits are not counted in the results, so set to 0
|
||||
hits[i].Score = 0.0
|
||||
}
|
||||
}
|
||||
|
||||
// Code from here is for calculating knn scores
|
||||
for queryIdx := 0; queryIdx < numKNNQueries; queryIdx++ {
|
||||
sortDocMatchesByBreakdown(hits, queryIdx)
|
||||
|
||||
// knnLimit holds the total number of knn hits retrieved for a specific knn query
|
||||
knnLimit := 0
|
||||
for _, hit := range hits {
|
||||
if _, ok := scoreBreakdownForQuery(hit, queryIdx); !ok {
|
||||
break
|
||||
}
|
||||
knnLimit++
|
||||
}
|
||||
knnLimit = min(knnLimit, windowSize)
|
||||
|
||||
// if limit is 0, skip calculating
|
||||
if knnLimit == 0 {
|
||||
continue
|
||||
}
|
||||
|
||||
max, _ := scoreBreakdownForQuery(hits[0], queryIdx)
|
||||
min, _ := scoreBreakdownForQuery(hits[knnLimit-1], queryIdx)
|
||||
denom := max - min
|
||||
weight := weights[queryIdx+1]
|
||||
|
||||
for i := 0; i < knnLimit; i++ {
|
||||
hit := hits[i]
|
||||
score, _ := scoreBreakdownForQuery(hit, queryIdx)
|
||||
norm := 1.0
|
||||
if denom > 0 {
|
||||
norm = (score - min) / denom
|
||||
}
|
||||
contrib := weight * norm
|
||||
if explain {
|
||||
expl := getFusionExplAt(
|
||||
hit,
|
||||
queryIdx+1,
|
||||
norm,
|
||||
formatRSFMessage(weight, norm, min, max),
|
||||
)
|
||||
fusionExpl[hit] = append(fusionExpl[hit], expl)
|
||||
}
|
||||
hit.Score += contrib
|
||||
}
|
||||
}
|
||||
|
||||
// Finalize scores
|
||||
var maxScore float64
|
||||
for _, hit := range hits {
|
||||
if explain {
|
||||
finalizeFusionExpl(hit, fusionExpl[hit])
|
||||
}
|
||||
if hit.Score > maxScore {
|
||||
maxScore = hit.Score
|
||||
}
|
||||
hit.ScoreBreakdown = nil
|
||||
}
|
||||
|
||||
sortDocMatchesByScore(hits)
|
||||
|
||||
if nHits > windowSize {
|
||||
hits = hits[:windowSize]
|
||||
}
|
||||
|
||||
return &FusionResult{
|
||||
Hits: hits,
|
||||
Total: uint64(len(hits)),
|
||||
MaxScore: maxScore,
|
||||
}
|
||||
}
|
||||
+111
@@ -0,0 +1,111 @@
|
||||
// Copyright (c) 2025 Couchbase, Inc.
|
||||
//
|
||||
// Licensed under the Apache License, Version 2.0 (the "License");
|
||||
// you may not use this file except in compliance with the License.
|
||||
// You may not use this file except in compliance with the License.
|
||||
// You may obtain a copy of the License at
|
||||
//
|
||||
// http://www.apache.org/licenses/LICENSE-2.0
|
||||
//
|
||||
// Unless required by applicable law or agreed to in writing, software
|
||||
// distributed under the License is distributed on an "AS IS" BASIS,
|
||||
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
// See the License for the specific language governing permissions and
|
||||
// limitations under the License.
|
||||
|
||||
package fusion
|
||||
|
||||
import (
|
||||
"sort"
|
||||
|
||||
"github.com/blevesearch/bleve/v2/search"
|
||||
)
|
||||
|
||||
// sortDocMatchesByScore orders the provided collection in-place by the primary
|
||||
// score in descending order, breaking ties with the original `HitNumber` to
|
||||
// ensure deterministic output.
|
||||
func sortDocMatchesByScore(hits search.DocumentMatchCollection) {
|
||||
if len(hits) < 2 {
|
||||
return
|
||||
}
|
||||
|
||||
sort.Slice(hits, func(a, b int) bool {
|
||||
i := hits[a]
|
||||
j := hits[b]
|
||||
if i.Score == j.Score {
|
||||
return i.HitNumber < j.HitNumber
|
||||
}
|
||||
return i.Score > j.Score
|
||||
})
|
||||
}
|
||||
|
||||
// scoreBreakdownForQuery fetches the score for a specific KNN query index from
|
||||
// the provided hit. The boolean return indicates whether the score is present.
|
||||
func scoreBreakdownForQuery(hit *search.DocumentMatch, idx int) (float64, bool) {
|
||||
if hit == nil || hit.ScoreBreakdown == nil {
|
||||
return 0, false
|
||||
}
|
||||
|
||||
score, ok := hit.ScoreBreakdown[idx]
|
||||
return score, ok
|
||||
}
|
||||
|
||||
// sortDocMatchesByBreakdown orders the hits in-place using the KNN score for
|
||||
// the supplied query index (descending), breaking ties with `HitNumber` and
|
||||
// placing hits without a score at the end.
|
||||
func sortDocMatchesByBreakdown(hits search.DocumentMatchCollection, queryIdx int) {
|
||||
if len(hits) < 2 {
|
||||
return
|
||||
}
|
||||
|
||||
sort.SliceStable(hits, func(a, b int) bool {
|
||||
left := hits[a]
|
||||
right := hits[b]
|
||||
|
||||
var leftScore float64
|
||||
leftOK := false
|
||||
if left != nil && left.ScoreBreakdown != nil {
|
||||
leftScore, leftOK = left.ScoreBreakdown[queryIdx]
|
||||
}
|
||||
|
||||
var rightScore float64
|
||||
rightOK := false
|
||||
if right != nil && right.ScoreBreakdown != nil {
|
||||
rightScore, rightOK = right.ScoreBreakdown[queryIdx]
|
||||
}
|
||||
|
||||
if leftOK && rightOK {
|
||||
if leftScore == rightScore {
|
||||
return left.HitNumber < right.HitNumber
|
||||
}
|
||||
return leftScore > rightScore
|
||||
}
|
||||
|
||||
if leftOK != rightOK {
|
||||
return leftOK
|
||||
}
|
||||
|
||||
return left.HitNumber < right.HitNumber
|
||||
})
|
||||
}
|
||||
|
||||
// getFusionExplAt copies the existing explanation child at the requested index
|
||||
// and wraps it in a new node describing how the fusion algorithm adjusted the
|
||||
// score.
|
||||
func getFusionExplAt(hit *search.DocumentMatch, i int, value float64, message string) *search.Explanation {
|
||||
return &search.Explanation{
|
||||
Value: value,
|
||||
Message: message,
|
||||
Children: []*search.Explanation{hit.Expl.Children[i]},
|
||||
}
|
||||
}
|
||||
|
||||
// finalizeFusionExpl installs the collection of fusion explanation children and
|
||||
// updates the root message so the caller sees the fused score as the sum of its
|
||||
// parts.
|
||||
func finalizeFusionExpl(hit *search.DocumentMatch, explChildren []*search.Explanation) {
|
||||
hit.Expl.Children = explChildren
|
||||
|
||||
hit.Expl.Value = hit.Score
|
||||
hit.Expl.Message = "sum of"
|
||||
}
|
||||
Reference in New Issue
Block a user