milvus/internal/util/function/rerank/decay_function.go
congqixia 7d13bdcf4c
fix: [2.6] correct field data offset calculation in rerank functions for bulk search (#45444) (#45482)
Cherry-pick from master
pr: #45444
Related to #45338

When using bulk vector search in hybrid search with rerank functions,
the output field values for different queries were all equal to the
values returned by the first query, instead of the correct values
belonging to each document ID. The document IDs were correct, but the
entity field values were wrong.

In rerank functions (RRF, weighted, decay, model), when processing
multiple queries in a batch, the `idLocations` stored only the relative
offset within each result set (`idx`), not accounting for the absolute
position within the entire batch. This caused `FillFieldData` to
retrieve field data from the wrong positions, always using offsets
relative to the first query.

This fix ensures that when processing bulk searches with rerank
functions, each result correctly retrieves its corresponding field data
based on the absolute offset within the entire batch, resolving the
issue where all queries returned the first query's field values.

Signed-off-by: Congqi Xia <congqi.xia@zilliz.com>
2025-11-11 18:15:37 +08:00

242 lines
8.0 KiB
Go

/*
* # Licensed to the LF AI & Data foundation under one
* # or more contributor license agreements. See the NOTICE file
* # distributed with this work for additional information
* # regarding copyright ownership. The ASF licenses this file
* # to you 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 rerank
import (
"context"
"fmt"
"math"
"strconv"
"strings"
"github.com/milvus-io/milvus-proto/go-api/v2/schemapb"
)
const (
originKey string = "origin"
scaleKey string = "scale"
offsetKey string = "offset"
decayKey string = "decay"
functionKey string = "function"
normsScorekey string = "norm_score"
scoreMode string = "score_mode"
)
const (
gaussFunction string = "gauss"
linearFunction string = "linear"
expFunction string = "exp"
)
type DecayFunction[T PKType, R int32 | int64 | float32 | float64] struct {
RerankBase
functionName string
origin float64
scale float64
offset float64
decay float64
needNorm bool
scoreFunc scoreMergeFunc[T]
reScorer decayReScorer
}
func newDecayFunction(collSchema *schemapb.CollectionSchema, funcSchema *schemapb.FunctionSchema) (Reranker, error) {
base, err := newRerankBase(collSchema, funcSchema, DecayFunctionName, true)
if err != nil {
return nil, err
}
if len(base.GetInputFieldNames()) != 1 {
return nil, fmt.Errorf("Decay function only supports single input, but gets [%s] input", base.GetInputFieldNames())
}
inputType := base.GetInputFieldTypes()[0]
if base.pkType == schemapb.DataType_Int64 {
switch inputType {
case schemapb.DataType_Int8, schemapb.DataType_Int16, schemapb.DataType_Int32:
return newFunction[int64, int32](base, funcSchema)
case schemapb.DataType_Int64, schemapb.DataType_Timestamptz:
return newFunction[int64, int64](base, funcSchema)
case schemapb.DataType_Float:
return newFunction[int64, float32](base, funcSchema)
case schemapb.DataType_Double:
return newFunction[int64, float64](base, funcSchema)
default:
return nil, fmt.Errorf("Decay rerank: unsupported input field type:%s, only support numberic field", inputType.String())
}
} else {
switch inputType {
case schemapb.DataType_Int8, schemapb.DataType_Int16, schemapb.DataType_Int32:
return newFunction[string, int32](base, funcSchema)
case schemapb.DataType_Int64, schemapb.DataType_Timestamptz:
return newFunction[string, int64](base, funcSchema)
case schemapb.DataType_Float:
return newFunction[string, float32](base, funcSchema)
case schemapb.DataType_Double:
return newFunction[string, float64](base, funcSchema)
default:
return nil, fmt.Errorf("Decay rerank: unsupported input field type:%s, only support numberic field", inputType.String())
}
}
}
// T: PK Type, R: field type
func newFunction[T PKType, R int32 | int64 | float32 | float64](base *RerankBase, funcSchema *schemapb.FunctionSchema) (Reranker, error) {
var err error
decayFunc := &DecayFunction[T, R]{RerankBase: *base, offset: 0, decay: 0.5, needNorm: false, scoreFunc: maxMerge[T]}
orginInit := false
scaleInit := false
for _, param := range funcSchema.Params {
switch strings.ToLower(param.Key) {
case functionKey:
decayFunc.functionName = param.Value
case originKey:
if decayFunc.origin, err = strconv.ParseFloat(param.Value, 64); err != nil {
return nil, fmt.Errorf("Param origin:%s is not a number", param.Value)
}
orginInit = true
case scaleKey:
if decayFunc.scale, err = strconv.ParseFloat(param.Value, 64); err != nil {
return nil, fmt.Errorf("Param scale:%s is not a number", param.Value)
}
scaleInit = true
case offsetKey:
if decayFunc.offset, err = strconv.ParseFloat(param.Value, 64); err != nil {
return nil, fmt.Errorf("Param offset:%s is not a number", param.Value)
}
case decayKey:
if decayFunc.decay, err = strconv.ParseFloat(param.Value, 64); err != nil {
return nil, fmt.Errorf("Param decay:%s is not a number", param.Value)
}
case normsScorekey:
if needNorm, err := strconv.ParseBool(param.Value); err != nil {
return nil, fmt.Errorf("%s params must be true/false, bug got %s", normsScorekey, param.Value)
} else {
decayFunc.needNorm = needNorm
}
case scoreMode:
if f, err := getMergeFunc[T](param.Value); err != nil {
return nil, err
} else {
decayFunc.scoreFunc = f
}
default:
}
}
if !orginInit {
return nil, fmt.Errorf("Decay function lost param: origin")
}
if !scaleInit {
return nil, fmt.Errorf("Decay function lost param: scale")
}
if decayFunc.scale <= 0 {
return nil, fmt.Errorf("Decay function param: scale must > 0, but got %f", decayFunc.scale)
}
if decayFunc.offset < 0 {
return nil, fmt.Errorf("Decay function param: offset must >= 0, but got %f", decayFunc.offset)
}
if decayFunc.decay <= 0 || decayFunc.decay >= 1 {
return nil, fmt.Errorf("Decay function param: decay must 0 < decay < 1, but got %f", decayFunc.decay)
}
switch decayFunc.functionName {
case gaussFunction:
decayFunc.reScorer = gaussianDecay
case expFunction:
decayFunc.reScorer = expDecay
case linearFunction:
decayFunc.reScorer = linearDecay
default:
return nil, fmt.Errorf("Invaild decay function: %s, only support [%s,%s,%s]", DecayFunctionName, gaussFunction, linearFunction, expFunction)
}
return decayFunc, nil
}
func (decay *DecayFunction[T, R]) processOneSearchData(ctx context.Context, searchParams *SearchParams, cols []*columns, idGroup map[any]any) (*IDScores[T], error) {
srcScores := decay.scoreFunc(cols)
decayScores := map[T]float32{}
idLocations := make(map[T]IDLoc)
for i, col := range cols {
if col.size == 0 {
continue
}
nums := col.data[0].([]R)
ids := col.ids.([]T)
for idx, id := range ids {
if _, ok := decayScores[id]; !ok {
idLocations[id] = IDLoc{batchIdx: i, offset: idx + int(col.nqOffset)}
decayScores[id] = float32(decay.reScorer(decay.origin, decay.scale, decay.decay, decay.offset, float64(nums[idx])))
}
}
}
for id := range decayScores {
decayScores[id] = decayScores[id] * srcScores[id]
}
if searchParams.isGrouping() {
return newGroupingIDScores(decayScores, idLocations, searchParams, idGroup)
}
return newIDScores(decayScores, idLocations, searchParams, true), nil
}
func (decay *DecayFunction[T, R]) Process(ctx context.Context, searchParams *SearchParams, inputs *rerankInputs) (*rerankOutputs, error) {
outputs := newRerankOutputs(inputs, searchParams)
for _, cols := range inputs.data {
for i, col := range cols {
normFunc := getNormalizeFunc(decay.needNorm, searchParams.searchMetrics[i], true)
for j, score := range col.scores {
col.scores[j] = normFunc(score)
}
}
idScore, err := decay.processOneSearchData(ctx, searchParams, cols, inputs.idGroupValue)
if err != nil {
return nil, err
}
appendResult(inputs, outputs, idScore)
}
return outputs, nil
}
type decayReScorer func(float64, float64, float64, float64, float64) float64
func gaussianDecay(origin, scale, decay, offset, distance float64) float64 {
adjustedDist := math.Max(0, math.Abs(distance-origin)-offset)
sigmaSquare := math.Pow(scale, 2.0) / math.Log(decay)
exponent := math.Pow(adjustedDist, 2.0) / sigmaSquare
return math.Exp(exponent)
}
func expDecay(origin, scale, decay, offset, distance float64) float64 {
adjustedDist := math.Max(0, math.Abs(distance-origin)-offset)
lambda := math.Log(decay) / scale
return math.Exp(lambda * adjustedDist)
}
func linearDecay(origin, scale, decay, offset, distance float64) float64 {
adjustedDist := math.Max(0, math.Abs(distance-origin)-offset)
slope := (1 - decay) / scale
return math.Max(decay, 1-slope*adjustedDist)
}